Open source and the EU AI Act
“We ship under Apache 2.0, so the AI Act does not apply” is the most confidently repeated wrong statement in this field. There are two separate carve-outs, at two different levels, and both are narrower than the sentence implies.
Two carve-outs, two levels
| System level — Article 2 | Model level — Chapter V | |
|---|---|---|
| What it covers | AI systems released under free and open-source licences | General-purpose AI models released under a free and open-source licence |
| Does not apply where | The system is placed on the market or put into service as high-risk, or falls under Article 5 or Article 50 | The model presents systemic risk |
| Conditions | Free and open-source licence | Licence allowing access, use, modification and distribution, plus public availability of parameters including weights, architecture information and usage information |
| What survives anyway | Everything in the three exclusions above | Copyright policy and training-content summary duties |
What that leaves
Read together, the Article 2 exemption covers minimal-risk open-source systems. Anything high-risk, anything prohibited, and anything that interacts with people or generates synthetic content is outside it. That is a much smaller space than the phrase “open source is exempt” suggests.
Where this usually goes next
Three situations account for most people reading this page. Each has a different answer.
A deal is blocked on an AI questionnaire
Legal will not sign until you can evidence how AI is governed. HumanAudit’s AI Trust Package is a fixed $3,500 over five business days: a public trust page, a pre-filled SIG Lite / CAIQ / SSPA Section K questionnaire bank, and your AI inventory and classification.
You need ISO/IEC 42001 documentation
23 clause-mapped AIMS documents with all 38 Annex A controls pre-populated, editable and yours to keep, from $199. Or score your gaps first: 18 questions, free, no signup to begin.
You are not sure what reaches you
Twenty minutes with the founder. No prep, no deck, straight to the person accountable for the work. If none of this applies to you, you get told that on the call.
This reference is published by HumanAudit Inc. Not a law firm, not an accredited certification body, not a registered auditor. We build documentation, your counsel interprets it, and an accredited body of your choosing certifies you. How this is funded →
Three ways teams get this wrong
- Assuming the licence travels to the application. Building a CV-screening tool on an open-source model does not make the tool exempt. You are the provider of a high-risk AI system under Article 25, with the full Chapter III burden, and the model’s licence is irrelevant to that. Roles →
- Assuming “open source” as the project describes it meets the conditions. The Chapter V relief is framed around a licence permitting access, use, modification and distribution, and public availability of weights, architecture and usage information. A licence restricting commercial use, or a release with withheld weights, is unlikely to qualify however the project labels itself.
- Assuming the relief is total. Even where it applies, the copyright policy and the sufficiently detailed summary of training content survive. Those are frequently the two obligations an open-source model provider finds hardest. GPAI obligations →
The systemic-risk cliff
Model-level relief evaporates entirely for models presenting systemic risk. Because that classification rests on a compute presumption amendable by delegated act, an open-source model that is comfortably outside today can cross into the full obligation set without the provider changing anything. Expert analysis: worth modelling before a training run, not after.
What actually applies to an open-source project
| If you… | Then… |
|---|---|
| Release a minimal-risk system under a FOSS licence | Art. 2 exemption likely applies |
| Release a system that chats with users | Article 50 applies: disclosure, live now |
| Release a system serving an Annex III purpose | High-risk. Exemption does not apply |
| Release a GPAI model, no systemic risk, weights public | Relief from some Ch. V duties; copyright policy and training summary still apply |
| Release a GPAI model with systemic risk | Full Chapter V, including evaluation, adversarial testing and incident reporting |
| Build a product on someone else’s open model | You are the provider of your system. The upstream licence is irrelevant to your obligations |
Status labels on this page
Verified fact: The Art. 2 system-level exclusion and its three carve-backs; the Chapter V model-level relief, its licence and public-availability conditions, its exclusion for systemic risk, and the survival of the copyright policy and training-content summary duties.
Expert analysis: The three failure modes, the systemic-risk cliff analysis, and the applicability table.
Unsettled: How the licence conditions will be assessed in practice, including for weight-available but use-restricted releases. No guidance or enforcement decisions exist.
The licence does not decide it, the purpose does
Classification turns on the intended purpose of your system, not on how the components inside it are licensed. Run the classification and record the reasoning, including where you conclude the exemption applies.
Not sure where you sit?
The classifier maps your system against Articles 5, 6, 50 and Annex III. Twelve questions, no email.
Frequently asked
Is open-source AI exempt from the EU AI Act?
Only partially, and the exemption is much narrower than commonly assumed. Article 2 provides that the Regulation does not apply to AI systems released under free and open-source licences, but that carve-out does not apply where the system is placed on the market or put into service as a high-risk AI system, or where it falls under Article 5 prohibited practices or the Article 50 transparency obligations. In practice this leaves the exemption covering minimal-risk open-source systems.
Are open-source AI models exempt from the GPAI obligations?
Partially. Providers of general-purpose AI models released under a free and open-source licence that allows access, use, modification and distribution, and whose parameters including weights, architecture and information on model usage are made publicly available, get relief from certain Chapter V obligations. That relief does not extend to models presenting systemic risk, and it does not remove the obligation to put in place a copyright policy or to publish a sufficiently detailed summary of training content.
Does releasing a model as open source avoid the AI Act?
No. Two limits matter. The system-level exemption in Article 2 does not apply to high-risk systems, Article 5 practices or Article 50 systems. The model-level relief in Chapter V does not apply to models with systemic risk and does not cover the copyright policy or training-data summary duties. Someone who builds a high-risk application on an open-source model is a provider of that high-risk system regardless of the licence.
What licence conditions does the GPAI relief require?
The relief is framed around a free and open-source licence that allows access, use, modification and distribution of the model, together with public availability of the model's parameters including weights, information on model architecture, and information on model usage. A licence that restricts commercial use, or a release that withholds weights, is unlikely to meet the conditions. Read the operative wording rather than relying on a project's self-description as open source.
Obligations, article by article
- Art. 5 prohibitions
- Art. 4 AI literacy
- Art. 50 transparency
- Art. 9 risk management
- Art. 10 data governance
- Art. 11 / Annex IV
- Arts. 12–13 logging
- Art. 14 human oversight
- Art. 15 accuracy & security
- Art. 17 QMS
- Arts. 43–48 conformity
- Art. 49 registration
- Art. 57 sandboxes
- Art. 72 monitoring
- Art. 73 incidents
- Arts. 51–56 GPAI
- Art. 99 penalties
- Compliance checklist
- FRIA template (Art. 27)
- When Annex III does not apply →