GPAI providers vs deployers: definitions, role-switch rules, and obligation split
As of 18 November 2026 · checked against Regulation (EU) 2024/1689, Regulation (EU) 2026/1744 and the Commission GPAI guidelines · Deutsch
Information only, not legal advice. The provider/deployer split has direct compliance consequences and changes when you fine-tune. Confirm role attribution with qualified counsel.
The AI Act allocates obligations by role. A provider builds and places on the market. A deployer puts the system to use. For general-purpose AI models the line gets blurry fast: most modern stacks involve fine-tuning, retrieval augmentation, agent scaffolding, and re-distribution under a new brand. Article 25 explicitly turns deployers into providers under defined conditions. This page sets out the definitions, the role-switch triggers, and what each side owes for GPAI specifically.
The two definitions, side by side
| Aspect | Provider (Art. 3(3)) | Deployer (Art. 3(4)) |
|---|---|---|
| Action | Develops or has developed an AI system or GPAI model | Uses the system under its own authority |
| Trigger | Places on the market or puts into service under its own name or trademark, for payment or free of charge | Any use except in the course of a personal, non-professional activity |
| Scope | Includes free of charge and open source where placed on the EU market | Includes public sector and private sector |
| Liability axis | Product-side (Art. 16, Art. 53 for GPAI, Art. 55 for systemic risk) | Use-side (Art. 26, Art. 27 FRIA where applicable) |
When a deployer becomes a provider: Article 25
Article 25 turns a downstream actor into a provider of a high-risk system in three situations:
- Rebranding (Art. 25(1)(a)). Putting its name or trademark on a high-risk system already on the market or in service, without prejudice to contracts that allocate the obligations differently.
- Substantial modification (Art. 25(1)(b)). A change not foreseen in the initial conformity assessment that affects compliance with the high-risk requirements or modifies the intended purpose (Art. 3(23)), where the system remains high-risk.
- Changing the intended purpose (Art. 25(1)(c)). Modifying the intended purpose of a system that is not high-risk, including a general-purpose AI system, so that it becomes high-risk.
When one of these cases occurs, the initial provider is no longer the provider of that specific system. As amended by the Digital Omnibus, Art. 25(2) requires it to cooperate with the new provider: technical documentation, known limitations and failure modes, and targeted technical access for testing. That duty does not apply if the initial provider clearly stated that its system is not to be changed into a high-risk system. A downstream provider that integrates a GPAI model becomes provider of its own system, not of the model. It relies on the information the GPAI provider must supply under Art. 53(1)(b) and Annex XII.
Obligation split for GPAI specifically
| Obligation | GPAI provider | Downstream provider integrating GPAI | Pure deployer |
|---|---|---|---|
| Technical documentation (Art. 53(1)(a)) | Yes | No; for its own high-risk system Art. 11 applies | No |
| Information for downstream providers (Art. 53(1)(b)) | Yes | Receives it | No |
| Copyright policy (Art. 53(1)(c)) | Yes | No | No |
| Training-data summary (Art. 53(1)(d)) | Yes | No | No |
| Systemic-risk obligations (Art. 55) | If above 10^25 FLOP threshold or designated | No | No |
| Conformity assessment of integrated high-risk system | If integrating | Yes | No |
| Article 26 deployer duties | Where also deployer | Where also deployer | Yes |
| FRIA (Art. 27) | Where also deployer in scope | Where also deployer in scope | Public bodies, private entities providing public services, and deployers under Annex III, points 5(b) and (c) |
Fine-tuning: when does it flip the role
This is the hard question for builders. The Commission guidelines on GPAI providers answer it with a compute criterion (Commission Q&A):
- Modification below one third of the original training compute. You do not become the provider of a GPAI model. Most fine-tuning, adaptation, retrieval and prompt work falls here.
- Modification above one third of the original training compute. You become the provider of the modified GPAI model. Your obligations are limited to documenting the modification, for example a training content summary covering only the new data.
- Systemic risk. A model is presumed to have high-impact capabilities above 10^25 FLOP cumulative training compute (Art. 51(2)). Check the guidelines for how a modifier is treated when the modified model is in that tier.
- Building a system with an Annex III purpose on top of a model. You become provider of a high-risk AI system, regardless of the compute you spent on the model.
Decision pattern: pin the role at release time
function classifyRole(release) {
if (release.isOriginalGPAI) return 'gpai-provider';
if (release.modificationComputeShare > 1/3) return 'gpai-provider-of-modified-model';
if (release.cumulativeFLOP >= 1e25) return 'gpai-provider-systemic';
if (release.integratedIntoAnnexIIIPurpose) return 'high-risk-provider';
if (release.usesUnderOwnAuthority) return 'deployer';
return 'undetermined-escalate';
} What we track as events
The gpai topic covers Article 53/55 developments, the roles sub-topic covers Article 25 interpretations and Commission guidance on substantial modification. Subscribe with both filters to see role-switch implications cleanly.
Sources
- Regulation (EU) 2024/1689, Articles 3, 25, 26, 27, 51, 53, 55: EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus), amended Article 25: EUR-Lex
- Commission Q&A on the guidelines for GPAI providers: digital-strategy.ec.europa.eu
→ See live updates: https://aiactradar.com
→ Subscribe to webhook: https://docs.aiactradar.com/quickstart