The Artificial Intelligence Act deal requires general-purpose model providers to publish summaries of training content

2023-12-08EU Policy

Provisional political agreement between the European Parliament and the Council of the European Union on the Artificial Intelligence Act, reached 8 December 2023; Council press release, "Artificial intelligence act: Council and Parliament strike a deal on the first rules for AI in the world" (9 December 2023); European Parliament press release, "Artificial Intelligence Act: deal on comprehensive rules for trustworthy AI" (9 December 2023)

The European Parliament and the Council of the European Union reached a provisional political agreement on the Artificial Intelligence Act on Friday 8 December 2023, and both institutions published summaries the next day. The Act keeps a risk-based structure, with an OECD-aligned definition of an artificial intelligence system, prohibited practices including untargeted scraping of facial images, emotion recognition in workplaces and schools, social scoring and biometric categorisation on sensitive characteristics, and a high-risk regime with a fundamental rights impact assessment before deployment. For general-purpose models the Parliament's summary states that providers must meet transparency requirements including "drawing up technical documentation," comply with European Union copyright law and disseminate detailed summaries about the content used for training. Models designated as high-impact with systemic risk must additionally conduct model evaluations, assess and mitigate systemic risks, conduct adversarial testing, report serious incidents to the Commission, ensure cybersecurity and report on energy efficiency, relying on codes of practice until harmonised standards exist. An AI Office inside the Commission will supervise the most advanced models. Fines run from 35 million euros or 7 percent of global annual turnover for prohibited applications down to 7.5 million euros or 1.5 percent for supplying incorrect information. The text goes to technical finalisation, endorsement by Member State representatives and committee votes before formal adoption, and applies two years after entry into force, with exceptions.

What it changesThe obligations on general-purpose model providers create two documents the patent program has to think about before they exist: the technical documentation handed to regulators and the public summary of training content. Decide now which parts of that material are also written description for European patent applications, since the European Patent Office's practice under G 1/19 rewards a specification that explains how the training data and architecture produce the technical effect, and which parts are trade secrets, because anything published in a summary no longer meets the reasonable-measures condition for secrecy, and documentation given to a regulator stays secret only on the confidentiality terms the final text provides. The data-provenance record the training-content summary will require is the same record a patent application on a model benefits from, so build it once. The deal is provisional; the compromise text is not public and the thresholds for systemic risk are not in either press release, so budget for a second review when the text is released.

More on AI regulation and policy · Copyright and training data

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.