The European Union has published a document setting out the code of conduct for companies developing large, general-purpose AI models.

The European Union yesterday published the first draft of the Code of Conduct for organizations training general AI models. Companies have policies in place to manage risk and avoid fines. The published text is a first version. The final version of the regulations is not expected to be available until spring 2025.

Code of conduct in four key areas

The basic principles are now clear. The document is the first attempt to define a code of conduct for developers of advanced AI models trained with computing power greater than 10²⁵ FLOPs. Companies likely to be covered by this policy include OpenAI, Google, Meta, Anthropic and Mistral. The European AI law, which came into force on August 1st, still left scope for the establishment of detailed GPAI regulations in the future. This has been done and the document is now submitted for feedback and refinement.

The Code of Conduct covers four core areas: transparency (e.g. providing information about web crawlers used), copyright compliance (attribution when training AI models), risk assessment (to prevent cybercrime, discrimination and loss of control) and technical risk mitigation (e.g. B fail-safe access controls and model data protection).

Violations can result in fines of up to 35 million euros or seven percent of annual global profits. Companies can submit feedback until November 28th and the final document will be published in May 2025.