What it is
The General-Purpose AI Code of Practice was published on 10 July 2025, written by independent experts through a multi-stakeholder process. It is a voluntary tool that GPAI providers can use to show compliance with AI Act Articles 53 and 55. On 1 August 2025 the Commission and the AI Board confirmed it is “an adequate voluntary tool”.
The key point: the Code is not the law. The obligations come from the AI Act; the Code is one way to show you meet them.
The three chapters
| Chapter | Applies to |
|---|---|
| Transparency | All GPAI providers (Art. 53) |
| Copyright | All GPAI providers (Art. 53) |
| Safety and Security | Providers of models with systemic risk (Art. 55) |
The copyright chapter supports Article 53(1)(c), which requires a copyright policy, including identifying and respecting text and data mining opt-outs under the DSM Directive.
Related Commission documents
- GPAI guidelines (18 July 2025): a model is general-purpose if trained with more than 10^23 FLOP and able to generate language, text-to-image, or text-to-video, with open-source exemptions.
- Systemic risk threshold: 10^25 FLOP (Art. 51).
- Training data summary template (24 July 2025): for the public summary under Art. 53(1)(d), partly to help rights holders.
- Fines for GPAI providers: up to 3% of global turnover or EUR 15 million, whichever is higher.
Signatories and dates
The Commission’s page lists 28 signatories, including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI, OpenAI, and WRITER; xAI signed only the Safety and Security chapter. GPAI duties apply from 2 August 2025, and models placed on the market before then have until 2 August 2027.
Reuse terms
Commission digital strategy pages are licensed CC BY 4.0 with attribution to the European Union and a link to the source. Third-party content and logos are excluded.
Using the Code in AI and RAG
- Index the Code apart from the AI Act and label it voluntary.
- Link each Code measure to the AI Act article it supports.
- Store guidelines and templates as separate documents with dates.
- Track signatory status if you rely on a vendor’s commitments.
- Attribute to the European Union in outputs.
Common mistakes
- Treating the Code as mandatory, or as the only way to comply.
- Assuming the Safety chapter applies to all GPAI models.
- Ignoring text and data mining opt-outs in training data.
- Relying on a vendor’s signature without checking which chapters they signed.
Checklist
- Code labeled voluntary and linked to articles
- Guidelines and template stored with dates
- Vendor signatory status checked
- Copyright policy and opt-out handling documented
- Attribution in outputs
How AskDona helps
AskDona checks your model documentation, copyright policy, and training data summary against the Code and the AI Act articles, and shows where each requirement is evidenced.
Run a sample GPAI documentation check
AskDona shows the evidence; your legal and policy teams decide.
Editor’s note. This article is general information, not legal advice. The “conditions to check” rating comes from our source list and is a first-pass judgment, not a legal opinion.
FAQ
Is the GPAI Code of Practice mandatory? No. It is a voluntary tool to show compliance with the AI Act.
What are its chapters? Transparency, Copyright, and Safety and Security.
Who signed it? The Commission lists 28 signatories, including major model providers.
When do GPAI duties apply? From 2 August 2025; existing models have until 2 August 2027.
References
All pages opened and checked by October 26, 2026.
| Source | What it supports |
|---|---|
| Commission: GPAI Code of Practice | Code, chapters, signatories |
| Commission: Assessment of the Code | Adequacy assessment |
| Commission: GPAI guidelines | 10^23 FLOP definition |
| Commission: Training data summary template | Template |
| Commission: GPAI Q&A | 10^25 FLOP, fines |
| Commission: GPAI rules start to apply | Dates |
| AI Act Service Desk: Article 53 | Copyright policy duty |