Lauer opened with the problem of units. For a language model it makes no difference whether a text comes from a journal, a book or a series, he said. Everything is cut up and fed into the machine. Libraries and publishers therefore need to ask which stable unit remains in which provenance, authorship and responsibility can still be tracked.

Balke added that responsibility is at the heart of the trust question. Peer review is "obviously breaking down" under the volume of AI-generated papers and preprints, he said. The only fallback he sees at the moment is trusting authors and groups one knows personally. That runs against what science is meant to be, he stressed, and it is exactly what needs to be preserved.

Carpenter described a more basic shift. For the first time in history, the main consumers of scholarly language are not human. Access control, provenance, citation and peer review all have to be rethought for machine readers, and he expects the standards community to spend years on it.

Crossick pushed back on the gloom. Curated, verifiable information is exactly what separates publishers from "AI slop", she argued, and doctors, lawyers and scientists cannot work with answers that are merely probable. Peer review is challenged but not over, and publishing survived the arrival of the internet too. Her recipe is to treat responsible AI as everyone's job, to collaborate across the sector and to show up in standards bodies, because standards are written by the people in the room.

At Voegler's request, Crossick sketched the Act. It is essentially market access legislation and product safety law, she said. Anyone who wants to operate in the EU market must conform, and the higher the risk of a system, the stricter the rules. Some practices, such as social scoring, are banned outright. Everyday tools like spell checkers and grammar checkers are not caught.

Why should publishers care if they build no AI products? Crossick pointed to the Act's history. It was drafted before generative AI, and lawmakers added rules for general-purpose models once ChatGPT arrived. In her view these rules give rightsholders new leverage.

Providers must be transparent about what they train on and what their systems retrieve in RAG setups, because rightsholders can hardly negotiate or litigate over content they cannot see. They must also have a copyright policy that explains how they respect rights reservations under Article 4 of the copyright directive. A third rule concerns labelling. Users must know when they talk to a chatbot and when they read AI-generated content, with an exception where editorial control exists. Publishers that build AI tools for their own products will have to apply it.

Asked whether regulation goes too far or not far enough, the panel leaned toward too weak and too slow.

Carpenter said he is not against regulation but worries about rules that cannot keep pace with technology nobody can yet foresee. Part of the answer is technical. NISO, together with STM, Ithaka and COUNTER, launched a project called Trace the day before, which looks at how the provenance of content in AI systems can be tracked and how its metadata can survive inference. He also raised a concern about cryptographic fingerprints that, as he described it, a major AI developer now embeds in its outputs in response to the Act. Such marks could misrepresent how much a machine contributed to a text, he warned, and could make users of the tools traceable.

Lauer criticised that the Act separates models from data, although today the two are almost the same thing. Copyright is only one part of the picture, he said. How data and metadata are structured decides how capable a model becomes. He also questioned what transparency can mean in practice. If a researcher uses a chatbot late at night as a sparring partner to find a starting point, should that appear in the paper? Recording every step of a research process would require infrastructure that nobody has built.

Balke traced the line from typewriter to word processor to spell checker to foundation model. The tools that polish a text were never a big problem, he said. The difference comes when a model generates content the author did not write. Authors remain responsible that only what they want to say ends up in the text. He hopes for community self-regulation along the lines of Wikipedia, since peer review no longer scales.

Crossick counted 39 implementing acts, guidelines, codes of practice and other instruments attached to the Act and called the result a "many headed monster". The transparency requirements are thin, she said, and without consequences AI companies have little reason to comply. Lauer made the mirror-image point for researchers. If filling in forms takes three days after publication, nobody will do it, and in the end it stays open how much AI went into a paper or book.

Voegler summed it up. The sector is stuck between no consequences for the companies and no incentives for the scientists. Lauer agreed.

Balke explained why science is hard to sanction. The Act speaks of products, and a bad paper does little harm until it is built into one. Publishers, whose product is science, therefore carry a special duty to curate. He also sees a democratising effect in open scholarship, especially for low and middle income countries, while noting that nobody yet knows what to cite when a model has absorbed vast amounts of published text.

Crossick argued that the answer does not have to be a general-purpose chatbot. Publishers can build specialised tools that are grounded in verified content and show where each citation comes from. She named Elsevier's LeapSpace as an example and acknowledged the plug. Europe lacks the compute and capital of the United States, she said, but it has domain expertise and a strong scholarly publishing community.

Carpenter made the business case for regulatory restraint. Customers building aircraft or buildings will not accept probabilistic answers, he said, so grounding and citation will become market requirements. He cited engineers who lost money by working from an outdated standard. In the United States, he added, the interests of media companies and the software industry are drifting apart on copyright, because AI companies do not want to pay for content.

Lauer drew a sobering picture for Europe. A handful of large model providers, mostly American with some Chinese players, dominate the field, and specialised European services such as legal publishers' databases sit in tension with them, because a large provider could simply scrape and rebuild them. Universities, meanwhile, lack basic infrastructure such as sandboxes where departments could safely test and train models. His verdict was "Wild West".

Crossick pointed to the Frankfurt Appeal, issued this week, in which publishers ask policymakers to hold off on new AI legislation until existing rules prove effective. The appeal calls for effective transparency requirements, licensing incentives and proper enforcement. She also wants a liability regime. The exemption of intermediaries from liability has caused many problems with social media, she said, and the liability directive once championed by MEP Axel Voss may return.

Carpenter countered that new rules are not needed where existing law, such as cybercrime legislation, is not enforced. Crossick replied that court cases come too late. Content that has been taken cannot be taken back, so she wants stricter rules up front and immediate consequences.