Are open-source AI systems exempt from the EU AI Act?

Only partially. Providers of free and open-source AI are exempt from most Chapter III obligations unless the system is high-risk or prohibited. Deployers bear full obligations regardless of source licensing. GPAI models with systemic risk are not exempt.

What security risks are specific to open-source AI models?

Key risks include pickle deserialization attacks in model files, model poisoning through compromised training data, dependency vulnerabilities, and lack of vendor security patching. Organizations must assume full security responsibility.

Can we use Llama models commercially?

The Llama Community License permits commercial use for organizations with fewer than 700 million monthly active users. Above that threshold, a separate commercial agreement with Meta is required. Always verify the specific license version of the model you are using.

What is an AI Bill of Materials and do we need one?

An AI BOM extends the Software Bill of Materials concept to include model components, training data sources, and dependencies. While not yet legally mandated, AI BOMs support EU AI Act technical documentation requirements and security best practices.

Does using an open-source AI model reduce our regulatory obligations as a deployer?

No. The EU AI Act open-source exemption applies only to providers. Deployers bear full obligations under Article 26 regardless of whether the AI system is open source, including human oversight, monitoring, and incident reporting for high-risk systems.

Quick answer

An open source AI policy governs the use, contribution, and distribution of open-source AI models and components, addressing licensing compatibility, security review requirements, contribution approval processes, and the specific EU AI Act obligations that apply even to open-source AI providers.

Updated June 2026 · MmowW AI Compliance

Open Source AI Policy: Licensing, Contribution Rules, and Security Requirements

Open Source AI in the Regulatory Landscape

Open-source AI models (Llama, Mistral, Stable Diffusion, and others) offer cost efficiency, transparency, and customization advantages. However, they introduce governance challenges: licensing complexity, security responsibility shifts from vendor to deployer, unclear liability chains, and regulatory obligations that apply regardless of whether software is open source.

The EU AI Act Article 2(12) provides a partial exemption for open-source AI: providers of free and open-source AI systems are exempt from most Chapter III obligations unless the system is high-risk, prohibited, or subject to Article 50 transparency requirements. This exemption is narrower than many organizations assume.

EU AI Act Open-Source Provisions

AI Act Obligation	Open-Source Exemption	Exception to Exemption
Chapter III high-risk requirements	Exempt for FOSS providers	Applies if system is in Annex III or is a safety component
Article 5 prohibited practices	No exemption	All providers regardless of licensing
Article 50 transparency	No exemption	All providers of deepfake-capable or emotion recognition systems
Articles 51-56 GPAI obligations	Exempt for open-weight models	Applies if model has systemic risk (above FLOP threshold)
Deployer obligations (Art. 26)	No exemption	Deployers bear full obligations regardless of source licensing

License Compliance for AI Models

AI model licenses differ from traditional software licenses. Common AI-specific licenses include:

Apache 2.0: Permissive, allows commercial use, modification, and distribution. Used by many Hugging Face models
MIT: Permissive, minimal restrictions. Common for AI libraries
Llama Community License: Permits commercial use below 700M monthly active users; requires separate agreement above that threshold
CreativeML Open RAIL-M: Includes use restrictions (no harmful applications). Used by Stable Diffusion
GPL v3: Copyleft, requires derivative works to be distributed under the same license. May affect AI systems that incorporate GPL components

Maintain a license inventory for all open-source AI components. Verify license compatibility before integrating components. Pay particular attention to use restriction clauses in RAIL-type licenses that may prohibit specific applications regardless of commercial viability.

Security Requirements for Open-Source AI

When you use open-source AI, you assume security responsibility. Implement:

Supply chain security: Verify model provenance, check for known vulnerabilities in model architectures, scan model files for embedded malicious code (pickle deserialization attacks are a documented risk)
Model integrity: Verify model checksums against trusted sources, use model signing where available, monitor for model poisoning indicators
Dependency management: Track all dependencies, apply security patches promptly, use SBOMs (Software Bill of Materials) extended to include model components (AI BOMs)
Access controls: Restrict who can deploy, modify, or retrain open-source models in production environments

Contribution Governance

If your organization contributes to open-source AI projects, establish approval procedures covering: intellectual property clearance (ensure contributions do not include proprietary data or algorithms), security review (no credentials, internal URLs, or sensitive configurations in contributions), license compatibility verification, and organizational attribution policies.

Require Contributor License Agreements (CLAs) from external contributors to your AI projects to clarify IP ownership and grant rights.

Deployer Obligations Are Not Affected

The EU AI Act's open-source exemption applies only to providers, not deployers. If your organization deploys an open-source AI system in a high-risk context (employment screening, credit scoring, etc.), you bear full deployer obligations under Article 26: fundamental rights impact assessment, human oversight, monitoring, and incident reporting. The fact that the model is open source does not reduce these obligations.

Due Diligence Checklist for Open-Source AI Adoption

License review: Verify commercial use is permitted, understand copyleft implications, check for use restrictions
Security assessment: Scan model files, verify provenance, review dependency chain
Regulatory classification: Determine if the intended use falls within EU AI Act high-risk categories
Performance validation: Conduct independent evaluation; do not rely solely on published benchmarks
Support assessment: Evaluate community activity, maintenance cadence, and whether commercial support is available
Documentation: Verify model card availability, training data documentation, and known limitations disclosure

Ongoing Governance

Open-source AI governance is continuous, not one-time. Monitor for license changes (Meta changed Llama licensing terms between versions), security advisories, model updates, and community health indicators. Maintain the ability to replace any open-source AI component with an alternative within a defined timeframe.

Check your AI compliance readiness — free.

Take the Readiness Check 3 minutes · 10 questions · no signup required

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.