Sometime last year, the New York Times, the Guardian, USA Today and a few other major publishers began blocking the Internet Archive’s Wayback Machine from scraping their sites. As I wrote back in May, the publishers’ beef with the Archive had to do with its APIs, which allow third parties to access content from the Wayback Machine (including of course journalists working for the very publishers now blocking it). In the publishers’ view, the Archive’s APIs had become a back door for AI companies to scrape their content without authorization.

“A lot of these AI businesses are looking for readily available, structured databases of content,” the Guardian’s head of business affairs and licensing, Robert Hahn, told Nieman Lab in January. “The Internet Archive’s API would have been an obvious place to plug their own machines into and suck out the IP.”

The particular beef was new, but the gripe is an old one. Complaints about unguarded doors providing disapproved electronic access to information long pre-date AI scraping and even the internet itself.

In the Before Time, when playback of digital media from physical sources or via transmission was first happening, the Hollywood studios in particular worked themselves into a lather over the so-called “analog hole.” While digital-to-digital connections between devices, or between devices and transmission lines, could be encrypted to prevent the data passing through them from being intercepted and copied, analog signals passing through analog connections could not be, which meant the signals could be recorded. Thus, the “analog hole.”

The studios’ solution to that perceived problem was to lobby the government to allow them to disable any unprotected analog outputs on devices by means of an embedded signal, and to require that devices recognize and comply with the flag. Counter lobbying by device makers and some AstroTurfed “grass roots” pushback foiled the scheme.

The latest dispute over an unprotected door pits Google against SerpApi. SerpApi sells subscriptions to a service it describes as a “Google Search API.” It generates automated searches, then organizes the resulting information and packages it into structured search-results data for clients. Google complained that the automated searches imposed significant processing costs on it without producing any advertising revenue for Google to offset those costs. So, in January 2025 Google introduced SearchGuard, which blocks automated search requests it does not recognize to prevent SerpAPI and similar services from scraping it search results.

An arms race then ensued, with SerpAPI tweaking its system to evade SearchGuard and Google tweaking back. In December 2025, Google finally sued, accusing Serp of circumventing a technical protection measure and trafficking in a circumvention technology in violation of §1201 of the DMCA. The complaint echoed an earlier lawsuit brought by Reddit accusing SerpAPI of illegally scraping its archive, as well as scraping its archive content from Google, which is licensed by Reddit to use and display its content.

Setting aside the irony of Google complaining about another service diverting its ad revenue, the DMCA claims in its complaint against SerpAPI stuck many legal observers as odd. Search results, being collections of facts, generally are not copyrightable, so why was Google bringing its case under the Copyright Act?

Google argued that the the “knowledge panels” it displays alongside some search results often include copyrighted as well as non-copyrightable elements. By scraping them, therefore, SerpApi was committing copyright infringement.

Last week, however, in unusually swift fashion, U.S. District Judge Yvonne Gonzalez Rogers of the Northern District of California, tossed Google’ DMCA claims. In a 21-page order, Gonzalez Rogers ruled that since Google is not the copyright owner of any of the copyrighted material in the knowledge panels, and had not shown that it had been expressly authorized to protect that material by any rights owner, it lacked standing to bring a DMCA claim. Further, she held, a platform cannot conjure a copyright right from uncopyrightable information simply by placing an anti-bot system in front of it.

She gave Google 21 days to file an amended complaint, however, if it can show that any of its licensing agreements with rights owners authorized it to protect the licensor’s content from third parties.

SerpApi, as well as some commentators, hailed the judge’s ruling as a major win, for SerpApi, and for the open internet generally. Others have argued that Judge Gonzalez Rogers’ ruling was more nuanced than that, and that Google has a plausible path to reinstating its claims.

I don’t have the expertise to evaluate the legal arguments of either side. But the ruling, if it stands up, is likely to have practical implications for rights owners thinking of licensing their content.

Given the huge increase in the amount of automated web scraping activity triggered by AI, rights holders not wanting to see their content exfiltrated through an unguarded door may need to make sure any agreements with licensees expressly authorize them to implement active measures to prevent such scraping. They might further want to demand the right to specify or approve those measures, and to audit their effectiveness.

Rights holders may also need to weigh the value of putting more third parties under an obligation to protect their content by initiating additional licensing deals vs. the value of preserving traffic to their own sites by limiting the number of licensees.

The lawyers on both sides may also need to clarify who is liable in the event scraping happens anyway and what rights of action the parties have.

For their part, licensees may seek offsetting concessions in the price or terms of the license if the agreement imposes significant financial or operational costs on them from the need to implement bot-blocking measures.

Following the ruling, Google vowed to renew the fight. In comments to Ars Technica, Google’s spokesperson, José Castañeda, said the company plans to amend the complaint and is “pleased to see that the Court rejected nearly all of SerpApi’s legal arguments” otherwise attempting to dispute Google’s standing.

“We look forward to filing an amended complaint, as the Court invited us to do,” he added. “We remain committed to protecting our services and partners from unauthorized access.”

In yet another irony of the case, rights owners may find themselves hoping Google succeeds in its DMCA claims and hands them a few legal bricks to plugs the API hole.

Share

Subscribe now

At least 128 cases involving AI and copyright have been filed in the U.S., according to the tally kept by Prof. Edward Lee on his invaluable ChatGPTIsEatingTheWorld blog, of which 116 are currently active. The remaining dozen were dismissed, settled or remain on appeal. Another 162 cases have been filed outside the U.S.

The largest number of U.S. cases—50 of the 128—have been filed by authors or book publishers, followed by news publishers (18), music publishers or record labels (17), video and film producers (15), and visual artists (12). OpenAI has been on the receiving end of the largest number of those cases, at 22, followed by Meta (16), Microsoft (12), and Anthropic (11).

The vast majority of those cases concern the alleged unlicensed use of copyright-protected works to train generative AI models, which the plaintiffs argue is infringing on their exclusive rights. In the vast majority of those, the AI company defendants have raised a defense of fair use, maintaining that the works included in their training datasets were made publicly available to read by the rights owner and that the process of training an AI model is transformative and does not implicate any of the five defined rights included in the copyright bundle, and therefore is fair use.

At the core of all of those legal disputes is the question of who should get paid for and from the fruits of generative AI. Leaving aside for now Ed Zitron’s argument that no one currently is reaping any fruit from AI, and may never, under the current structure of the industry, AI developers are benefitting (if not profiting) from their use of copyrighted works to train their models while creators and rights owners, by and large, are not. What the public is gaining or losing is a matter of debate.

Unacknowledged in all those cases, however, is the question of whether resolving the fair use vs. infringement debate would—or even could—actually resolve the question of who gets paid. That’s the view advanced by Profs. Frank Pasquale of Cornell Law School and Thomas W. Malone of MIT Sloan School of Management, and Andrew Ting, a lecturer in law at George Washington University, in a provocative paper published last fall in the Journal of Technology and Intellectual Property.

The paper runs through the pro and con arguments to both sides of the debate and the objections to each, weighing them all against both the law and equity. It concludes that the law, as currently written and applied by courts, cannot fairly balance the equities involved in generative AI training, no matter how the current lawsuits are decided, because it cannot support a viable system of licensing that would allow for a market-based solution.

Instead, the authors propose adding a new independently licensable right to the current copyright bundle that they refer to as the learning right, or “learnright.”

“Thoughtful policy demands respect for the humans whose works are now fueling this surfeit of computerized output,” the authors write. “Given that the myriad copyright cases now slowly making their way through courts may never vindicate these creators, or may do so inadequately, there is a strong normative case for policymakers to support a “learnright” to supplement existing intellectual property protections.”

Since their paper was published, Malone and Pasquale have also published condensed versions of their argument in non-scholarly outlets, including here and here.

As described by Malone, et. al., the learnright would consist of an exclusive right governing machine learning from protected expressive works. It would not extend to any form of human learning.

Unlike the automatic protection provided to the fixation of expression, the authors propose a system of opt-in registration for creators to invoke the new right, somewhat analogous to patent and copyright registries. Registration would identify the rightsholder and provide sufficient information for AI companies to locate the owner and negotiate payment. Any work registered would require permission to be used to train AI models. Developers would also be required to disclose the contents of their datasets.

The authors assume that many, perhaps even most, creators whose work is not likely to have meaningful commercial value as training data would not bother to register their works. That would reduce transaction costs for AI companies by eliminating the need to clear every work they potentially could encounter, allowing them to license a curated set of the most valuable registered data.

Such a statutory structure, the authors contend, would allow for the development of a market-based licensing system based on blanket licensing , without the need of a compulsory license or government-set fees. They envision an ecosystem comprised of four components:

• Creators/rightsholders: authors, artists, publishers, news organizations and other content owners.

• AI companies: entities training and operating generative AI systems.

• Users: consumers and businesses using AI outputs.

• Agents/content brokers: intermediaries aggregating and licensing rights.

They envision the fourth stakeholder—agents and content brokers—as analogous to music PROs, offering blanket licenses for a particular category of works and responsible for dispensing the fees collected to rights holders.

“By aggregating various kinds of content from multiple sources, these Agents can greatly simplify the amount of negotiating that needs to go on in this market,” they write. “For example, one Agent might represent many of the major news publishers, another might represent sets of fiction authors (or their publishers), and so forth.”

The concept is not entirely new. Other analysts and commentators have proposed creating some sort of AI training right, either sui generis or as an addition to existing copyright law. But Malone, Pasquale and Ting deserve credit for offering a fully developed proposal for getting past the fair use vs. infringement defile by envisioning a legal structure to support a market-based licensing system that doesn’t require settling the question definitively.

Their proposed system would still face challenges, however. The first is definitional. How exactly would the legally covered act of “learning” be defined such that it doesn’t sweep up currently legal activity? The paper does not include any suggested statutory text.

The harder problem would be antitrust. If there is a single entity offering a license to content from all major news outlets, or all major book publishers, the potential for abuse would be significant. Without competition there would be nothing to restrain the price of a license absent government intervention. ASCAP and BMI were forced to operate under court supervision for decades due to exactly that problem.

There is also no discussion in the paper about how startups or open-source AI developers would fare under the proposed regime. They potentially could be priced out of the market for the most valuable data.

Depending on how the market ultimately shakes out the proposed structure could also face the inverse problem of monopsony, in which a small number of large buyers would be able to dictate prices, preventing rights holders from realizing the full value of their rights.

The paper is also silent on whether retraining, fine-tuning or retrieval-augmented generation would trigger the right. Nor does it specify whether the envisioned licensed would be for one-time use in training, or for a model’s ongoing use of its “learning” as it generates new outputs. If the latter, it’s unclear what would happen when the licensed has expired or been revoked. Would developers be forced to retrain their models without the expired content?

Also not clear is how the authors envision how large, multinational datasets would be handled across jurisdictions, including where the “learnright” has not been adopted.

Some of those challenges, of course, can likely be met through stakeholder negotiation and some clever lawyering. The bigger challenge would be political: actually enacting the necessary amendment to the Copyright Act.

The deeper potential challenge is conceptual. Locating the learnright within the domain of copyright may make a certain intuitive sense. But it’s a bit of odd fit. Unlike the reproduction, distribution, performance, display and derivative work rights, all of which are agnostic as to the method or technology used to exploit them, the envisioned learnright is use-case specific. It applies to the use of otherwise copyrighted works for the purpose of machine learning and in no other circumstance or technological context.

The closest analogy would be the digital audio performance right introduced in 1995. It applies to the public performance of non-dramatic music recordings via digital transmission, but not to the performance of the same recordings via over-the-air broadcast. In that case, the difference in application between use cases reflects the relative political influence of broadcasters and digital platform providers at the time of passage, not the balance of equities or any underlying principle related to the progress of useful arts and sciences. But political influence waxes and wanes with shifts in the market and culture.

The learnright might have a better claim on the balance of equities, but unlike the digital audio performance right it could also hold the potential for conflict with other elements of copyright. Should the Supreme Court ultimately decide that the use of copyrighted works to train AI models qualifies as fair use, for instance, would the learnright override that judgment, or be precluded by it? It’s not clear.

Careful legislative drafting might be able to avoid such potential conflicts to wedge the learnright into the Copyright Act, but it won’t be easy. And pity the first courts asked to draw the line.

To be clear, none of those caveats is intended as criticism of Malone, Pasquale and Ting. As noted, they deserve credit wrestling seriously with the complexity of the problem and offering a serious proposal to resolve it. Their paper is highly recommended reading. What it mostly reveals, however, is just how complex and difficult the challenge generative AI is to copyright.

Share

Subscribe now

A little over a year ago, the once and future Register of Copyrights, Shira Perlmutter, engineered the usual release of a pre-publication draft of the final section of the Copyright Office’s massive study on the relationship between copyright law and artificial intelligence. The next morning she was fired, in a one-sentence email, at the direction of so-called president Donald Trump. Perlmutter knew at the time of the hurried release that her time in her position would soon be up. Her boss, who had appointed her, Librarian of Congress Dr. Carla Hayden, had been sacked the day before, evidently for the sin of having been appointed by President Barack Obama, as well as being an accomplished woman and African-American—the trifecta of Trump animus—and Perlmutter was guilty by association.

Perlmutter also knew that elements of the report were not in tune with Trump’s unhusbanded AI boosterism and had reason to fear the final section of the report would be suppressed once she was gone. So she made sure it wouldn’t be.

Once fired, Perlmutter sued the government claiming that, as an appointee of the Librarian and therefore a Congressional officer, the president did not have the authority to fire her. The U.S. District Court for the District of Columbia denied Perlmutter’s request for an injunction that would have allowed her to remain in her post. She then appealed to the DC Circuit Court of Appeals, which found her firing to be “likely unlawful” and granted the injunction pending the outcome of her appeal of the district court’s ruling.

Last week, the Supreme Court denied the government’s application for a stay of the circuit court’s injunction, effectively allowing Perlmutter to remain as Register of Copyright’s, at least for now.

With the preliminaries out of the way, the courts can now turn to the underlying legal question of whether the president has the authority to fire the Register of Copyrights. In what could be an extensive meantime, however, Perlmutter looks secure in her post, raising the possibility that the Copyright Office might formally publish Part 3 of its report on AI and copyright.

Part 3, the final section of the report, was the most eagerly anticipated because it deals with the most legally and politically charged question regarding the use of copyrighted works in the training of generative AI models (the first two sections, on digital replicas and the copyrightability of AI-generated works, were published in July 2024 and January 2025, respectively). As I wrote at the time of the pre-publication release, the draft version concluded that the unlicensed use of copyrighted works to train generative AI models, in many if not most cases, cannot be excused as fair use.

That conclusion is very much at odds with the position taken by the technology companies fending off claims of copyright infringement in court brought by a wide swath of creators and rights holders. So, too, is it at odds with the Trump administration’s pro-AI agenda, which may have contributed to Perlmutter’s ostensible defenestration.

The opinions of the Copyright Office, by themselves, do not carry the force of law. But courts, historically, have been deferential to the office’s interpretation of the law it administers.

Because it was never formally published, however, the final section of the AI report has had limited impact in court. The two courts that have issued final opinions in AI-related infringement cases, in fact, ruled that unlicensed use of copyrighted works in training qualified as transformative fair use, although one also held that the means by which the works were obtained could still be infringing. That led to the largest ever copyright settlement, with Anthropic agreement to pay $1.5 billion to the class-action author plaintiffs in Bartz v. Anthropic.

The court’s analysis, at least, comported with the draft report, even if the opinion did not reference it.

“The steps required to produce a training dataset containing copyrighted works clearly implicate the right of reproduction,” the report said. “Developers make multiple copies of works by downloading them; transferring them across storage mediums; converting them to different formats; and creating modified versions or including them in filtered subsets. In many cases, the first step is downloading data from publicly available locations, but whatever the source, copies are made—often repeatedly.”

In Bartz, the court distinguished Anthropic’s downloading of publicly available libraries of pirated e-books from their use in training, finding the latter to be fair use but not the former.

Formal publication of Part 3 of the report could make its conclusions even more persuasive with courts, providing new ammunition for AI-copyright plaintiffs. That could prove particularly telling with respect to the question of market harm, frequently the most dispositive of the four statutory factors in fair use analyses.

Traditionally, courts have confined their analysis of the fourth fair use factor to direct, cognizable harms to a plaintiff, such as lost sales, lost licensing opportunities, or the potential for direct substitution of an original work by the contested one. Those generally have proved difficult for plaintiffs to establish regarding the use of their works in training generative AI models where the harm is more stochastic than empirical.

In the draft section of the report, however, the Copyright Office seemed to expand the parameters of market harm to include the concepts of market dilution and stylistic imitation even where an AI output is not substantially similar—another critical benchmark in a fair use analysis—to a plaintiff’s work.

“While we acknowledge this is uncharted territory, in the Office’s view, the fourth factor should not be read so narrowly,” the report said. “The statute on its face encompasses any ‘effect’ upon the potential market. The speed and scale at which AI systems generate content pose a serious risk of diluting markets for works of the same kind as in their training data. That means more competition for sales of an author’s works and more difficulty for audiences in finding them. If thousands of AI-generated romance novels are put on the market, fewer of the human-authored romance novels that the AI was trained on are likely to be sold.”

If courts were to embrace that view, AI companies could find their fair use shield showing cracks.

Share

Subscribe now

Google is out this week with a set of proposals for AI governance that includes a number of recommendations that, wait for it…would be good for Google. Titled “A Pragmatic Approach to AI Governance in America,” the 21-page outline touches on six areas of concern involving so-called “frontier” AI models: the future of work; protection of minors; information integrity; copyright and creativity; energy infrastructure; and privacy. The pragmatism is in the eye of the beholder.

The proposal takes its biggest swing at the question of who should regulate frontier AI models. Google’s answer: an independent self-regulatory body under the supervision of an unspecified federal agency. It calls for creating a federally supervised frontier AI regulatory organization (FARO) that would “standardize and verify frontier AI labs’ safety and security practices prior to launching frontier models on the U.S. market, including by ensuring that companies are publishing and complying with frontier AI frameworks and engaging with annual governance process audits.”

The FARO board, according to the proposal, would include “a combination of independent directors and industry representatives, balancing considerations of safety and security against the need for bold innovation and speed to market in a global race.”

Google likens the structure to the Financial Industry Regulatory Authority (FINRA), a self-regulatory body comprised of financial broker/dealer members overseen by the Securities and Exchange Commission (SEC), or state bar associations that are overseen by state supreme courts.

Unmentioned by Google is that FINRA, at least, was established by act of Congress and given explicit authority to enforce SEC regulations over its members. The FARO, as described in the proposal, would be more of a standards-setting organization than an enforcement agency.

Working in concert with federal government experts in different safety domains, the FARO should maintain a repository of standards and benchmarks for building, testing, and deploying the most advanced AI systems.

It could also establish other safety rules based on industry best practices, guidelines from industry bodies like the Frontier Model Forum, and standards from the American National Standards Institute and the International Standards Organization. Ideally, governments around the world would implement international reciprocity agreements with the U.S. government, such that models verified by the FARO would be automatically verified in reciprocal jurisdictions.

Google’s other big idea is to explore “new types of value-exchange” with publishers—so long as it doesn’t involve mandatory licensing of publishers’ content for training AI models.

Using publicly available web data for training models is a transformative, non-expressive use — like an art student taking inspiration from walking through a gallery — that should remain protected under fair use in the U.S. and text-and-data-mining exceptions abroad. While this use is legally protected, responsible model developers should give website owners choice and control over whether their content is used in model development via simple, machine-readable robots.txt tags, like the Google-Extended control. Such controls enable markets to work and allocate value where appropriate.

“Where appropriate” is doing a lot of work in that paragraph. It also elides the fact that robots.txt is a voluntary protocol with no technical enforcement mechanism to back up “website owners choice and control” over the use of their content. It’s up to the bots whether to honor robots.txt tags.

As for any further consideration of copyright, “the focus should again be on outputs — in this case, whether a specific image or piece of text actually copies an existing work, regardless of how it was created,” the proposal says.

It’s hard to know who the intended audience is for the proposal. The copyright portion simply restates a position Google and other tech companies have long held and advocated for, both publicly and in court.

The FARO proposal seems intended to imply a concession on Google’s part to the need for regulation of AI. But substantively, it seems more like a bid to preempt state AI regulations without actually creating formal federal regulations beyond industry-developed standards and “best practices.”

To be fair, there is something to be said for involving outside expertise in crafting any federal standards regarding highly technical and fast-moving developments such as AI. That’s especially so now, after Donald Trump and Elon Musk have effectively denuded the federal government of expertise of any kind in favor of political hackery or simply no governance at all.

Beyond that, however, it seems there’s less here than meets the eye.

Share

Subscribe now

The Senate Judiciary Committee recently approved the NO FAKES Act unanimously, sending it to the full Senate for consideration. While no floor vote has yet been scheduled, the bill’s passage through committee, by unopposed voice vote, takes us one big stop closer to creating a new species of intellectual property in the U.S.

The Nurture Originals, Foster Art, and Keep Entertainment Safe of 2026, a revised version of NO FAKES Act of 2025, grants individuals an intellectual property right in their voice and visual likeness, and would prohibit the distribution of unauthorized AI-generated deepfakes or offering a product or service primarily designed to create them. According to the bill’s text, “each individual or right holder shall have the right to authorize the use of the voice or visual likeness of the individual—(A) in a digital replica; or (B) in connection with a product or service for which authorization of the individual or right holder is required to avoid liability“.

The right is not assignable to a third party during the life of the individual, but is licensable, “in whole or in part, exclusively or non-exclusively, by the right holder.” Nor does it expire with the death of the individual but extends for 10 years after death, renewable in five-year chunks if the voice or likeness is being used in interstate or foreign commerce, up to a maximum of 70 years post-mortem.

“This bill is about protecting what’s most personal to us, what makes us us: our voice and our likeness,” Sen. Chris Coons (D-DE), one of the bill’s sponsors, said. “Every American should have the right to decide how their voice and likeness will be used, especially when generative AI can create highly realistic digital replicas that make it look like we said things we never said and did things we never did.”

The bill now includes exemptions for parody, news and documentaries, in addition to libraries and archives, in a bid to quell the concerns that dogged previous versions of the legislation over its potential to trample on the First Amendment’s guarantee of free speech. Although, those concerns still loomed over the committee vote.

“The legislation as drafted now raises some potentially significant concerns regarding free speech,” Sen. Mike Lee (R-UT) said before voting to advance the measure. “We do need to ensure that in protecting content creators’ rights, we don’t inadvertently chill free speech or undermine long-standing First Amendment principles.”

The panel also adopted an amendment to limit liability in cases where a deepfake bore a coincidental resemblance to a non-famous person.

Not surprisingly, the committee’s action was cheered by the creative industries.

“The Human Artistry Campaign applauds the Senate Judiciary Committee for advancing the NO FAKES Act.,” the coalition said in a statement emailed to reporters. “Creativity is rooted in human experience – perspectives, faces and stories that connect us and move culture forward. As AI evolves, everyone deserves the right to control how their voice, likeness and identity are used. The NO FAKES Act provides those protections without inhibiting innovation, and Congress must act now for the benefit of each American.”

Added the Recording Industry Association of America, “As more policymakers join, a broadening coalition of creators, advocacy groups and technology stakeholders is driving nationwide protection for every American to maintain control over their own voice and likeness from exploitive unauthorized AI-generated voice and image clones while also encouraging innovation and protecting free speech, creating clear momentum for Congressional passage this year.”

I wasn’t able to find a formal statement from the IP bar but I imagine they’re pretty pumped as well. If it becomes law, the NO FAKES Act will usher in a whole new flavor of intellectual property—a virgin field of law waiting to go under the plow of dealmaking and litigation.

Clearly, AI-enabled deepfakes are a scourge on modern life, not just for celebrities, politicians and other high-profile victims, but for ordinary folks who have fallen victim to deepfake scams: the grandparent deceived into thinking they’re speaking by phone with a relative in trouble; the victims, often underage, of deepfake porn; the middle manager duped by their “CFO” on a video call into transferring money to a fake account. Policymakers and regulators have a clear obligation to take meaningful action to try to limit the damage—to say nothing of the companies that unleash the tools for creating deepfakes.

But it’s no trivial matter to conjure an entirely new form of intellectual property out of something as commonplace—and for the most part as unauthored—as faces and voices.

Historically, and in the U.S. constitutionally, intellectual property rights have been reserved for the fruits of human creativity, whether inventorship or artistry, or for the distinctive indicia of a brand. But there is nothing particularly creative about having a face or a voice. Most of us have them. The enumerated intellectual property power in the Constitution is also quite explicit as to its purpose: to promote the progress and useful arts. It would be difficult to argue that a person’s likeness or voice are are the products of science or art.

There are, of course, creative uses to which a human face or voice can be put. But that’s not what the NO FAKES Act is about. The products of creative uses of a voice or likeness, if fixed in a tangible medium of expression, are already protected by copyright, or if performed live, by Ticketmaster (sorry about that).

Activity that trades on the notability of a likeness or voice without authorization should certainly be proscribed. But that can be done without creating a new property right, with all that entails. We sanction a lot of unauthorized uses of things without bestowing new property rights. It’s illegal to make unauthorized use of financial services to launder money, or evade trade restrictions, but not because it infringes anyone’s property rights. There are other plausible legal authorities on which a prohibition on deepfakes of prominent persons could be based—the interstate commerce clause, perhaps, or anti-fraud statutes, identity theft laws or computer crime statutes. It might even make sense to establish a national right-of-publicity law that would not necessarily require the administrative infrastructure of property laws. But the NO FAKES Act doesn’t do that.

Administratively, in fact, the NO FAKES Act is a bit of an oddball. Conceptually, the new IP right envisioned by the bill is most similar to trademarks. It imagines the likeness of a person, or the sound of their voice, as distinctive indicia of that person’s being. But the law would assign administrative responsibility to the Register of Copyrights, not the Patent & Trademark Office. And it charges the Register with developing and supervising procedures for registering to extend the post-mortem term of the right.

The law also envisions a DMCA-like notice-and-takedown regimen and a provision for counter notices, ironically a system many copyright owners find frustrating and ineffective.

But the real potential for confusion lies in the would-be law’s application in cases involving individual’s who have not otherwise made commercial or notable use of their likeness or voice. Certainly, the malicious use of another person’s identity ought to be a crime. So, too, should providing tools predominantly designed to make such use. But investing an intellectual property right in something of no commercial value, even to the possessor, as the foundation for prohibiting such activity seems like a long way around and could even be counterproductive.

You don’t need specially designed tools to create deepfakes. Midjourney or Gemini will do just fine. And, insofar as the new right has at least one foot in the Copyright Office, it could lead courts to apply the substantial non-infringing uses standard in copyright law to decline to impose liability on general-purpose AI tools.

It’s perfectly understandable that those in the creative industries who make their living from their likeness or their voice would desire legal tools to wield against their misuse. It might even be justifiable as a form of intellectual property in those cases. Their appearance and the sound of their voices might plausibly be construed as intangible assets that have value due to the creative efforts of those individuals.

The problem, of course, is where do you draw the line? How creative or notable would a voice or likeness need to be to earn protection? The NO FAKES Act attempts to route around the problem of line drawing by extending the right to everyone—the notable and anonymous alike. But is it still intellectual property if its application is so indiscriminate.

It’s perfectly understandable that Congress should seek to protect the rest of us anonymous, no-talent schlubs from fraud and abuse. But investing every face in the crowd with a new property right steers a route into unchartered waters.

This is the part where it’s incumbent on me to suggest a workable alternative. And I freely admit I do not (yet) have one. But with more debate likely to come before the NO FAKES Act reaches its final form I hope someone smarter than me can figure out how to tailor its solution to the nature of the problem without creating new problems in its wake.

Share

Subscribe now

The U.K.-based news-publisher coalition SPUR last week unveiled what it describes as a “common language” for tracking the use of content by AI systems. It also took the wraps off a draft “signal format for content usage reporting.”

The goal, according to the announcement, is to enable publishers to assess the value of their works within an AI environment, and to eventually be able to license and price their content at a more granular level than merely yes/no.

“Tracking how AI systems use content creates benefits for everyone involved. A common standard means AI systems and their users benefit from better, more relevant outputs,” Financial Times chief executive Jon Slade, who is leading the effort, said in the press release. “For content creators, seeing how and where their work brings value within the AI environment helps them understand where to focus energy and resources for that audience.”

SPUR was launched in late February and included the BBC, Guardian, Sky News and Telegraph Media in addition to the Financial Times. Since then, roughly 20 other publishers have joined the coalition.

The common language defines five stages of news content’s journey through a generative AI model:

• Retrieval from a website, which content owners can currently track based on bot activity;

• Grounding, referring to when content is being loaded into an AI system’s generation machinery;

• Citing, where content is being explicitly referenced in an AI-generated response;

• Display, where a user actually sees the content, either via reference or it being embedded in a generated answer; and

• Engagement, when a user clicks on, copies, shares or directs an AI agent to act on the content.

The next, and far harder step, will be developing a technical standard to implement the taxonomy at the machine-to-machine level. Especially so in that SPUR, for now at least, remains a publisher-only forum. There currently are no technology company members of the coalition, although last week’s announcement did include publication of the SPUR Content Telemetry Profile spelling out the proposed conditions a technology company must meet to be considered compliant with the standard.

Still, the taxonomy offers a potentially useful framing of a fundamental challenge rights owners of all kinds face in reckoning with generative AI—the problem of gauging the value of their content inside an enormous black box.

From a macro perspective, the value to an AI model of any particular work is negligible to the point of unmeasurable. When a model is being trained on hundreds of billions of individual works, and the training is based on statistical associations within the total dataset, excluding any one work—or even hundreds or thousands of works—is not going to materially affect the weights within the finished model.

Certain types of training inputs—substantial archives of high-quality, structured data—can affect the quality and reliability of a model, at least at the margins, and rights owners sitting on those sorts of archives have been able command fees for their use in training. But those fees are based on an imputed aggregate value of the archive rather than on any consistent metric.

Measuring the value of any one work to a particular generated output faces a different problem—that of attribution. Isolating and weighing the ingredients that go into an output effectively is impossible, for reasons both of the scale of the model and its inputs, and because the ingredients are so transformed in the training process that reliable identification is no longer feasible.

There are so-called attribution engines their developers claim can reverse engineer an output to yield a probability score for whether a known piece of content or collection of content was a source for a generated work, such as the Sureel.ai tool acquired by Warner Music Group last week. But those are based on a probabilistic analysis of a probabilistic event, which is not easily reduced to a deterministic metric.

The SPUR taxonomy attempts to route around that problem by breaking down the generative process—from the model to the end user—into a sequence of discrete events that, in theory at least, can be detected and measured in real time if (still a big “if”) the necessary technical engineering proves feasible. Its object, moreover, is to measure value as it’s being created rather than imputed before hand or estimated after the fact.

A few caveats: the SPUR model is use-case specific, in this instance the scraping of news content for use in generating a summary of the original or a response to a particular search query. Its goal is to put a hard number on how publishers’ content is being used in a generative AI environment, much as Google Analytics let publishers put a number on how the audience is using their content online.

It would not be useful for standing up a licensing model for including content in an initial training dataset before any output is generated.

Nor is it obviously transferable to other types of media that may be used very differently by AI models and AI users. It is a bespoke solution to a particular problem.

That said, it’s an interesting attempt to restore a measure of information symmetry to an industry that had it stripped away by generative AI.

Share

Subscribe now

In another uncanny echo of the Luddite era in today’s anti-data center protests, Wired reported this week on a tranche of leaked documents from the Department of Homeland Security and the FBI discussing a new domestic threat: anti-technology extremists.

Among the trove is a report prepared by the New York Intelligence and Counterterrorism Bureau warning of the potential for widespread upheaval in response to AI adoption. "The chaotic atmosphere that may result from emergent AI technology in the next five years may fuel large-scale protests that devolve into civil unrest and anti-tech violent extremist activity, especially in large urban areas such as New York City," the report reads.

The Luddite movement ultimately was undone by a combination of state surveillance and state suppression in response to Luddites’ own proclivities for what might be described as “anti-tech violent extremist activity,” in their case, smashing up textile machinery. In the documents obtained by Wired, we see the beginnings of state surveillance of anti-data center protestors.

According to Wired, the concerns discussed in the documents likely relate to President Trump’s September 2025 National Security Memorandum directing the Justice Department to target anyone expressing “anti-American,” “anti-Christian,” or “anti-capitalist” views. Like the British authorities’ response to the Luddite rebellion, that is, the Trump administration views “anti-tech” sentiment as of a piece with a broader threat to the imperial order.

Today’s authorities have even found their own “Ned Lud.” The same New York Intelligence Bureau report describes a new threat emerging in the wake of the arrest and trial in Maryland of Jack LaSota, a transgender woman who goes by “Ziz,” and is purported to be the leader of a group of mostly young computer scientists who go by the name “Zizians,” and who allegedly hold extreme views about veganism, animal rights, gender identity and AI.

Like those acting in the name of Ned Lud, the Zizians pose a genuine a threat of violence. Members of the group have been connected to the death of a California landlord, the shooting death of one of the member’s parents in Pennsylvania and a highway shootout in Vermont that led to the deaths of a border agent as well as a member of the group. Despite the apparent connections, no member has yet been charged in any of the deaths.

Yet, while “Ned Lud” was an apocryphal figure, conjured to strike fear in mill owners and the authorities, “Ziz” may simply be crazy. In March, the judge overseeing her trial ordered her held in custody while a mental competency exam is performed. The self-described “Zizians” appear to be closer to a violent cult than to a political or popular movement—Manson Family than Montgomery bus boycott.

Most of the anti-data center protests, of course, have nothing to do with Ziz or the Zizians, although there have been sporadic incidents of violence. An Indianapolis city councilman who supported a local data center had 13 rounds fired at his house, and OpenAI CEO Sam Altman had a Molotov cocktail thrown at his home. Threats of violence against individuals, policymakers and AI companies have proliferated online, along with threats of physical sabotage of data centers already in operation.

The threat of political violence is unfortunately very real in the U.S. right now. And anti-AI sentiment, focused on data centers, appears to be intensifying, not abating. So, the possibility of more anti-AI related violence can’t be discounted.

And if history is anything to go by, the specter of “anti-tech violent extremist activity,” will be invoked to justify increasingly invasive and forceful countermeasures.

Share

Subscribe now

On the afternoon of March 11, 1811, in Nottingham England, British troops broke up a group of protesting mill workers demanding more work and better wages. That night, angry workers smashed up textile machinery in a mill in a nearby village. The attack sparked a movement.

Soon, similar attacks began to occur throughout the textile centers of the English Midlands and north into Yorkshire, first happening nightly and then in sporadic waves. Efforts to suppress them, both by mill owners and the crown, grew increasingly violent as the attacks spread. One of the bloodiest incidents occurred in April 1812, when a large group of protestors stormed a mill near Manchester. The owner of the mill ordered his men to fire on the protestors, killing three and wounding 18. British soldiers called in the quell the unrest killed five more the next day.

The attacks nonetheless kept up, on and off, for another five years, smashing machinery and even setting some mills ablaze. Mill owners as well as the authorities responded with increasingly forceful tactics to try to suppress the movement. Dozens of protesters ultimately were hanged and others sent to penal colonies in Australia.

Yet for all the panic the attacks engendered, the “movement” was never truly organized and had no clear leader. Much of its mystique—at the time and since—stemmed from the shadowy figure of Ned Ludd, sometimes called Captain Ludd, General Ludd, and even King Ludd. He made is first appearance at an attack in Nottingham in November 1811, and popped up at protests all across the Midlands.

But there was no such person, at least not at that time. He was an apocryphal figure cooked up by the protesters and loosely based on an 18th Century figure named Ned Lud, or perhaps Ludham, who himself may or may not have existed. He was purported to be a young apprentice in a textile mill in Leicester in the 1780s. According to legend, he was working one day at a stocking frame—a kind of mechanical knitting machine—when he was admonished by his boss for knitting too loosely and ordered to “square his needles.” Rather than comply, Ned supposedly grabbed a hammer and smashed the stocking frame to bits.

The 19th Century textile workers invoked his legend and took the name “Luddites.” They sent bellicose letters and posted billboards pseudonymously signed Ned Ludd. So effective was their propaganda that British soldiers sent to put down the protests more than once swore they had seen him, sometimes as a spectral figure.

Today, the term “Luddite” is used mainly as a pejorative, to denote someone opposed to technology and resistant to technological innovation. But the Luddites themselves, by and large, were not anti-technology or anti-innovation. Many, in fact, were skilled machine operators in the mills they attacked. Although they’re often remembered today for smashing the newly developed steam-powered looms then being introduced into mills, giving rise to their later reputation for opposing technological innovation, among the most frequent targets of their attacks were in fact stocking frames, a technology developed during the reign of Queen Elizabeth I more than two centuries earlier.

Instead, their real grievance was with production-obsessed mill owners using machines in what the Luddites called “a fraudulent and deceitful manner” to get around standard labor practices, suppress wages, and churn out cheaper, inferior quality goods to feed new and expanding export-driven markets in the aftermath of the Napoleonic Wars. Their principal demands were for machines that made better quality goods, run by workers who had gone through a traditional apprenticeship and paid a better wage, and an end to child labor.

The steam looms and stocking frames the Luddites destroyed were merely the smashable proxies for a new capitalist, mercantile economic paradigm that characterized the British Industrial Revolution, as both a commercial and political project, and into which they had little input and to which they did not consent. It was an approach to automation that was almost entirely extractive, appropriating value from both inputs and outputs, in which the fruits of the increased productivity it yielded fell disproportionately on one side of the industrial-labor fence.

The intercession of state power in support of that new paradigm only deepened their grievance. Rather than addressing the underlying causes of the unrest, the authorities as much as the mill owner regarded the Luddites as mere vandals and criminals, and responded accordingly.

Today, more than two centuries hence, Luddism is again abroad in the land. The focus of its ire this time is not the steam loom or the stocking frame but the data center. Protests against their construction have broken out nationwide, catching developers and the local officials approving granting their permits and zoning waivers off guard. Like the Luddite attacks of the past, the anti-data center protests have erupted spontaneously in disparate communities, with little if any coordination or clear leadership. According to Data Center Watch, which tracks data center development across the U.S., at least 48 such projects, worth $156 billion, were cancelled or delayed last year due to community opposition.

The Baltimore City Council this month became one or more than 60 local governments to impose a moratorium on data center construction. Last month, the Maine legislature became the first to pass a state-wide moratorium on development of new data centers (although the bill was ultimately vetoed by Gov. Janet Mills). A group of more than 60 organizations, ranging from environmental groups to community organizations to labor unions last week sent a letter to Gov. Mikie Sherrill calling for a state-wide moratorium. At least a dozen other states are considering similar moves. On May 10th, Colleton County South Carolina passed a six month moratorium on data center construction with an option to extend it.

Unlike the Luddite movement, however, anti-data center sentiment is widely shared. According to a recent Gallup poll, seven in 10 Americans are opposed to data center development in the local area, and nearly half (48%) are strongly opposed. The opposition cuts across party lines, regional differences, and generational divides. More Americans support building nuclear power plants in their area than support building data centers.

Much of the opposition stems from the very tangible impacts data centers can have on the communities where they’re located. They use enormous amounts of energy, causing utility prices to spike for consumers and businesses in the area; they have myriad baleful environmental impacts, including on local water quality a producing copious amounts of e-waste; they generate noise at a decibel level and frequency range many find noxious; and they undermine nearby property values.

But the ferocity of the data-center backlash points to grievances that go deeper than higher electric bills. Like the steam looms and stocking frames that took the brunt of the Luddite’s ire, the data centers now being rushed into development around the country, often with little input or consent from the community, have become physical proxies for disruptive new economic and political forces unleased by the AI Revolution.

Around the country this spring commencement speakers have been showered with boos at the mere mention of AI, as graduates face grim employment prospects as AI automates away many entry-level jobs and companies lay off middle level workers by the thousands. A recent New York Times-Siena poll found that more than twice as many voters feel AI is “mostly bad” than feel it is “mostly good.” That same poll found that nearly half (47%) of voters under age 30 have a mostly negative opinion of AI, the highest share of any age group, suggesting the most tech-savvy generation is the most resistant.

Tech companies that have sunk hundreds of billions of dollars into building AI infrastructure have been reduced to forcing AI tools into everything from operating systems to office productivity tools to search bars to create the illusion of adoption in the face of customer resistance. The privacy-focused search engine Duck Duck Go, which promotes its lack of AI features, has seen a 30% surge in installs since Google announced an AI-centric overhaul of its core search platform last week, prompting long-time users to flee.

Just as empire and industrialization were inextricably linked in 19th Century Britain, moreover, the current American regime’s techno-imperialist ambitions have aligned state power with the interests of AI makers. The Trump administration has used tariffs and sanctions to try to coerce European countries into altering their digital rules to make them more accommodating to U.S. technology companies. Domestically, he has sought to use federal power to coerce state and local governments into relaxing or repealing laws regulating AI systems and companies.

No less a figure than Pope Leo XIV made the connection between the two eras explicit in his first encyclical since becoming pontiff. His 42,000-word pastoral treatise, Magnifica Humanitas, denounces the “culture of power” driving the AI race and warns against “opaque algorithms,” controlled by “a few” companies.

In his Encyclical Laudato Si’, Pope Francis denounced the growing dominance of a technocratic paradigm in our globalized world: the tendency to let the logic of efficiency, control and profit alone shape personal, social and economic decisions. This makes it clear that technology is not simply a tool. When it becomes the standard by which everything is judged, it begins to dictate what matters and what can be discarded, reducing creation to an object of exploitation and human beings to mere cogs in a system driven toward ever greater efficiency.

For emphasis, Leo signed his encyclical on May 19th, the 135th anniversary of his namesake Leo XIII’s encyclical Rerum Novarum, which called for fair wages and workers’ rights in response to the Industrial Revolution.

For many in the creative community, of course, the AI revolution has always felt extractive and coercive. AI models extract value from all that they ingest in training, generally with neither compensation or consent of whoever may have created it, and return mediocrity at best, to say nothing of the more pernicious and abusive output of which they’re capable.

The question for creators is whether they can effectively channel the broader cultural anxiety over AI into policies and practices that materially improve their lot.

While the Industrial Revolution did eventually create enormous wealth and lifted living standards broadly, the Luddite project was a failure in its own time, undone by a combination of self-defeating tactics and official repression. It did little immediately to improve the material circumstances of the Luddites themselves or of the average laborer, and the Industrial Revolution rolled on.

Whether the AI Revolution will have the same world-changing impact as its predecessor is yet to be seen. While the productivity gains of the Industrial Revolution were fairly obvious and concrete early on, at least for the industrialists, companies that have adopted AI, by and large, are yet to see meaningful productivity gains from the deployment. According to a recent study by the National Bureau of Economic Research, 89% of companies across four countries surveyed reported no productivity gains from AI over the last three years.

With the initial deployments teetering on the edge of the hype cycle’s Trough of Disillusionment, and the cultural backlash to the AI juggernaut’s coercive and disruptive ethos hardening, the moment may be ripe for meaningful change to the prevailing balance of equities of the AI Revolution, and to avoid the fate of the Luddites.

Share

Subscribe now

When actor Matthew McConaughey filed to register his likeness as a trademark with the U.S. Patent & Trademark Office in January, his lawyers acknowledged it was an untested strategy. U.S. trademark law does not explicitly recognize an individual’s likeness as a protectable mark.

A few celebrities have managed over the years to successfully assert a claim against an unauthorized use of an identifiable feature of their image or persona. In 1988, the Ninth Circuit Court of Appeals ruled that the “distinctive voice of a professional singer” is a protectable element of an individual’s identity under a common law right to control your image in a case brought by Bette Midler against Ford Motor over the use of an imitator in an ad (one of Midler’s own backup singers at the time, in fact).

Two years later, Tom Waits successfully challenged Frito-Lay over a copy-cat use of his signature gravel-throated vocal delivery in an ad, citing both voice misappropriation under California’s right-of-publicity law and false endorsement under the federal Lanham Act.

Other celebrities have succeeded in bringing similar cases under various state right-of-publicity or unfair trade practices law. But for now at least, federal law does not regard a person’s likeness as a form of intellectual property on par with a patent or copyright.

Bills to create such a right have been introduced in both the U.S. House and Senate, most notably the NO FAKES Act, but none have made it to a vote. Objections to the idea run from the fact that your face is not something you produced yourself (generally speaking) and therefore not a creative expression, to potential conflicts with free speech, such as in criticism or parody, to whether and where to draw the line between notable faces and everyone else.

The introduction of powerful AI applications, however, has made it trivially easy to create a digital replica of your own or anyone else’s face or voice and insert their apparent likeness into any scenario, real or imagined. that has altered the quantitative, if not necessarily qualitative terms of the debate, as malicious deepfakes of politicians, celebrities and athletes have poured out of AI-powered image generators.

That was certainly a factor in prompting Taylor Swift to file an application to trademark her likeness and voice last month. As one of the most popular artists on the planet, whose every utterance, outfit, gesture and romantic engagement is minutely scrutinized by the celebrity press and her myriad legions of devoted fans, Swift has been a frequent target of deepfakes. She is also acutely aware of the monetary value of her celebrity as well as her music and attends to it assiduously. If there were ever an ideal test case for trying to trademark a person’s image it is Taylor Swift.

Even Taylor Swift, however, could only enforce her trademark against unauthorized uses as swiftly as the courts move, which is not swift, even for Swift. Given the sheer volume of unauthorized uses AI has unleashed, the case-by-case approach is neither practical nor effective. To be effective in the context of AI, rights need to be asserted, managed and enforced at the network level, on a machine-to-machine basis. That’s the goal of projects like Cloudflare’s managed-robots.txt system launched last year, WIPO’s Artificial Intelligence Infrastructure Interchange initiative and the RSL 1.0 technical standard with respect to copyrights.

Now, group of A-listers from the movie business took are trying to do the same for a person’s likeness, voice, and other personal and personality traits. Co-founded by actress Cate Blanchett, the group, calling itself RSL Media, this week launched a new non-profit that will administer a registry of works, personal identities, characters and protected marks linked to the new RSL Human Consent Standard.

The new standard — technically a module within the existing RSL (Real Simple Licensing) standard — will express “licensing terms that apply to identified Rights Subjects as abstract legal entities, rather than to individual digital assets,” in machine readable form, thereby “enabling consistent interpretation across training, reuse, redistribution, and AI-generated representations,” according to the group’s website.

Part of that interpretation will involve the “Evaluation of Rights Subject licensing terms by AI Systems and, where required, linkage to external authorization, licensing, or Clearance processes through license servers implementing the RSL-MEDIA Open License Protocol.”

“AI technologies are expanding rampantly, essentially unchecked and unregulated. In order for humans to remain in front of these technologies, consent must be the first consideration,” Blanchett said in a statement. “RSL Media is a simple, effective and free solutions-based technology for facilitating and activating consent. It’s also the industry’s first practical solution where people everywhere, not just public figures, can assert control over how their work is used by AI.”

Others involved at launch comprise a Mt. Rushmore of film and TV performers, including Javier Bardem, George Clooney, Viola Davis, Tom Hanks, Dame Helen Mirren, Steven Soderbergh, Kristen Stewart, Meryl Streep and Dame Emma Thompson, along with the Creative Artists Agency and the Music Artists Coalition.

Though new, RSL Media is not the first effort to create such a standard. The HAND ID standard, developed by HAND (Human and Digital) launched in 2024 and was recognized as an International Standards Organization (ISO) technical standard the following year. HAND ID is a digital object identifier (DOI) associated with a registry of legal-entity humans, their human likenesses and their animated characters for those seeking to license those assets for use in cultural products.

RSL Media is on the right track, by associating an individual with a machine-readable signal. The challenge it will face is compliance. Whereas HAND is designed for those figures seeking to proactively license their image and likeness, such as athletes with contractual rights to their name, image and likeness (NIL), RSL Media is designed to combat unauthorized uses.

RSL works in conjunction with robots.txt, which is a voluntary protocol that some bots honor and some bots don’t. As part of the RSL Consortium, Cloudflare and its managed-robots platform puts some teeth into robots by controlling access to websites at the network level. But at least in that case the right being asserted is unambiguously a legally defined and protected intellectual property right: copyright.

RSL Media faces the inverse of Taylor Swift’s dilemma. Even if Swift is granted a trademark, trademarks are not self-enforcing in a digital environment. Without a machine-readable mechanism to assert and enforce her trademark she is at the mercy of the legal system to protect it. RSL Media has developed a machine-readable mechanism for expressing rights in a person’s likeness, but that purported right is nowhere to be found in U.S. intellectual property law.

Share

Subscribe now

Image CC

Most economists would tell you that if you want less of something, tax it. In classic economic theory, taxes, no matter how justified on other grounds, act as a disincentive to produce. So, in the aggregate, you’ll get less overall production.

Some commentators have lately proposed applying that principle to AI “slop.” In a widely read op-ed in the Guardian last month, technologist Mike Pepi, author of Against Platforms: Surviving Digital Utopia, proposed “a minuscule tax levied on the largest AI companies,” and using the money to fund cultural institutions, artists and researchers. “As it stands, AI slop is a malicious manipulation of human cognitive labor and the institutions that support it – something akin to a cognitive pollutant,” he wrote. “A ‘slop tax’ would ensure robust institutional support structures for human creativity forced to compete in a sea of meaningless content.”

The idea, and Pepi’s essay, have lately gained considerable traction on Reddit, suggesting he stuck a chord, at least among the type of folks who spend a lot of time debating things online.

There’s no doubt that AI slop has become the scourge of many in the creative industries. French music streamer Deezer recently reported that roughly 75,000 AI-generated tracks per day are uploaded to its platform, some 44% of all the songs uploaded, while capturing less than 3% of total streams. Amazon is awash in AI-generated e-books and the Podcast Index recently reported that fewer than half the new shows released on various platforms within the previous 24 hours were identifiably produced by human podcasters. Nearly all the rest appeared to be AI-generated, for which the industry has coined the term, “podslop” (about 10% could not be firmly classified).

The proposal to tax slop comes amid a larger debate around taxing AI more broadly as a means to offset its expected impact on the labor market, among other economic disruptions. Proponents of the idea also note that experts expect that AI will increase returns on capital over time while reducing returns from labor. In the U.S., which derives 85% of its income tax revenue from labor, that shift would greatly reduce overall revenue. No less a proponent of AI than OpenAI last month released a 13-page “Industrial Policy for the Intelligence Age” paper calling for “modernizing the tax base” to account for the changing composition of economic activity being wrought by the technology from mostly labor-based to mostly capital-based.

Critics of the idea of taxing AI outputs argue the idea is ultimately unworkable due to a basic definitional problem: what distinguishes “good” content from “bad” content, and who should do the distinguishing? It’s the same problem as with content moderation on social media platforms: what gets filtered and what gets through, and who decides?

There is another approach that might be more workable, however. As I discussed in my previous post, AI is good at nothing so much as generating negative externalities. In economics, a negative externality is one where parties that did not content to, or profit from, a transaction nonetheless end up shouldering some of the cost. Air pollution and carbon emissions from power generation and other industrial activity are classic examples. The market for power generation does not internalize a price on the negative downstream effects of that pollution, so the costs fall on humans, plants and animals that have to breathe the air in the form of harms to human health and climate change.

Rather than trying to define the specific outputs to tax in the case of AI slop—rather than trying to distinguish between “good” outputs and “sloppy” outputs—tax policy could treat the sheer volume of outputs as the externality and impose a levy on producers according to the volume of what they produce. It’s a concept developed by the British economist Arthur Cecil Pigou in the 1920s in his book The Economics of Welfare. In it, Pigou proposed imposing a tax on producers to deal with the problem of over-production.

The idea in the case of AI slop is that if generating 75,000 low-value music tracks cost a few dollars more, and the extra cost increased in proportion to the volume, there would be less aggregate incentive to generate them, and you would end up with less low-value production.

There would still be questions about who should pay how much of the tax. Should OpenAI be taxed on all the volume produced by all the applications and integrations that used ChatGPT, or should the tax burden be more evenly shared? But those are basically political questions that could be debated and decided just as income tax policy is now, rather than questions of taste or cultural preference around “good” content and “bad” content that are inherently fraught.

Moreover, in an attention-based competitive context such as the media and entertainment industries, volume equates to market power. Netflix spends tens of billions of dollars to churn out content because it wants to capture and hold the attention of as much of the video streaming audience as possible, and generates revenue by monetizing that attention. To whatever extent AI slop captures the audience’s attention it detracts from human producers competing for the attention of the same audience, at an artificially lower marginal cost, suppressing returns on human production.

Putting a price on the production of slop could help rebalance those scales.

Share

Subscribe now

Back in January, the Nieman Foundation’s Nieman Lab reported that the New York Times, The Guardian, USA Today and a few other news organizations have taken steps to prevent the Internet Archive’s Wayback Machine from scraping their sites. Since then, additional reporting has uncovered other publishers blocking the archive’s crawlers. Tom’s Hardware counted at least 23 as of mid-April.

The news organizations’ stated reason for shutting out the Archive’s crawlers is to prevent AI companies from scraping their content from the Wayback Machine to use as model-training fodder.

“We believe in the value of The New York Times’s human-led journalism and always want to ensure that our IP is being accessed and used lawfully,” a Times spokesperson told Nieman Lab. “We are blocking the Internet Archive’s bot from accessing the Times because the Wayback Machine provides unfettered access to Times content — including by AI companies — without authorization.”

It’s a perverse policy. The Internet Archive has been collecting, archiving and preserving digital content for some 30-odd years. The Wayback Machine is said to hold more than a trillion web pages and remains the only unbroken chain of custody of that content over that time. It is also an invaluable resource for researchers, historians and journalists, including those employed by the outlets that are now blocking the Archive’s crawlers.

Just a few weeks ago, for instance, USA Today published a sensational investigative report on efforts by Immigration and Customs Enforcement—ICE—to withhold information on its scandalously inhumane detention policies. The piece itself acknowledged relying on the Wayback Machine to compare “every spreadsheet available” in the archive back to 2021 with current data on ICE’s website, “to show how things are changing under the Trump administration and beyond.”

Archives, like libraries and museums, are an essential pillar of the information commons on which all human discourse relies. Preserving that commons and making it accessible is an essential civic function. That’s especially true with respect to information that exists only in incorporeal, technologically-dependent digital formats. Linguists and archeologists can still decipher information stored on 5,000-year old cuneiform tablets because the hardened clay tablets themselves turned out to be a remarkably durable medium. But information created in digital form mere decades ago is in constant danger of being lost due to its ephemeral, fungible nature or because the devices needed to interpret the particular formatted arrangement of ones and zeros are no longer available or functional.

There is something fundamentally incongruous about organizations whose raison d’être is enhancing the commons by gathering information, including otherwise undisclosed or inaccessible information, and making it accessible then intentionally preventing the collection and archiving of what they publish. A newspaper of record declining to leave a record.

That’s not to say the actions of the publishers are incomprehensible. Generative AI systems’ participation in the information commons is almost entirely a one-way street. They gather the organic fruits of the commons on an immense scale and process it into synthetic, inorganic simulacrums that add bulk but not much enrichment. They are the products not of intention or purpose but of blind probability and used parts, and they can crowd out the genuine articles. There’s a logic, of sorts, to organizations that participate fully in the information commons not wanting to participate in the one-way traffic of AI, however self-negating the outcome.

It’s an example, though, of how generative AI’s disruptive impact can also fall on innocent bystanders, in this case the Wayback Machine. Although the Internet Archive has had its differences with copyright owners on other issues, it’s not a direct party to the squabble between news organizations and AI companies. Yet, it’s become the battleground on which the parties are fighting, and the net result is the impoverishing of the information commons by eroding its only comprehensive digital archive.

The Wayback Machine is hardly the only bystander to be caught in AI’s crossfire, however. Many local communities playing host to AI data centers see themselves as unwilling parties to shouldering the costs of the centers through higher utility prices and environmental impacts. But the boom in data center construction is also driving up prices in sectors of the economy that are not directly involved in AI at all.

As AI data centers gobble up all manner of server components, the prices of RAM, graphics cards, hard drives and SSDs, fiber optic cables and more have all skyrocketed, impacting all manner of consumer devices that also rely on those components. Earlier this year, then Apple CEO Tim Cook acknowledged having difficultly sourcing memory components for iPhones due to the high demand for the chips for AI data centers.

A 2 terabyte external SSD that cost around $160 a year ago, is currently listed at $575 on B&H Photo’s website. Samsung, a leading provider of RAM, just this week warned that the current shortage of memory chips is likely to continue well into next year. Meta this week announced it will run some of its non-AI servers longer than in the past due to the shortage of memory chips for replacements.

The great irony of the current AI boom is that, while spending on AI is buoying the overall economy by masking the worst effects of President Trump’s destructive tariffs, it is also bleeding a growing number of sectors of the economy.

As the legal and cultural debate over AI’s use and abuse of copyrighted works continues to rage, it’s worth bearing in mind that its happening within a broader context, and that the technology’s disruptive impact is being felt well beyond the creative industries.

Share

Subscribe now

The repercussions from the Supreme Court’s unanimous decision in Cox Communications v. Sony Music are rumbling through copyright cases in courtrooms around the country. SCOTUS itself quickly wiped the piracy liability verdict against Grande Communications in a case brought by Warner and Sony Music, sending it back to the Fifth Circuit for reconsideration in under the new liability standard. Stream-ripper Yout.com seized on the ruling in a case brought against it by the RIAA. Altice and the the record labels suing the ISP mutually asked the court hearing their case for more time to file status briefs in light of the Cox ruling. And Meta is considering embracing Cox to fend off the torrenting charge still pending against it in the Kadrey case.

Some of those efforts might be a stretch. But none of them is surprising. You would expect any lawyer worth their billable hours to at least try to capitalize on the new state of play. Especially so as the majority opinion in Cox penned by Justice Clarence Thomas (there were two concurrences) appears to have meaningfully narrowed the definition of “contributory infringement” by service providers in copyright cases, holding that mere knowledge that a service was being used for piracy is not sufficient to establish liability.

The fallout may not stop there, however. The ruling seemingly has reignited interest on Capitol Hill in site-blocking legislation. Bills ostensibly targeting foreign piracy sites have been kicking around in both the House and the Senate for a while without really going anywhere. In the wake of the Cox rulings, however, Rep. Zoe Lofgren (D-CA) and Sen. Thom Tillis (R-NC), sponsors respectively of separate House and Senate site-blocking bills, reportedly have begun working together to draft unified, bi-cameral legislation that would require ISPs and DNS resolvers to block foreign piracy when ordered to by a court.

The inclusion of DNS resolvers is significant. Those like me who have been at this awhile may recall the battles over the Stop Online Piracy Act (SOPA) in the House and its companion Senate bill, the Protect IP Act (PIPA). Those two bills, largely cooked up by the major studios and record labels, pioneered the idea of enlisting DNS resolvers in the battle against alleged piracy, primarily as a way to route around a series of unfavorable court decisions in cases brought against online platforms and ISPs. If you can’t beat ‘em, block ‘em.

The two bills each enjoyed broad bipartisan support when introduced in 2011. But they derailed in 2012 amid widespread (and well-funded) protests that included a coordinated day-long black out of hundreds of websites to highlight opponents’ claim that messing with the DNS system to block access to websites would “break the internet.”

Ironically (if predictably), the enshittification of the internet ensued anyway in the decade and a half since, such that mandatory DNS blocking perhaps no longer seems like a third rail to policymakers. The Lofgren-Tillis bill, in fact, resembles regulatory regimes already in place in Europe, where, for instance, the Paris Judicial Court last month ordered ISPs, VPN providers and DNS resolvers, including Cloudflare and Google, to block access to 35 illegal sports-streaming sites in a series of simultaneous rulings. Like SOPA and PIPA, including VPNs and DNS resolvers in the order was intended to close off the most popular routes for circumventing ISP-level blocks. The Cloudflare order even covers its CDN and reverse proxy service.

Nor is France the only European Union country where courts can issue such orders. Jurisdictions throughout the 27-nation bloc have site-blocking regulations targeting various levels of the web’s tech-stack.

That’s not to say that opposition to the measures has disappeared. A recent report by the Center for European Policy Studies (CEPS), which is not affiliated with the EU, reviewed the impact of site-blocking policies across all 27 countries and found that it carries a substantial risk of unintended consequences and negative side effects, including the blocking of many legitimate sites and services along with alleged pirates.

In one incident, for instance, a site-blocking order targeting DNS resolvers in Belgium led at least one such provider to leave the country.

“The episode illustrates how overreaching court orders can have unintended consequences for the broader digital ecosystem, and how disproportionate liability exposure can sometimes incentivise service provider withdrawal rather than compliance,” the report said.

It’s the same argument opponents of SOPA and PIPA made 15 years ago.

Share

Subscribe now

The White House last week released its long-awaited framework for developing a national policy on AI. You could be forgiven for wondering what took so long. What was released was a bare four-page outline sketching out, in very broad strokes, what the Trump administration would like Congress to do—and not do—in crafting specific federal legislation to govern the technology.

The release came the day after Senator Marsha Blackburn (R-Tenn.) plopped her own, massive oar in the water, with a 300-page “discussion draft” of a comprehensive federal regulatory scheme, including specific proposed legislative language. Obviously anticipating the release of the White House proposal, which had been telegraphed for a week, and not wanting to provoke the king, Blackburn shrewdly, if transparently, titled her draft the “TRUMP AMERICA AI Act.” Bold.

The two proposals actually do share many points in common. Both want to include protections for children using chatbots, both would preempt many existing state AI regulations while permitting states and localities some measure of discretion over data-center construction, utility costs and enforcement of consumer protection laws on AI systems. And both nod to the canard of censorship of conservative voices online by prohibiting viewpoint “discrimination” by chatbots, and by sunsetting Section 230 of the Communications Act.

Where the two notably part ways concerns their treatment of intellectual property. Whereas Blackburn urges Congress to clarify through legislation that the unauthorized use of copyrighted works to train AI models is not fair use under the Copyright Act, the White House wants Congress to stay out of it.

“Although the Administration believes that training of AI models on copyrighted material does not violate copyright laws, it acknowledges arguments to the contrary exist and therefore supports allowing the Courts to resolve this issue,” the White House outline says. “Similarly, Congress should not take any actions that would impact the judiciary’s resolution of whether training on copyrighted material constitutes fair use.”

The important point of divergence is not so much in the specific positions the two proposal stake out, but in the split over whether the question of fair use in AI training is best decided in the arena of politics, through legislation, or in the arena of equity law, in the courts.

In the U.S. the debate has played out almost entirely within the courts, with dozens of copyright infringement actions being filed around the country by artists and rightsholders against AI developers. That’s partly a reflection of the fact that ours here in the U.S. is a particularly litigious culture. You can sue anyone over pretty much anything if you pay a lawyer enough.

But it also reflects the fact there is a fair use doctrine enshrined in U.S. copyright law and the history how it got there. What we know today as “fair use,” was originally a common law concept developed by the courts. Most legal historians trace its origins to Folsom v. Marsh decided in 1841 by Supreme Court Justice William Story riding circuit. He didn’t use the term “fair use” in that case but discussed the need to consider whether a defendant in an infringement action had made “a justifiable use of the original materials.”

The question remained in the courts until 1976 when Congress finally codified the common law doctrine as §107 of the Copyright Act. Yet, even in its statutory form, the doctrine retained the the essence of Folsom v. Marsh. It lays out the factors for courts to consider but leaves to them the weighing of competing equities in particular cases to decide whether a defendant has made a justifiable use of the original materials notwithstanding the original author’s copyright. It is open-ended by both design and intent.

Many other jurisdictions recognize exemptions and limitations to copyright, to be sure, but in more cases than not they are fundamentally statutory in nature. In most European countries, for instance, those limitations are 9specifically enumerated in the statute and do not involve a weighing of equities: this is permitted, that is not. In that sense, Blackburn’s proposal hues more closely to the European model than the U.S. model, which straddles both statutory and common law, neither all one, or all the other.

There are many things to recommend in the U.S. model. It allows the law to be flexible and able to respond as needed to new use cases for copyrighted works brought about, for instance, by technological innovations. And it acts as an escape valve for free speech from the monopoly of copyright.

The question in the case of AI is whether that model is still up to the task of arriving at an equitable outcome. When tens of millions of works are being boiled down at once to create an AI model, can any court or any authority practically and fairly provide the sort of case-specific, fact-based weighing of equities the doctrine envisions for each one of those works? Probably not, which might support an argument for a uniform statutory treatment of the question.

On the other hand, would excluding a specific use case from the possibility of fair-use adjudication wreck the careful balance of the doctrine and statute? Possibly, which might support an argument for letting the courts do their thing as best they can.

More pressing, even, than the question itself of what to do about AI’s use of copyrighted works is the problem of deciding how to answer it. Until we figure that out, the question will linger.

Share

Subscribe now

Q

Ever ask yourself what Roland Barthes would make of the the debate over AI and copyright? I don’t recommend it without consulting a physician. It can make your brain hurt, or worse. But if you make the mistake that I made and actually go there, you could find yourself imagining he might have some interesting things to say about it. Once the brain pain wears off that is.

Barthes is best known for his 1967 essay “La mort de l'auteur” (“The Death of the Author”), in which he argued against relying on the intentions of an author to definitively explain the "ultimate meaning" of a text. Instead, he insisted in that classically discursive, verbose and confounding French philosopher sort of way, that it was only in the individual reader’s interpretation of the text that meaning could be found, and therefore was never fixed. "To give a text an author,” with a definitive intention or interpretation, he wrote, “is to impose a limit on that text.”

His approached shared something in common with his fellow Frenchman Jacques Derrida’s Deconstructionist theory (although Derrida was ot altogether a fan of Barthes), in that both dispensed with the structuralist notion of a single, definitive and discernable meaning of a text. But Barthesism’s closest American cousin was probably the “New Criticism” that flourished in certain academic literary circles in the middle decades of the 20th Century.

The central New Criticism idea was that of the “intentional fallacy,” first described by William Wimsatt and Monroe Beardsley in a 1946 essay. It declared that a text is “detached from the author at birth and goes about the world beyond his power to intend about it or control it. The [text] belongs to the public."

Post-structuralism, as those various modernist theories are collectively known, is a poor fit with the pre-modern notion of copyright. Although not a critical literary concept, copyright is in many ways the type specimen of a structuralist theory. It is premised on texts having discrete, identifiable authors and it invests those authors with the “power to intend about” the uses of those texts, at least up to a point.

I don’t know whether Sam Altman or Dario Amodei may have read Barthes at some point, or the New Criticism theorists (the evidence is weak). But the argument they make with respect to the uses to which they put texts carries a distinct, albeit cynical, echo of the intentional fallacy. They treat the texts they use to train their LLMs essentially as if “detached from the author” and going about the world “beyond his power to intend about” its uses, belonging to the public.

It’s not so much an argument from fair use as an argument from post-structuralism, i.e., once a text is made available to be read, it is free (or should be) to be scraped, tokenized, and reinterpreted as a collection of mathematical representations to float in multidimensional vector space. If that’s your thing.

The output of their large language models is an even more perfect avatar of post-structuralism. The text itself is not written but constructed, probabilistically and in real time, detached from any intention or any source of independent agency. Any “meaning” it may embody did not—really cannot—exist a priori, simply awaiting the spark of human intention to be expressed. It’s an algorithmic emanation embodying only the dumb laws of probability.

The problem I have always found with post-structuralism is that it is fatalistic to the point of nihilism. It is a negation of human agency. You cannot say what you mean because you cannot mean what you say. It thereby absolves all of responsibility for their actions. Or any joy or agony of shared experience for that matter.

The irony in the case of AI, is that its developers take a very deterministic view of technology: Technological innovation is good per se, and advancing AI technology in particular is essential, the end. It’s its own falsification of the intentional fallacy. Which, puts them very squarely back in the structuralist realm of copyright.

It’s enough to give an old English major a headache.

We’ll now return to our regularly scheduled programming.

Share

Subscribe now

Curious move by a group of leading British news publishers last week regarding AI licensing. The BBC, Financial Times, Guardian, Sky News and Telegraph Media announced the launch of SPUR, the Standards for Publishers’ Usage Rights, coalition. According to an open letter posted on the group’s website and signed by the respective chief executives inviting other publishers to join the initiative, SPUR will:

• Develop shared industry standards, creating responsible pathways for original journalism to be used sustainably

• Reduce friction in licensing and bridge the gap between publishers and AI developers

• Identify gaps in the technical tools needed to protect intellectual property, and support their creation

• Ensure high value content can be accessed through rights cleared, accountable channels

• Evaluate existing industry infrastructure and assess where new technologies or approaches are needed

• Enable transparent, scalable use of journalistic content

It’s curious because there are already multiple initiatives underway, involving multiple coalitions, to develop technical standards, ensure access through cleared, accountable channels, deploy infrastructure, and enable transparent and scalable use of content in AI, as discussed here, here, here and here, to note just a few recent ones.

There are even a couple (see here and here) that have produced published ISO standards. Is another such coalition really needed? Especially where, as with the BBC and Guardian, members are already involved in some of those other initiatives and coalitions? Or when a coalition planning to “identify gaps in technical tools” has no technical-tool developers as members, at least for now?

SPUR declined to elaborate on its plans and referred a media inquiry to the website’s FAQ. Under “Why are you doing this?” the FAQ says, “SPUR aims to help shape technical and commercial infrastructure in ways that deliver and support efficient marketplaces where content can be licensed on fair commercial terms.”

But it’s not immediately clear what the members of SPUR bring to that table that others haven’t already.

The fact that no technology companies are involved at the outset may be tell as to what’s going on here. Many of the other coalitions and initiatives working on AI licensing standards were organized or spearheaded by technology company, from Adobe-let C2PA to Cloudflare’s pay-per-crawl system to Amazon and Microsoft’s licensing marketplaces. SPUR, on the other hand, seems intentionally set up to be a publisher-led effort.

While the FAQ does say the group will “engage extensively with the buy-side of the market to ensure our solution creates value for AI licensees,” and invites “any developers who may wish to engage with the coalition to contact us directly,” that language seems carefully chosen. Contrast the invitation to developers to “engage with” the group with the invitation to any news organization “interested in joining the coalition” to apply for membership. Some invitees are more equal than others.

There is something to be said for licensing standards development being led by rightsholders rather than technology developers. All too often in the history of copyright’s encounters with digital technology, rightsholders have found themselves beholden to the technology rather than the other way around. The classic example is the introduction of Apple’s iTunes Music Store, and more specifically Apple’s proprietary DRM.

So relieved were the record companies in the wake of Napster to be offered a licensed, paid, and copy-protected outlet for MP3s they eagerly signed up. What they failed to realize—and came to regret—was that Apple’s proprietary digital rights management (DRM) meant the iTunes store was the only outlet they had and to reach consumers willing to pay for music they had to go through Apple. That allowed Apple to set the price of a download and dictate the terms of engagement. Which it did.

The classic counter-example is the DVD format. While the hardware engineering came largely out of Japan, the major Hollywood studios controlled the CSS DRM system and decreed the the exclusive DRM for the format. That allowed the studios to dictate the design of DVD players and drives. If a hardware maker wanted its players to be able to play Hollywood movies it had to license CSS from Hollywood, which meant adhering to the design specifications Hollywood set.

The result was uniform hardware/software compatibility, standardized disc manufacturing, rapid consumer adoption, and a much larger retail base than video rental stores, including mass market outlets like Walmart and K-Mart (at the time) and grocery stores. The fact that CSS is was almost immediately hacked was ultimately immaterial to the success of the format. Those other market factors yielded a huge windfall for Hollywood.

The founding members of SPUR do not enjoy the sort of monopoly leverage the seven major studios had in 1997 when the DVD format was introduced. But three of them—BBC, Sky and Financial Times—are among the top 20 news media companies by revenue in the U.K, according to Press Gazette’s latest ranking. And they all produce the kind of high-quality, fact-checked and timely content that is particularly useful to developers of large language models, which should give them some measure of leverage.

All five SPUR founders have, or easily could have, direct licensing deals with select AI developers. But those are necessarily limited. To achieve scale, standardization is essential, just as it was for DVDs.

Another notable feature of SPUR’s announcement is that the founders are all publishers of news. No book publishers, and no non-publishing companies, at least at this stage. “At this stage,” according to the FAQ, “SPUR coalition members are focused on journalistic content only, but the coalition is open to expansion to other media sectors.”

I haven’t discussed it much here, but I’ve long thought some degree of specialization in AI licensing systems will be necessary. While standardization is essential, different types of media have different business models, different distribution systems, and different constituencies to which they are accountable. Books, images, music and journalism may each require a licensing model tailored to their individual economic and operational needs.

Starting with a single vertical before seeing whether its solutions and infrastructure are transferrable could help keep the effort focused on practical, executable plans without biting off more than can immediately be chewed.

Share

Subscribe now

Slightly off topic, but given AI’s prominence in virtually every discussion in the media industries these days, and the turmoil caused therein, it seems worth taking brief note of the turmoil currently afoot in the AI industry.

On Wednesday, Nvidia, the world’s leading provider of GPU chips needed to power AI servers by a long shot, on Wednesday posted a 73% year-on-year jump in revenue in its most recent quarter, beating the Street, and guided to even stronger earnings in the current quarter—the sort of numbers that ordinarily should make Wall Street traders drool. Yet, Nvidia’s stock fell 4% on Thursday.

What gives? It wasn’t so long ago that every boffo quarter Nvidia posted sent its share price skyward. Why the tepid reaction this time?

It was reflection of growing concern on Wall Street that a near-term return on the massive investments tech companies have made to build AI infrastructure may be illusory, fueled in part by businesses failing to report productivity gains from deploying AI and heightened by strengthening public and political headwinds against its social and economic impacts.

Consider:

The Numbers

• No return: A February study by the National Bureau of Economic Research found that, while 70% of large enterprises are using AI in some capacity, 80% of those reported that A.I. was having no impact on their productivity or their headcount. By itself that’s not terribly surprising; most of those deployments were relatively new. But it suggests AI is yet to find a meaningful use case within companies and organizations, despite significant investment.

• AI zero: In a February interview with the Atlantic Council, Goldman Sachs chief economist Jan Hatzius claimed that AI investment spending has had “basically zero” contribution to the U.S. GDP growth in 2025. “We don’t actually view AI investment as strongly growth positive,” said Hatzius. “I think there’s a lot of misreporting, actually, of the impact AI investment had on U.S. GDP growth in 2025, and it’s much smaller than is often perceived.” The reason? Most of the spending on AI infrastructure is for chips and servers, most of which are sources from China and Taiwan, not the U.S.

• Bears are out: AI bears are beginning to smell blood in the Wall Street forest and have started looking for ways to short the likes of Oracle, Meta, Microsoft and yes, even Nvidia.

• ‘Bot’lash: Public resistance to data centers is building rapidly over concerns about their impact on utility costs and environmental factors, and has become a major political factor at the state and local levels. Of 770 centers on the drawing board in 2025, and another 200 under constructure, research firm Heatmap found 40% are facing meaningful local opposition, and 25 were cancelled altogether. Governors and state legislatures that once clamored to attract the centers to their states are now backpedaling furiously. In February, New York became the sixth state to propose a moratorium on construction. And the governors in Illinois and Connecticut who once championed the centers have been forced to retreat.

• More ‘bot’lash: An analysis by CNBC found that 42 states provide full or partial exemptions from sales tax to data centers. Of those, 37 have passed legislation specifically granting sales tax exemptions, and 16 of those have granted nearly $6 billion in exemptions over the past five years. Those breaks have sparked a growing debate over whether giant tech companies should receive such largess. “There was a giant transfer of wealth from taxpayers to shareholders,” Greg LeRoy, executive director of the nonprofit research group Good Jobs First, told the outlet.

• Super PACs: AI companies and investors have taken notice of the opposition and are dumping money in super PACs by the bucketful to try to keep the contagion out of Congress and state houses. The Financial Times tallied $265 million in AI related super PAC fundraising, with pro-AI groups vastly out-spending pro-regulation groups. The race to fill the congressional seat to replace retiring Jerry Nadler in NY has become ground zero in the lobbying battle, with a super PAC funded by A16z’s Marc Andreessen has dropped in $1 million into the Democratic primary against Assemblymember Alex Bores for his temerity of authoring the NY RAISE Act requiring large AI companies to disclose risk assessments and mitigation plans.

• Labor pains: The Federal Reserve, which is mandated by Congress to maintain maximum employment, has started looking at AI’s impact on the labor force. In a February speech to the NY Association for Business Economics, Fed governor Michael Barr said the central bank had begun to game-out what would happen if AI capabilities and adoption accelerated and triggered a “jobless boom” in which AI agents “replace or displace a range of professional and service occupations.”

The Vibe

• Doomsday report: A report by Citrini Research titled The 2028 Global Intelligence Crisis went viral last weekend by positing a scenario in which AI upends white-collar work and triggers mass employment and plunging stock values. Other analysts challenged the report, and to be fair, Citrini described as more a thought experiment than a forecast. But the report seemed to catch the new Wall Street vibe, sending the S&P down on Monday and leaving a feeling of psychological dread in its wake that lingered like a bachelor party hangover.

• Over the moon: A year ago you would not have seen stories like this February piece in the Wall Street Journal with the headline, “Big Tech’s AI Push Is Costing a Lot More Than the Moon Landing.” It reported that four tech giants—Microsoft, Amazon, Meta and Alphabet (Google), plan to spend more than $670 billion on AI infrastructure in 2026 alone. That’s bigger in real terms than the railroad expansion of the 1850s, the Apollo space program, and the decades-long build-out of the interstate highway system. As a percentage of GDP, in fact, it is bettered only by the Louisiana Purchase in 1803, which more than doubled the territory of the United States. Those sorts of comparisons would not have found their way into AI stories even six months ago. Something in the zeitgeist has changed.

• Gloom and doom: At the Cisco AI Summit in February, OpenAI CEO and AI’s chief pitchman Sam Altman, admitted there has been more resistance to AI’s “diffusion” and “absorbtion” than he anticipated. “Looking at what’s possible, it does feel sort of surprisingly slow,” he said. Meanwhile, Nvidia CEO Jensen Huang lamented in a recent podcast that “the battle of narratives” is being won by AI’s critics. “It’s extremely hurtful, frankly,” he said. “A lot of damage” has been done by “very well-respected people who have painted a doomer narrative, end-of-the-world narrative,” that is “scaring people from making the investments in A.I.”

• Feeling the pressure: Sam Altman is not handling it well. At the AI Impact Summit in India in February, Altman whined about the complaints about AI’s energy consumption and its impact on utility prices. “People talk about how much energy it takes to train an AI model – but it also takes a lot of energy to train a human,” he said. “It takes about 20 years of life – and all the food you consume during that time – before you become smart,” perhaps the stupidest analogy ever made about anything, although also perhaps revealing of how AI bros actually view meat engines.

Neither the growing doubts among investors, nor the shifting vibe, will fundamentally change the dynamic between AI developers and the creative community in the short term. But the more disenthralled the culture grows with the hucksterism about “transformation,” “super-human intelligence” (whatever that’s supposed to mean) and the miraculous but inchoate benefits of “innovation,” and the more skeptical AI companies’ financial backers grow of their claims, the more open both could become to the pleas of those who feel aggrieved by it.

End of rant.

Share

Subscribe now

So far as is known to science, light sabers are found only in a galaxy far, far away and have never been deployed in our own Milky Way. So, if you see a character wielding one in an AI-generated video, it would not be a huge leap to conclude that the AI model used had been exposed during training to some Disney-owned Star Wars footage. There would be no real-world examples of light sabers in use by humanoids that could have served as source material for the generated video.

Even if the user of the model that generated the video did not explicitly ask for a “light saber,” but instead described it as a “hand-held, sword-like weapon in which a constrained beam of light replaces steel as the blade material,” if it looked like Luke with a light saber there would still not be a lot of ambiguity around whether the model had been exposed to Star Wars footage. Again, there would be no real-world examples of such a weapon that could have served as training material. And, if there was no business relationship between Disney and the model’s developer, it would be a better-than-fair bet that the use of Star Wars footage was unauthorized.

Setting aside for now the question of law as to whether that use might still meet the test for fair use, the basic fact pattern would be hard to dispute. It might even convince a jury.

But what if the relationship between input and output is more ambiguous? What if a generated output was merely suggestive of a protected character, image or song? Is it possible to determine whether that image or song had been used to train the model that produced the output?

Not definitively, no. Not by working backwards from the output.

For one thing, the datasets used to train generative AI models are so massive that any relationship between an input and an output is so attenuated as to defy definition. More to the point, AI models, by nature and design, are probabilistic. A model cannot even be relied on to produce the same output from the same input twice, to say nothing of a different model. Short of outright copying, there will always be a certain irreducible amount of uncertainty trying to connect an output to an input.

I’ve written before that the harm AI poses to creators and rights owners is stochastic; the likelihood that a particular work or collection of works used in training will be implicated in a particular output is not zero, but it falls within a probability distribution rather than as a deterministic outcome.

But what if it were possible to produce probabilistic evidence of training use? That is, to establish the likelihood that a particular piece of IP was implicated in producing a certain output?

That’s what New York City-based Lightbar is working on. It is a tool the company claims can analyze an image generated by an AI model to determine whether a particular piece of IP was likely to have been used in training the model, which it expresses as a confidence score.

“[I]f you can call a character directly in a model and it generates something that is very close to accurate of what that character actually looks like, it was used in some way shape or form as part of their training data,” a Lightbar spokesperson explained in an email. “If you don't call a character by name and just describe the character and it then still generates something that is also very close to accurate of what a character looks like, it had to have been used in some way as part of their training data. This applies especially to more esoteric characters that are not prevalent on the public web as there's no real reference images to train from just by scraping the internet.”

Lightbar has set up what it describes as a research project that anyone can use to test whether Universal Studios’ characters were used to train Google’s Gemini models by entering prompts. Anyone can enter prompts designed to elicit evidence that Universal’s IP was used in training and then submit their test results to Lightbar. Users are awarded $1.00 for every approved submission.

According to the project’s terms & conditions, acceptable submissions involve “prompts and outputs” that have “clear evidentiary value,” of “the model’s capability to generate content highly consistent with having been trained on specific copyrighted works,” and/or that “show patterns of memorization, substantial reproduction, or detailed knowledge indicating training data inclusion.”

“We analyze the outputs and establish a confidence score based on the likeness, style, and prominence of the generation relative to the character itself,” the spokesperson told me.

Lightbar is not the only, or even the first, startup to try to peer into the black box that is an AI model to figure out what went into it.

I wrote a year ago, in a report for the Variety Intelligence Platform (paywalled, I’m afraid) about Palo Alto-based Sureel. It analyzes what it calls the "AI DNA” of a work or catalog of works, which can then be used to match with an AI-generated work.

Sureel founder and CEO Dr. Tamay Aykut described AI DNA to me at the time as those features of a work that an AI model typically would extract during training. For each work, he told me, the system computes an “attribution likelihood” estimate of how much of the work has been used by an AI system, or the degree of influence by a catalog of works. According to Aykut, the system yields an estimate of 85%-90% likelihood at the works level.

New York-based Personal Digital Spaces has also developed a process that computes the likelihood that a work has been used by an AI model.

Sony’s AI division recently announced it has developed a tool it claims can calculate a training-data attribution share from AI-generated music, although it does not appear to have immediate plans to commercialize the technology. Sony Music is currently suing AI-music generator Suno over its unauthorized use of Sony’s works to train its model, which may have provided the impetus to develop the attribution engine.

The question hanging over all of of those efforts is whether the probability estimates the tools produce will stand up in court. Unlike AI, the law is deterministic, not probabilistic. Effects must have identifiable causes to be actionable, and the harm produced must be actual and quantifiable, not merely stochastic. Even copyright law, where “substantial similarity” can be one indicia of infringing reproduction, other elements must be present to establish liability, such as access to the original and actual harm to the market for the original.

According to the Lightbar spokesperson, the impetus for developing its technology was the Studio Ghibli moment on ChatGPT a year ago. “The founder saw the Studio Ghibli moment… and realized that what was going on in the generative AI space was fundamentally wrong,” the spokesperson said. (The founder, who previously founded a successful fintech company, has chosen to remain anonymous for now due to “the sensitivities around what Lightbar is doing against big tech.”) But the Ghibli case did not result in litigation, at least not yet.

Not withstanding Lightbar’s refence to seeking submissions with “clear evidentiary value” in its T&C, the spokesperson acknowledged its results may not yield clear legal proof of infringement.

“Lightbar is actually just a microproduct… and it's one piece of a broader IP platform the team is building,” the spokesperson told me. “This particular use case was developed because we realized private models like Gemini and ChatGPT are really difficult for studios to gauge just how much infringement is happening or even understand how brand guidelines are being respected (or not). This is designed to give them a bit more leverage in their negotiations with the AI model companies by having definitive proof.”

That may strike some hoping for hard evidence of infringement as small beer. But any licensing of copyrighted works for use by AI models will involve negotiations. And in negotiations, leverage often comes down to information. If one party has relevant information the other party lacks, the party with the info will invariable have the upper hand.

Copyright owners currently are at a significant disadvantage in treating with AI developers because a gross information asymmetry greatly favors developers. They know exactly what datasets went into training their models and where they came from while copyright owners can only infer whether their works were used from indirect sources.

Having defensible probability data to lean on could meaningfully help creators rebalance those scales, putting them on stronger footing for when the discussion moves from the court room to a conference room.

Share

Subscribe now

Last week, Amazon Web Services quietly revealed plans to launch a marketplace where publishers can offer their content technology companies developing AI systems. According to a report by the Information, Amazon circulated slides ahead of an AWS conference that included reference to the plan in a presentation on its AI developer platforms including Bedrock and Quick Suite.

The announcement followed on the heels of Microsoft’s somewhat less quiet unveiling its own AI licensing platform a week earlier, which it dubbed Publishers Content Marketplace (PCM).

Unlike similar content marketplaces such as the recently acquired Human Native, which are largely focused on providing datasets for training AI models, the Microsoft and Amazon offerings are positioned as platforms for securing licensed access to real-time data retrieval to support AI-powered search applications.

In a blog post, Microsoft described PCM as “a solution that gives publishers a new revenue stream, provides AI systems with scaled access to premium content, and delivers better responses for consumers.” The goal, according to the post, is to make “authoritative content [that] lives behind paywalls or within specialized archives,” available to AI developers while providing publishers with “sustainable, transparent ways to govern how their premium content is used and to license it when it makes the most sense.”

Microsoft has lined up a charter group of publishers for the launch, including AP, Business Insider, Condé Nast, Hearst Magazines, People, USA TODAY, and Vox Media (notably missing from the list: the New York Times, still locked in litigation with Microsoft over unlicensed scrapping of Times content), and it designated its own Copilot system as the initial buyer in the marketplace.

Amazon’s announcement offered fewer details, but its slides aligned its marketplace with Bedrock, which is its platform for developing so-called agentic-AI applications, which would include AI-powered search tools. Given how many websites are hosted on AWS servers Amazon also has with a large stable of potential publisher clients to pitch.

I suggested a year ago that real-time, AI-powered search—what folks used to call “retrieval-augmented generation” (RAG) but now seem to prefer “conversational search”—presents both a greater potential threat to publishers than traditional search, but could also offer a more clear-cut licensing opportunity for their content than merely selling data to pre-train large language models.

With training, once the tokenized data have been extracted and incorporated into a model’s weights, the dataset has little to no further value to the model. For the data provider, therefore, a deal for training data is effectively a one-and-done proposition, not a recurring a revenue stream.

AI-powered search on the other hand, is…well, search. As with traditional search, it puts a premium on perpetual access to up-to-date, authoritative content. For certain types of publishers—those with the most up-to-date, authoritative content—that need for perpetual access by automated agents could potentially translate into recurring revenue.

As Microsoft put it in its blog post, “The open web was built on an implicit value exchange where publishers made content accessible, and distribution channels - like search - helped people find it. That model does not translate cleanly to an AI-first world, where answers are increasingly delivered in a conversation… As the AI web grows, publishers need sustainable, transparent ways to govern how their premium content is used and to license it when it makes the most sense.”

What’s needed, in other words, is the technical infrastructure to enable those “sustainable, transparent ways to govern” how content is used. That could include a scaleable marketplace, robust usage analytics, and some sort of gating mechanism to regulate access, among other things.

Microsoft and Amazon are two of the largest cloud-computing providers on the planet and can certainly deliver scale. They also have the means to generate and report analytics and can no doubt devise a technical means to effectively regulate access to content archives (something like Cloudflare’s managed-robots.txt platform).

What Microsoft and Amazon need is the incentive to go to the trouble of providing the infrastructure. And that incentive is revealed in Microsoft’s blog post: “PCM also provides usage-based reporting, enabling publishers to understand how content has been valued in the past and where it can provide increased value in the future.”

It’s the data. Particularly search-related data. Data on usage and user intent. The kind of data that made Google the undisputed champion of the previous generation of search.

Google did not invent the search engine. But it devised a superior strategy for indexing websites and organizing information. That one weird trick led more people to use Google to find the information they were searching for, and the more people who used Google the more data Google collected on their interests and intentions, further improving the results it delivered. The more data Google had on consumers’ interests and intention the better it also became at targeting advertising to them, which in turn drew more advertisers to use its platform.

That feedback loop ultimately allowed Google to achieve an effective monopoly over both search and online advertising.

Microsoft and Amazon are betting that the next generation of search, powered by AI, will be less about the best algorithm for finding information and more about superior access to premium content to answer queries directly.

By providing access to premium content via their marketplaces,and collecting data on how that content is being used, they’re hoping they will be the ones that end up with the data on consumers’ interests and intent.

Just as Google’s superior algorithm initially drew more users, who in turn drew more advertisers, Microsoft and Amazon are further betting that marketplaces with the best content will attract the most developers looking for premium content, and the more buyers they draw into the market the more sellers they will attract—in this case rightsholders—creating the same sort of feedback loop that powered Google.

As the market makers, Microsoft and Amazon will benefit directly from transactions that happen there, taking a small percentage of whatever payments developers end up making to rightsholders. But the greater treasure could be the mountain of usage data they end up sitting atop.

For all Microsoft’s talk of providing publishers with “sustainable, transparent ways to govern how their premium content is used,” the arrangement’s benefit to publishers is incidental to the strategy. But that doesn’t make it any less real. It’s the price to be paid to ensure the infrastructure is built to benefit at all.

Share

Subscribe now

In Brussels last week, the Legal Affairs Committee of the European Parliament, by a vote of 17-3, approved a series of proposals for ensuring the “fair remuneration” of rightsholders for the use of copyrighted works by generative AI. The proposals will be put before the full Parliament at the next plenary session scheduled for March.

While no doubt well-intentioned, like so many similar pronouncements the proposals are long on what the committee would like to see happen, and short on how to make it happen in the real world.

The committee calls, for instance, for rightsholders to have “full control over the digital use of their content by AI systems and models for training purposes,” by means of a “robust and functioning” (i.e. machine-readable) opt-out mechanism.

All well and good, but a meaningful opt-out mechanism needs to be machine-enforceable, not just readable. That is, the opt-out signal, once detected, must functionally disable the requesting machine’s access to the copyrighted work at the network level, and that disability must persist through downstream copies of the work.

My point is not to pick on the Legal Affairs Committee. The problem is fundamental to the digital media economy: the how is almost entirely in the network. It cannot be mandated or imposed from the outside. Any solution for realizing the what must become part of the network’s architecture.

That’s not easy to do as a technical matter. Reproducing files comprised of binary code is fundamental to how von Neumann machines and the networks that connect them operate.

Nor is it an easy sale to shareholders or prospective investors. Inhibiting the fundamental operation of the network is not nearly as sexy a proposition as building the network itself or the applications and capabilities it supports. Those are easier to do as a technical matter and therefore more likely to produce a near-term return on investment.

That’s one basic reason why you see hundreds of billions of dollars in investment capital going into building AI data centers today. Investing in infrastructure pencils out even if the business case for AI applications is yet to. It’s also why, historically, far less investment has gone into figuring out how to architect IP principles into digital networks, and where it has been tried the results have often been discouraging, as any music blockchain investor can tell you.

That under-investment is starting to change, however. Last week, the World Intellectual Property Organization (WIPO) announced the planned March 17 launch of the Artificial Intelligence Infrastructure Interchange (AIII, pronounced “A triple I), and opened applications for “any creator, rightsholder, tech professional, AI developer, or academic with the expertise” to join the project’s Technical Exchange Network (TEN). The goal of the initiative is to advance a “global dialogue on the technical and operational aspects of the intellectual property system in the context of artificial intelligence.”

That is, to figure out how to “enable digital content to circulate globally while helping to ensure that creators and rightsholders are properly identified and, where applicable, compensated by including “copyright infrastructure” within the architecture of AI.

Also last week, the internet infrastructure provider Cloudflare announced the acquisition of Human Native, a company I profiled back in 2024. Cloudflare is a major content delivery network (CDN) and cybersecurity provider whose core business is securing corporate and government networks against attacks. But as discussed here last year, it also developed the pay-per-crawl economic model for web publishers based on its managed-robots.txt platform.

The platform makes AI opt-outs expressed in a website’s robots.txt functionally enforceable at the network level by blocking a web crawler’s access to the site if the bot’s heading does not signal an intent-to-pay or an existing payment relationship with the publisher on the part of the bot’s source. Publishers and AI developers can register their payment credentials with Cloudflare, which then acts as the merchant-of-record to execute the transaction.

What sets pay-per-crawl apart from other opt-out based approaches to licensing is that it is based on access to the content, not the content itself. It does not rely on embedding a signal into the content, which may or may not prove persistent. And it removes the transaction from the abstract realm of copyright to the concrete world of network engineering.

Pay-per-crawl’s other critical feature is scalability. Plenty of direct AI licensing deals are being struck these days between individual rightsholders and AI developers. Scarcely a week goes by, in seems, that Universal Music Group doesn’t announce a new one. But as I and many other analysts have noted, that’s not a model that can scale beyond the relative handful of publishers with archives valuable enough that developers will be willing to pay for them. As pay-per-crawl is not based on the content itself, the size or perceived value of an archive is irrelevant.

With its acquisition of Human Native, Cloudflare adds another dimension to its AI licensing arsenal. Human Native’s business is aggregating datasets from rightsowners and making them available to AI developers. In between, it cleans and structures the data, and ensures necessary clearances are in place, to make it easily usable for training AI models. It’s a different approach to creating a model that can scale.

“I think the two approaches are actually very complementary,” Human Native co-founder James Smith told me last week. “It’s about giving publishers control.”

The goal, Smith says, is to “create an alternative to scraping.” Scraping, he notes yields “messy data” that is unstructured and often unreliable. The combination of Cloudflare’s access-control capability and a marketplace of pre-cleaned and cleared datasets is intended to steer AI developers away from relying on blind scraping for data to looking for ready-to-use quality datasets.

“What developers need from data,” Smith tells me, is “access, preparation, legality and quality. And it needs to be accessible in a marketplace.”

It doesn’t hurt, he adds, that Cloudflare’s CDN carries roughly 20% of global internet traffic, including all those blind scraper bots, giving the project a built-in degree of scale. The ultimate goal, Smith says, is to build “an internet-scale infrastructure for data commerce.”

Watch this space.

Share

Subscribe now

The Nieman Lab at Harvard University recently published its Predictions for Journalism 2026 report, featuring contribution from an array of leading digital media executives. Among those was one from David Skok, CEO and editor-in-chief of the Canadian business and technology outlet The Logic.

His prediction? Publishers will see no meaningful AI licensing revenue.

While Skok acknowledges the recent flurry of direct deals between publishers and AI developers, such those announced by Meta, Google, OpenAI and Microsoft, he argues they are not enough to create a genuinely liquid market for rights and that the model won’t scale.

“They will be more like public relations exercises than true economic signals,” per Skok.

The fundamental problem, in Skok’s telling, is that “[a]s long as Google’s search crawler and its AI-training crawler are functioning as a single system, the market for licensing journalism into AI models is effectively stalled.”

The tools available to publishers to regulate access to their content online, principally the robots.txt protocol, he suggests, are no longer up to the task in the face of AI.

With the workflows of AI crawling and search indexing intertwined, “blocking one means blocking both,” Skok notes. The effect is to force a choice on publishers between open access to AI crawlers and invisibility in search. “Most publishers can’t afford the latter. So the wall they assumed they could put up is gone,” Skok writes. “This removal of friction destroys the scarcity upon which a licensing market depends.”

His complaint is not a new one. It is a view widely shared among publishers. But the market may not be quite as stalled as he claims.

Last week, the RSL Technical Steering Committee published the first iteration of the Real Simple Licensing technical specification. First announced in September, RSL is an open standard, like the Real Simple Syndication (RSS) standard it’s based on, allowing anyone to implement or support it.

It is designed to work with robots.txt, augmenting robots’ simple yes/no permission gate by adding new permission categories, including “ai-all”, “ai-input”, and “ai-index.” The aim is to give publishers finer control over how their content can be used in AI-related applications, such as freely, as input for training, or for indexing purposes only. It also enables their disentangling search indexing from AI scraping.

“Today’s release of RSL 1.0 marks an inflection point for the open internet,” Eckart Walther, chair of the RSL Technical Steering Committee, said in a press release. “RSL establishes clarity, transparency, and the foundation for new economic frameworks for publishers and AI systems, ensuring that internet innovation can continue to flourish, underpinned by clear, accountable content rights.”

Although voluntary, like robots.txt, RSL has at least nominal support from some 1,500 publishers, technology companies, platforms and open standards organizations, according to the RSL Collective. Among those is Cloudflare, the web’s largest CDN, which has also attracted broad support for its recently announced managed-robots.txt platform, making the voluntary protocol enforceable at the network level and enabling a pay-per-crawl access model.

The RSL standard also includes support for a new “contribution” licensing option developed in collaboration with Creative Commons and meant to help strengthen the non-commercial web publishing ecosystem.

A voluntary technical standard is not a panacea for beleaguered web publishers, of course. There is no legal or contractual enforcement mechanism and no penalty for evading or ignoring it. Tinkering with robots.txt also does nothing to control downstream copies of content, nor does it solve the problems facing every kind of rights owners. It does nothing for music rights holders, for instance, whose works are readily scraped from streaming services.

But the publication of the standard and the broad initial support signal real movement within the media and technology industries toward standardizing the technical foundations of a licensing system, which is the sine qua non for getting any such system to scale.

Another approach to getting an AI licensing system to scale, of course, would be to make it mandatory. While that would take the issue out of the orderly, empirical world of engineering into the cut-and-thrust realm of policymaking, a new report commissioned by the European Parliament deigns to go there.

The report, The Economics of Copyright and AI, was commissioned by Parliament’s Policy Department for Justice, Civil Liberties and Institutional Affairs at the request of the Committee on Legal Affairs. It views the controversy over the use of copyrighted works to train AI models through the lens of earlier technology-enabled disruptions such as the Napster firestorm, and concludes the most viable solution is to implement a compulsory AI license for copyrighted works and statutory royalty.

“Among realistic options, a statutory licence with a modest royalty outperforms licensing markets, or copyright exceptions with or without opt-out” in balancing the interests of rights owners, technology developers and consumers, the report says. “Our calibration exercise suggests that statutory licensing generates about $14 billion more annual welfare than the next best alternative.”

Specifically, the report recommends policymakers “Avoid opt-out and adopt statutory licensing as the default framework. It is the most robust option for aligning private incentives with societal welfare. The royalty rate should be set at the lowest level that restores maintenance of creative supply.”

The proposal, especially the low royalty rate, is not likely to be music to the ears of artists and rights holders. Music publishers have chafed at statutory mechanical licensing since its inception, for instance, which they argue deprives them of the true market value of their works.

Yet, the fact that it’s being floated at the highest level of policymaking within the EU, which has led the world in digital regulation, is an indication that movement toward getting AI licensing to scale is happening at the policy level as well as the technical level.

It may not get there quickly (certainly not in 2026, in fairness to Skok). And it may never protect all participants’ interests all the time. But neither is it completely stalled.

Share

Subscribe now