Open-Source AI and the EU AI Act: Two Carve-Outs, Neither a Free Pass
Open-source AI has two EU AI Act carve-outs: Article 2(12) for systems, Article 53(2) for GPAI models. Neither exempts high-risk or prohibited uses.
"It's open source, so the AI Act doesn't apply" is one of the most common — and most expensive — misreadings of Regulation (EU) 2024/1689. The Act does give open-source AI special treatment, but through two separate, narrow carve-outs, and neither exempts the things that carry the heaviest obligations. This page separates them and shows what survives.
Carve-out one: open-source AI systems (Article 2(12))
Article 2(12) provides that the Regulation does not apply to AI systems released under free and open-source licences — except where they are placed on the market or put into service as high-risk systems, or as prohibited practices under Article 5, or as systems subject to Article 50 transparency duties.
Read the exception carefully, because it swallows most of what matters. The carve-out helps genuinely low-risk open-source systems. It does not apply if the system is high-risk, prohibited, or a limited-risk system that interacts with people or generates content. So an open-source CV-screening system is still high-risk; an open-source chatbot still owes Article 50 disclosure.
Carve-out two: open-source GPAI models (Article 53(2))
The second carve-out is for models, not systems. Article 53(2) relieves providers of GPAI models released under a free and open-source licence from two specific duties: the Annex XI technical-documentation obligations and the Annex XII downstream-information obligations under Article 53(1)(a) and (b).
Two limits make this narrow:
- It does not apply to GPAI models with systemic risk (the Article 51 tier) — those keep the full Article 55 obligations.
- The copyright-compliance policy and the training-data summary always survive (Article 53(1)(c) and (d)). Open source never relieves those.
Do not conflate the two
The frequent error is to treat one carve-out as the other. Article 2(12) is about systems and is disapplied for high-risk/prohibited/Article 50 systems. Article 53(2) is about GPAI models and only lifts two documentation duties for non-systemic open models. A provider of an open-source foundation model and a company shipping a product built on it are in different positions under different provisions.
What this means if you build on open-source models
If you take an open-source model — Llama, Mistral or another — and build a system on it, two layers apply. The model layer: the model's provider may benefit from the Article 53(2) carve-out (unless systemic), but the copyright policy and training-data summary remain. The system layer: you are building and shipping the system, so under Article 25 and Article 16 you are likely its provider, and you classify it by its use. Running on an open-source model does not lower your system's risk tier — an open-source-based hiring tool is still Annex III high-risk.
Worked example: a startup fine-tuning an open model
Verda AI, a 25-person company, fine-tunes an open-source language model and ships a customer-facing assistant plus, separately, an internal CV-ranking feature. The open-source licence does not exempt either system: the assistant is limited-risk and owes Article 50 disclosure (Article 2(12) carve-out is disapplied for Article 50 systems), and the CV-ranking feature is Annex III point 4 high-risk with the full provider stack (the carve-out is disapplied for high-risk systems). Verda is the provider of both under Article 25. The model's open-source status helped its original maker with two documentation duties — it did nothing for Verda's obligations.
Timeline and penalties
GPAI model obligations, including the Article 53 duties and the Article 53(2) carve-out, have applied since 2 August 2025. System-level obligations follow the use: Article 5 prohibitions and Article 4 literacy are in force (2 February 2025); Article 50 transparency from 2 August 2026; stand-alone high-risk Annex III from 2 December 2027 under the Digital Omnibus, adopted in June 2026. Penalties: high-risk and most obligations under Article 99(4) (€15 million or 3%); prohibitions under Article 99(3) (€35 million or 7%); GPAI-provider fines under Article 101; the Article 99(6) SME cap applies.
How Confir helps
Confir classifies the system you build on an open-source model by its use, and tells you whether the Article 2(12) carve-out is available or disapplied — so an open-source-based high-risk or Article 50 system is scoped correctly rather than waved through. It keeps the model layer and the system layer distinct, the way the Act does. The engine is deterministic and rule-based, with each finding cited.
Frequently asked questions
Does the EU AI Act exempt open-source AI?
Only narrowly. Article 2(12) exempts open-source AI systems — but not where they are high-risk, prohibited, or subject to Article 50 transparency. Article 53(2) lifts two documentation duties for open-source GPAI models, but not for systemic-risk models, and the copyright policy and training-data summary always survive. Neither is a blanket exemption.
What is the difference between the two open-source carve-outs?
Article 2(12) is about AI systems and is disapplied for high-risk, prohibited, and Article 50 systems. Article 53(2) is about GPAI models and only removes the Annex XI and Annex XII documentation duties for non-systemic open models. One concerns systems, the other concerns models.
If I build on an open-source model, am I exempt?
No. You are building a system, so you are likely its provider under Articles 25 and 16, and you classify it by its use. Running on an open-source model does not lower your system's risk tier. An open-source-based hiring tool is still Annex III high-risk.
Does open source remove the copyright and training-data duties?
No. Even for open-source GPAI models, the Article 53(1)(c) copyright-compliance policy and the Article 53(1)(d) training-data summary always survive. The Article 53(2) carve-out only touches the Annex XI and Annex XII documentation duties.
Are open-source systemic-risk models exempt?
No. The Article 53(2) carve-out does not apply to GPAI models with systemic risk under Article 51. Those keep the full Article 55 obligations regardless of licence.
Related guides
- GPAI model, defined
- GPAI obligations explained
- Provider versus deployer roles
- How AI risk classification works
- Article 50 transparency obligations
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →