ChatGPT launched just under four years ago, Reid noted. On the Gartner hype curve, the market has passed the peak of inflated expectations and the trough of disillusionment and is now climbing the slope of enlightenment, with strong productivity gains and growing user engagement. Three developments back the claim that this is the end of the beginning.

The first is infrastructure. Huge numbers of machines and bots are hitting publishers' websites, and the volume keeps accelerating. Infrastructure providers such as Cloudflare mitigate part of it. The second is end users. Wiley Scholar Gateway has been in pilot for a year, and researchers using it have shown where content proves useful when combined with different models and tasks.

The third is the traditional institutional customer, which Reid called a lagging indicator. Libraries and institutions see their users working with AI tools and want to connect them to their subscriptions, while dealing with budget pressure and information security concerns.

Among publishers, Reid sees three stages. Acceptance is now broad, even if some still call AI a bubble. The financial argument may have merit, but as a technology AI is transformative enough to stay. Understanding is the next stage, meaning the implications and the growth trajectory, not just the technology. The data points to a move from agentic discovery, where machines find, read and synthesise content, to agentic consumption, where machines carry out specific tasks, for example in self-driving labs.

This is more than a technology change, Reid stressed. It is a business model and commercial change. A frequent request from publishers is help translating technical details into commercial terms. One example is a disallow rule in a robots.txt file, which can seriously affect a publisher's ability to do deals.

The third stage is action. Publishers have agency and should experiment and put offers into the market. Reid's plea was to talk to end users, above all researchers and power users. How people access and consume content changed little for a long time, even in the move from print to digital, and it is now changing extremely fast.

Kaufman opened the discussion by pointing out that AI is not one thing, any more than "computers" or "the internet" are. How law, licensing and technology apply depends on the use case. Asked how much legacy technology holds publishers back, Reid said less than one might think, and the problem is not unique to publishers.

The web was built for humans, Reid said, and machines need to be directed into the right channels. Different machines play different roles and should not be treated as one lump. The aim is a web that stays built for humans, who will keep reading, plus a route for machines that is authenticated, measurable and attributable, so that publishers get the benefits without carrying too much of the cost.

Sutton said that STM's members share the copyright and licensing concerns of other rightsholders, with one addition. Scholarly content is verified, but it is also corrected and sometimes retracted, and AI brings a flood of fraudulent or manipulated material. Academic publishers have been among the most welcoming of AI, she said, and some members both supply training content and build AI tools. In her view, licensing gives publishers leverage to say that access comes with expectations of good behaviour. STM has worked on considerations for training language models on scholarly content.

Open access licences need a rethink too, Sutton argued. When the community pushed CC BY, it did not foresee this use. Creative Commons published new thinking in September, including preference signals, but Sutton doubts that AI developers will heed what individual authors would like. She also stressed accountability. Science needs a genealogy of knowledge that shows where an idea came from, and that is not yet traceable in AI systems.

Toral, whose company Frontiers is a pure-play open access publisher that leaves copyright with its authors, was equally unconvinced by Creative Commons' guidance on AI training, which he summed up as "it depends". He noted that the Berlin Declaration treats attribution as a community matter and that attribution lawsuits in science are rare. That does not mean AI companies deserve a free lunch, he said. Machines accessing content cost money, and that cost ends up with research institutions.

Kaufman added the notion of high stakes. Where a business decision, a cure or the safety of an aircraft depends on AI output, accuracy that is right most of the time is not enough. He sees that as a reason for optimism about peer-reviewed content.

Kaufman framed the cost problem. In scholarly publishing, costs circulate in a closed loop between publisher, funder, institution and researcher, while AI companies take value out without contributing to it.

Reid confirmed the pressure from the platform side. Infrastructure costs keep rising, with no revenue coming back from publishers to pass them on. He still sees green shoots and called himself a cup-half-full person. Wiley Scholar Gateway, launched with Anthropic, includes a bring-your-own-licence approach in the connector. A researcher can read perhaps 100 to 150 papers a year, he noted, while a machine has no such limit. A machine that fact-checks for journalists would be a new use of scholarly content.

Toral said publishers have become part of someone else's value chain, which feels alien. Frontiers focuses on costs and attribution rather than on extracting revenue from open science. He expects AI companies to want proper attribution in order to sell a scientifically viable product, and some companies have already approached Frontiers with offers to ensure it. Sutton wants to make cooperating with publishers more valuable than scraping. She cautioned that the point of contact with content may move away from publishers' platforms, and that researchers who now push back on prices may balk when expensive domain-specific AI tools are sold back to libraries.

The flood of AI-generated submissions is a cost as well, Kaufman said, because someone has to reject them. "It's not just price, it's cost," he added. At Frontiers, Toral said, the rejection rate is around 60 percent, much of it desk rejection. The company invests heavily in AI tools, since it has become an arms race and papers can no longer be filtered without them. Legitimate research polished by AI sits next to papers with synthetic data. Frontiers offers authors AI tools in a closed ecosystem. Kaufman's advice for talking to tech companies was to lead with how licensing helps them do their jobs, not with copyright.

Kaufman asked whether the value proposition for AI companies is less about access to content and more about reduced uncertainty over rights, cleared content, provenance, metadata and a single route to license at scale. Sutton agreed. Beyond legal certainty, she said, developers get a validated record that can be corrected. Publishers still have to build the mechanisms to pass corrections on, for example through tagging, and she pointed to the Trace initiative announced this week. Kaufman added that the corporate clients of his collective licensing organisation need to know exactly what was used and in which version, and are willing to pay for it.

Toral gave the retraction example. A bot that scrapes a study has no idea which version it saw or whether the study was later withdrawn, which matters for something like drug development. A structured feed agreed with the companies that ingest the data would make sure the science stays fit for purpose.

Reid called the scholarly record a living thing that changes through corrections and retractions. He recalled asking a national laboratory that was training a model what it would do about retractions and getting no answer. Researchers also tell Wiley that an answer from a chatbot without Scholar Gateway is one they do not trust, while connected answers are cited and referenced. It is as much a market issue as a copyright issue, Reid said. Sutton added that the corrected version lives on the publisher's platform, which is part of the value proposition.

The only audience question came from Patty of the Australian Publishers Association. Australian universities are telling researchers to let AI rip, she said, which raises copyright concerns, and she asked what advice publishers give authors about AI-generated material.

Toral stressed that this was not legal advice. Research funded by US federal money is not copyrightable and falls into the public domain, so open access publishers like Frontiers are comfortable without a copyright fight. Authors are told to use the tools they like and to use AI responsibly. With international author teams, he said, there is no simple percentage of permitted AI use, because the legal analysis of where copyright sits is too complex. Reid added that Wiley publishes more than 4,000 journals, many owned by professional societies with their own views, so guidance has to balance advising, educating and listening.