AI Act · Article

Transparency obligations for general-purpose AI

You do not have to train a model to inherit obligations under the AI Act. If you build on, or ship, general-purpose AI, transparency duties apply — here is what you must disclose and record.

Plenty of teams assume the AI Act only touches the labs that train large models. It does not. The moment you put an AI feature in front of users — even one built on someone else’s model — transparency obligations can apply. They are among the easiest duties to meet and the easiest to forget.

Transparency is about not deceiving people. If someone is talking to a machine, or looking at something a machine generated, they are entitled to know.

When transparency applies

Transparency duties attach to specific situations rather than to risk tier. The common triggers are systems that interact with people, systems that generate or manipulate content, and certain biometric or emotion-recognition uses.

”Transparency is the cheapest obligation in the Act to satisfy — and the most visible one to get wrong.”

Telling people they are dealing with AI

Where people interact with an AI system — a chatbot, a voice agent — they should be informed they are dealing with a machine, unless it is obvious from the context. The disclosure should be clear and timely, not buried in terms.

Marking synthetic content

AI-generated or manipulated audio, image, video or text — including deepfakes — generally needs to be marked as artificially generated. For users, that means a visible label; technically, it can also mean machine-readable marking so the provenance travels with the content.

Practical step

Decide your labelling pattern once — wording, placement, and any machine-readable marker — and apply it consistently across every surface that ships generated content.

If you ship a general-purpose model

Providers of general-purpose AI models carry additional duties — including technical documentation, information for downstream providers, and a policy to respect copyright. If you only build on a third-party model, much of this sits with the upstream provider, but you should confirm what they supply.

Keeping records

Treat your disclosures as evidence: record where AI is used, what users are told, and how generated content is marked. When duties are spread across many features, a central register stops gaps appearing. Keep it with your other compliance records so nothing drifts out of scope unnoticed.

Key takeaways

  • Transparency duties apply to deployers, not just model trainers.
  • Tell people when they interact with AI, and mark AI-generated content.
  • General-purpose model providers carry extra documentation and copyright duties.
  • Keep a register of where AI is used and what you disclose.

Not sure which features trigger disclosure? Book a 30-minute demo.

This article is general guidance, not legal advice. Verify obligations and deadlines for your organisation with qualified counsel.

AI ActTransparencyGPAI
Put it into practice

Ship AI features
that disclose clearly

Book a 30-minute demo and we will map your transparency duties across the AI you build and the AI you buy.