Training data, copyright and privacy in AI

In short: AI models are trained on huge amounts of data, often from the internet. This raises questions about copyright, personality rights and privacy that courts and legislators are currently clarifying.

In more detail: Two levels: may data be used for training? And: what happens to data users enter into a model?

In detail

• Text and data mining: in the EU copyright law allows automatic analysis of works, rights holders can object (opt-out, for example via robots.txt or metadata).
• Lawsuits: authors, image agencies, music labels and publishers are suing providers. Decisions vary.
• Outputs: models can partially reproduce rare training texts verbatim (memorisation).
• Personal data: the GDPR also applies to training data. Rights of access and deletion are hard to implement when knowledge sits in weights.
• User inputs: some providers use chats for training unless you object. Never enter passwords, health data or trade secrets into third-party services, see Using AI assistants sensibly and safely.
• Local models: run on your own computer without data leaving, see GGUF and llama.cpp.
• Licences: open weights have their own licences, see Open models and their licences.
• Note: not legal advice.

See also: EU AI Act and rules for AI, Data protection in a small business