As generative AI continues to reshape content discovery and consumption, industry leaders are increasingly advocating for licensing frameworks that protect creators while enabling technological innovation. At a recent panel discussion titled "Licensing Is a Win-Win: The Exciting AI Partnerships Between Creators and Tech"—part of the "The Story Starts With Us," a day-long forum cohosted by the Association of American Publishers and the Copyright Alliance—executives from publishing, entertainment, and AI development highlighted the commercial and ethical imperatives of establishing sustainable licensing models.

Geoff Campbell, SVP of strategy and business development at Condé Nast, emphasized the shifting landscape of content discovery. "Consumer behavior is changing, so the front door of the internet is changing," Campbell said. "As a publisher, it's going to change the way that we need to make money off our content." He cited recent data showing that "referrals from generative search results have gone up by five and a half million" while "the corresponding decline in public top 100 publisher searches is 64 million."

Campbell noted that Condé Nast has already completed "two public deals in this space, both with OpenAI and with Amazon," while working on "several more" direct licensing agreements with both major and smaller AI companies.

The panelists unanimously rejected the notion that fair use provisions could adequately address AI training concerns. Michael D. Smith, professor of Information Technology and Public Policy at Carnegie Mellon University, warned that without proper licensing frameworks, "we're going to send inefficiently low incentives for the market to create things." Drawing parallels to past technological disruptions, Smith referenced his research on music piracy during the Napster era, asserting that "if the courts hadn't made the right decision in that case, we would not have any of the music licensing systems we have today.” He added, "this is publishing’s Napster moment."

The panel, moderated by Catie Zaller Rowland, general counsel at Copyright Clearance Center, emphasized that the market for licensing content to AI companies is developing rapidly. Davis, general manager of Protege Media, a company that offers a licensing platform for media, acknowledged that many rights holders may have "understandable fear" about AI, but his company, and others like it, were working to make licensing both viable and equitable. So far, he said, his company has worked with more than 50 “large catalogs of content.” He also pointed out that there are several startups securing significant funding for ethical AI models, such as Moonvalley, which he said has raised more than $100 million from "top-tier venture funds" to build video models based on ethically licensed content.

Vered Horesh, chief of strategic AI partnerships at the visual generative AI company Bria.ai described how Bria developed attribution technology that "measures the impact of any authentic asset being provided into the training catalog on any synthetic output being generated." This system "guides the allocation of payouts among data partners," creating what Horesh called "an ecosystem that is sustainable that creates the right incentives for all players to long-term continue to collaborate."

The panelists identified multiple advantages beyond monetary compensation for content owners who engage with AI companies. Campbell explained that licensing leads to collaboration: "You actually end up collaborating with that company around innovation. New things come out on product side [and] engineering side...through these partnerships." He added that licensing must be voluntary, not compulsory, the latter of which risks prompting AI companies to "pull back and stifles innovation."

Horesh reported that many of Bria's content partners have become "our customers...and also our resellers and channels, because now they want to make this technology accessible to their audiences as well."

Davis highlighted the legal uncertainty inhibiting AI adoption in creative industries. "There's a real downstream reluctance to use these AI models in a lot of businesses," Davis said, noting that a "senior legal executive from one of the major studios" recently disclosed that his company avoids using AI in production workflows due to concerns about "copyright liability" and the risk of being "enjoined from releasing a movie" after investing millions in production.

While the panel was optimistic about the development of voluntary licensing markets, Smith expressed concern about short-term challenges from companies "trained on moving fast and breaking things" who are "trying to get as much content as possible so that they can be the winner-take-all in this game." He called for legislators to "step in and give us all some rules that we can live by."

Campbell revealed that Condé Nast is "exploring three different kinds of commercial licenses in terms of use and distribution—B2B, B2C, B2B2C—[with] different use cases depending on what the licensor is looking for, and then also training, RAG [retrieval augmented generation], and access."

As Campbell noted in response to a question about data valuation, the success of content licensing depends on maintaining a sustainable ecosystem: "If AI companies bring down the value of companies too much, who are going to be the human creators that are going to create the content that is going to teach these models something new? As AI developers, our main goal is to make sure there's a sustainable ecosystem."

New firm promises to detect content's use in AI training

One of the most significant challenges in establishing effective licensing frameworks is determining when and how much copyrighted material has been used without permission in AI training datasets. At a presentation titled "Tech Innovation to Protect Creators Against AI Abuses," Louis Hunt, cofounder and CEO of Valent, discussed his newly launched company's emerging technology, which aims to address this critical issue.

Hunt, whose background includes serving as CFO and head of business development at the MIT-based AI company Liquid AI, explained that Valent aims to solve what he called "critical challenges in IP” as regards AI.

"If alpha in AI is in the data, how do you identify alpha?" Hunt asked during his presentation. "How do you identify the most creative sources of data and how do you facilitate legal permission to access to it in a way that is efficient?" He noted that the space around intellectual property and data in AI is "rife with challenges and information asymmetries," particularly the fact that rights holders often don't know what content has been used in AI models.

Hunt demonstrated two algorithms his company has developed: one that can calculate with "up to 98% confidence interval whether any given data sample was used to train an AI system," and another he called their "evident algorithm," which can generate evidence "with 99.99999999% certainty" that a model was trained on specific content.

The technology produced striking results when analyzing various content types. For one customer, Valent identified 7.54 million URLs and nearly 2.5 million web pages of their content available in commonly used AI training datasets. When examining books, including the Harry Potter series, Hunt's team found that 31-35% of the content was "memorized verbatim" by certain AI models. Valent's technology also detected unauthorized use of song lyrics, screenplays, and proprietary software code. He said that this can all be done externally, without direct access to the core code of the models themselves.

Beyond detection, Hunt explained that Valent has developed algorithms that can quantify how specific data would improve an AI model's performance, giving content owners leverage in licensing negotiations. The aim of the technology is to help content creators identify unauthorized use of their works and negotiate fair compensation, representing a shift in the current AI ecosystem where creators often lack the technical means to prove their content has been used without permission.

"We've solved one of the major problems still confounding the model developers themselves today," Hunt claimed, asserting that Valent’s technology can evaluate data quality for AI systems better than what model developers currently use internally.

In response to questions from the audience, Hunt addressed several technical points, including confirming that their analysis shows AI models are often cumulative rather than completely retrained, contradicting claims that content can simply be removed from future updates should rights holders demand it. He also noted that "unlearning"—removing content from a model that has already used it for training—"has not been successful" despite claims to the contrary.