EU AI Act
Is my model a General Purpose AI Model under Article 53?
6 questions to determine whether your AI model qualifies as a GPAI under the EU AI Act and what obligations Article 53 imposes on you.
Last updated: 1 May 2025
Do GPAI model obligations need to comply with EU AI Act?
6 questions to determine whether your AI model qualifies as a GPAI under the EU AI Act and what obligations Article 53 imposes on you. If yes: GPAI with systemic risk — enhanced obligations apply. If not: Model likely below GPAI threshold — standard rules apply. Use the interacti…
- Yes path: GPAI with systemic risk — enhanced obligations apply
- No path: Model likely below GPAI threshold — standard rules apply
- Use the step-by-step decision tree below for your exact situation
EU AI Act · Question 1
Is your model capable of serving a wide range of distinct tasks across many domains, rather than being trained for a single specific purpose?
GPAIs can write code, summarise documents, answer questions, generate images — across domains. A fraud-detection model for banking only is not a GPAI. GPT-4, Claude, Llama, Mistral, Gemini are GPAIs.
For informational purposes only. Consult qualified legal counsel before making compliance decisions.
Decision Criteria & Full Path Breakdown
Review all branch conditions, regulatory criteria, and outcomes assessed by this decision tree.
- 1
Is your model capable of serving a wide range of distinct tasks across many domains, rather than being trained for a single specific purpose?
GPAIs can write code, summarise documents, answer questions, generate images — across domains. A fraud-detection model for banking only is not a GPAI. GPT-4, Claude, Llama, Mistral, Gemini are GPAIs.
If Yes: Proceeds to: Was your model trained using significant computational resources — more than 10^23 FLOPs?If No: Not a GPAI — narrow task-specific model - 2
Was your model trained using significant computational resources — more than 10^23 FLOPs?
10^23 FLOPs is the lower threshold triggering GPAI classification. Most frontier models exceed this. For reference, training GPT-3 used approximately 3.14 × 10^23 FLOPs.
If Yes: Proceeds to: Is your model released as open source with publicly available weights?If No: Model likely below GPAI threshold — standard rules apply - 3
Is your model released as open source with publicly available weights?
Open-source GPAI models with published weights have a lighter compliance path — they are exempt from some documentation requirements under Art. 53(2), unless they present systemic risk.
If Yes: Proceeds to: Was your model trained with more than 10^25 FLOPs — indicating potential systemic risk?If No: Proceeds to: Do you make your GPAI available to other businesses or developers who build downstream applications on top of it? - 4
Was your model trained with more than 10^25 FLOPs — indicating potential systemic risk?
10^25 FLOPs is the systemic-risk threshold under Art. 51. Models above this threshold face enhanced obligations regardless of open-source status.
If Yes: GPAI with systemic risk — enhanced obligations applyIf No: Open-source GPAI — reduced obligations, baseline compliance needed - 5
Do you make your GPAI available to other businesses or developers who build downstream applications on top of it?
If you provide an API, SDK, or model weights to third-party providers who embed your model in their products, you have additional information obligations toward those downstream providers.
If Yes: GPAI provider with downstream users — full Art. 53 obligationsIf No: GPAI for internal use — Art. 53 still applies