At Frankfurt, publishers test Europe's AI law against machine-made science
Publishers at Frankfurter Buchmesse on 7 October 2026 debated how Europe's AI law affects trust in science.
The session, listed in the fair's official timetable as Preserving Trust in Science, brought together Max Voegler, Gerhard Lauer, professor of book studies at Mainz University, Wolf-Tilo Balke of TU Braunschweig, Todd Carpenter, executive director of the US National Information Standards Organization, known as NISO, and Elizabeth Crossick, head of EU government affairs and global AI policy lead at RELX Frankfurter Buchmesse. The fair describes the European AI Act as landmark legislation with far-reaching impact on publishing. Artificial intelligence is a cross-cutting theme across its trade programme.
Crossick described the Act as market-access legislation built on risk tiers, meaning higher-risk uses face stricter duties. For providers of general-purpose AI models, meaning systems trained on broad data that can perform many tasks, she said the law requires transparency about training data Publishing Perspectives. The European Commission uses similar terms, saying the Act puts in place transparency and copyright rules for those providers European Commission.
She said providers must also publish a copyright policy that respects rights reservations under Article 4 of the EU copyright directive, meaning machine-readable opt-outs that tell data miners not to use a work. The law also requires labelling of AI-generated content, with an exception where editorial control applies.
Enforcement is complicated. Crossick counted 39 implementing and related texts around the Act, including nine implementing acts, nine guidelines and eight delegated acts. A separate General-Purpose AI Code of Practice is intended to help industry meet duties on safety, transparency and copyright.
Balke said peer review can no longer keep pace with submission volumes, including AI-generated preprints, meaning early research papers shared before formal review. Volume is the problem. Referees cannot read fast enough to filter everything by hand.
Carpenter pointed to one technical response. NISO, with STM, Ithaka and Counter, has launched the Trace project to track content provenance, meaning a record of where material came from, in AI systems and to keep metadata intact through inference, meaning the stage when a model generates an answer. The aim is to keep source and rights information attached as text moves through machines.
Crossick cited the Frankfurt Appeal, issued the day before the panel, in which publishers ask policymakers to make existing rules effective before adding new ones. For readers, this means the fight is less about new bans than about making current permissions work. Leverage exists on paper. Making it stick is the next chapter.
