Does the EU AI Act Apply to Open-Source AI? A Plain-English Guide to the Article 2(12) Exemption

The headline rule

Article 2(12) of the EU AI Act says the Regulation does not apply to AI systems released under free and open-source licenses - unless they are placed on the market or put into service as a high-risk AI system, or as a system falling under Article 5 (prohibited practices) or Article 50 (transparency). That "unless" clause does most of the work, and it's narrower than most maintainers assume. The Linux Foundation's explainer and Hugging Face's guide for open-source developers are both useful starting points if you maintain or ship open models.

Where the exemption actually holds

For a plain AI system - not a general-purpose AI model - released under a genuine free or open-source license, with no monetization and no use in a prohibited or Annex III high-risk context, Article 2(12) means the AI Act largely does not apply. A hobbyist image classifier posted on GitHub under an MIT license, used for a non-high-risk purpose, sits outside the Act's scope entirely.

Where it stops - and it stops fast

The exemption evaporates the moment any of the following is true:

• The system is high-risk under Annex III. Biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and administration of justice all remain fully in scope even when the underlying code is open source. An open-source CV-screening tool used for recruitment (Annex III(4)) gets zero benefit from being open source.
• The system falls under Article 5. Absolute prohibitions - subliminal manipulation, exploitation of vulnerabilities, social scoring, and similar practices - apply regardless of licensing model. There is no open-source carve-out for banned practices.
• The system falls under Article 50. Transparency obligations for systems that interact with natural persons, generate synthetic content (including deepfakes), or perform emotion recognition and biometric categorisation apply even to open-source releases.

Open Future's AI Act and Open Source observatory and Orrick's analysis of the Act's application to open-source projects both map these boundaries in more detail, including edge cases around fine-tuned derivatives.

The monetization trap

The exemption is conditioned on the release being genuinely free and open source. If you monetize an open-source AI component - dual-licensing, paid hosting tied to the model, commercial support bundled with access - regulators and legal commentators treat that as taking the release outside the spirit of Article 2(12), and the ordinary AI Act obligations re-attach. This is a real trap for open-core business models: the moment "open source" becomes a funnel into a paid product built on the same system, don't assume the exemption still covers you. Rephrase's explainer on the open-source exemption walks through this distinction.

General-purpose AI models get a different, narrower carve-out

Open-source AI systems and open-source general-purpose AI (GPAI) models are governed by different provisions. For GPAI models released under a free and open-source license, Article 53(2) exempts the provider from only two of the GPAI obligations: drawing up the full technical documentation otherwise required under Article 53(1)(a), and (in part) the requirement to make information available to downstream integrators in the same depth as a closed model would require. Every other GPAI obligation - the training-content summary requirement, copyright-compliance policy, and cooperation with the AI Office - still applies to open-source model providers.

Critically, GPAI models classified as carrying systemic risk get no exemption at all, open source or not. If your open-weight model crosses the systemic-risk compute threshold, Article 55's additional obligations (adversarial testing, incident reporting, cybersecurity) apply in full. AO Shearman's "Zooming in on AI" series on open-source AI and inno³'s open-source and AI Act analysis go into the systemic-risk threshold in more depth.

The practical test

Before assuming your open-source project is exempt, ask three questions: Is it a high-risk Annex III use case? Does it touch a prohibited practice or an Article 50 transparency trigger? Is there any monetization attached to the release? If the answer to all three is no, Article 2(12) likely covers you. If any answer is yes, treat the AI Act as fully applicable and start the compliance workstream regardless of your license file.