Opt-Out Flops Out At WIPO Meeting on AI and IP

Opting out of allowing copyrighted IP works from being used to train generative AI models currently is one of the hottest topics in technology policy circles. And it was the topic of a debate hosted by the World Intellectual Property Organization (WIPO) Wednesday between Andrew Gass, a partner at Latham & Watkins and lecturer at UC Berkley School of Law, and Ed Newton-Rex, a former Stability AI executive who left the company last year and formed Fairly Trained, an organization that certifies fully licensed training datasets.

The two disagreed sharply on the ethics of opt-out regimes and other issues, but there was one area where they mostly agreed: requiring IP copyright owners to affirmatively opt out of AI training is neither technically feasible nor practical as a matter of policy.

For Gass, who currently represents five technology companies being sued over the development of Large Language Models (LLM), the feasibility question is primarily a problem of scale. “The starting point for this conversation has to be an understanding of the volume of data that is a technological requirement for creating the kinds of tools that we’re talking about here,” he said.

Gass noted that Meta’s Llama 4 model was trained on 32 million tokens, or roughly 30 trillion words. Even the widely touted “small” Chinese model DeepSeek used 14.8 trillion tokens. By comparison, he said, “The entire 100-year archive of the New York Times has 5.5 billion words in it. So a very small fraction of the tokens that are required.” The question of what information a model is really extracting from all those words aside, Gass contended it is simply not feasible to develop and deploy a technology to reliably screen that volume of text for every instance in which an opted-out work might occur, from whatever source, and remove it.

Newton-Rex, a choral music composer who led development of Stability’s StableAudio music generator before leaving the company, agreed, but argued that it is unrealistic and unfair to put the burden of opting out on creators, regardless of the technological means used.

“It’s virtually impossible, I think, to picture an opt-out scheme that becomes well known among the majority of the people who would actually be eligible to use it,” he said. “Even the most widely used AI opt-out method to date, robots.txt is little known among IP creators and poorly understood.”

He cited a 2024 study by Cloudflare that found that even among the top 100 most visited sites hosted on its platform only 16% blocked any web crawlers at all. “And  actually, if you look at the top 1,000 sites, they’re still very popular sites, only 8.8% of those blocked any AI callers, and that was a full 15 months after the release of ChatGPT when you’d really expect that people would understand by then that they had this right to opt out.” Newton-Rex said.

As for the technical feasibility of an opt-out regime, Newton-Rex shared Gass’ skepticism, and offered his own experience to illustrate the problem.

“Location-based opt-outs only work for web domains that you control. I have no control over the web domains where recordings of my compositions are shared, except my own website, and that doesn’t have the most visitors,” he said. “Unit-based opt-outs add metadata to indicate the opt-out status, but there’s no metadata on any of my compositions. And even if there were it wouldn’t be transferred to a choir’s recordings. That’s a totally different media format. And even if by some miracle it actually were, it’s then easily removed, sometimes intentionally, sometimes unintentionally when you upload it to a new platform.”

That’s where the agreement ended, however. To Newton-Rex, the other fair solution is to adopt an opt-in regime and to require licensing of data for use in training.

“I strongly believe that the only way to effectively ensure that IP rights holders’ works are not used for generative AI training against their wishes, in a way that I think is fair to both rights holders and AI companies, is for training to be based on opt-in consent on licensing.”

Gass took exception. “I will exercise all of the fortitude and discipline that I have not to rebut Mr. Newton Rex’s comments, which I sort of categorically disagree with.”

To Gass, there are fundamentally conflicting views about the nature of creativity and the purpose of copyright law that need to be addressed before a policy consensus can be reached.

“Behavior that we have seen from AI companies has been largely predicated on the idea that it is permissible as a matter of law to take pre-existing instances of expression in a given genre and analyze them to extract uncopyrightable facts about the world and about the genres in which a given instance of expression participates.,” he said. “And it’s hard, I think, to have a constructive dialogue about these issues without grappling with that underlying dispute. We have on one hand sort of this reflexive notion that I wrote it, it’s mine, no one gets to do anything with it unless I say so. And on the other hand, we have this idea that the purpose of incentivizing the creation of new works is specifically and precisely to enrich our creative culture and further the progress of scientific advancement.

“And it’s really that conflict that I think you see playing out here in a lot of different respects.”