The EU AI Act Gives Publishers Leverage Without Teeth

When the Frankfurt Book Fair addressed the impact of artificial intelligence on publishing, the discussion turned quickly from creative disruption to regulatory reality. Publishers find themselves negotiating with major technology companies over the use of their copyrighted catalogues to train large language models. The conversation reveals a recurring problem in AI governance. Laws exist on paper, but the practical leverage for rightsholders to audit, restrict, or be compensated for training data remains weak.

The enforcement gap in content licensing

The core issue is that training an artificial intelligence model requires massive datasets, often scraped from the open web without permission or payment. While European Union rules require transparency about training data and respect for copyright opt-outs, enforcement relies heavily on complaints, regulatory bandwidth, and costly legal action. For an individual publisher, taking on a multinational technology company over data scraping is economically prohibitive. This creates a regulatory gap where legal rights exist, but the mechanisms to enforce them swiftly and fairly do not.

The governance failure pattern

This situation represents a classic AI regulation and enforcement failure. It occurs when a legal framework establishes clear prohibitions or obligations, such as copyright respect or biometric-use restrictions under the European Union Artificial Intelligence Act Annex III, but fails to provide accessible, rapid mechanisms for enforcement. Similar patterns occurred when Clearview AI scraped billions of facial images, drawing multiple data protection fines across Europe, yet took years of regulatory proceedings to restrict. The governance lesson is that rules without automated, verifiable compliance checks leave affected parties with all the risk and little recourse.

What the rules require

Under current European Union law, providers of general-purpose artificial intelligence models must publish sufficiently detailed summaries of the content used for training. They must also respect copyright laws, including opt-outs registered by rightsholders via machine-readable formats. However, knowing that your data was used is very different from proving it in a way that triggers a remedy or a fine. Governance teams inside publishing houses are realizing that waiting for regulators to step in is not an effective risk strategy.

Measuring your exposure

To manage this kind of regulatory exposure, organizations need to track specific indicators rather than hoping the law will protect their assets. Useful metrics include the percentage of your digital catalogue with machine-readable opt-out tags applied, the number of days of buffer your legal team has before statutory enforcement deadlines, and the time it takes your organization to produce a complete evidence pack when demanding an audit from a technology vendor.

What to do

• Audit your digital catalogues to ensure machine-readable copyright opt-out tags are correctly implemented across all web-facing assets.
• Establish a standard vendor assessment questionnaire that asks generative artificial intelligence providers to disclose their training data sources with verifiable proof.
• Track your regulatory compliance deadlines and maintain a ready-to-file documentation pack for any unauthorized data scraping incidents.
• Consider using automated governance platforms like autogovern.io to monitor your digital assets and maintain clear audit trails for regulatory reporting.

More from our platforms

These sister platforms cover the parts of this problem that sit outside governance.

• Argus (argus.threatclaw.ai) records every trace an AI application produces and scans it for prompt injection, jailbreaks and data leaks, including the attacks hidden inside retrieved documents and tool results rather than in what the user typed. Governance decides what an AI agent is allowed to do. Argus shows what it actually did.
• ThreatClaw (www.threatclaw.ai) tracks the threat side of the same systems: 22 live intelligence feeds, exploitation predicted before it is officially confirmed, threat actor profiles, and detection rules you can deploy straight away. A control is only as good as the threat it is sized against.