EU AI Act GPAI obligations: model provider or downstream user?
By Sam Rivera, Founder, SentinelPanda · October 24, 2026 · 3 min read · AI Governance
If you call a model through an API, Chapter V is mostly someone else's problem. If you modify one, read on.
What counts as a general-purpose AI model
The Act defines a general-purpose AI model by its generality: it performs a wide range of distinct tasks competently and can be integrated into many downstream systems. The Commission's guidelines on GPAI providers, published in July 2025, use an indicative threshold of training compute above 10^23 FLOP combined with the ability to generate language, or images or video from text. Large language models and image generators from the major labs are squarely inside; a narrow classifier you trained is not.
Provider obligations under Article 53
- Technical documentation of the model, including training, testing and evaluation results, available to the AI Office and national authorities on request.
- Information and documentation for downstream providers who integrate the model, covering its capabilities and limitations.
- A policy for complying with EU copyright law, including respecting machine-readable text-and-data-mining opt-outs.
- A publicly available, sufficiently detailed summary of the content used for training, using the AI Office template.
- Free and open-source models with public weights are exempt from the first two, unless they carry systemic risk.
Systemic-risk models
A model is presumed to have systemic risk when its cumulative training compute exceeds 10^25 FLOP, or when the Commission designates it. Its provider must notify the Commission within two weeks of meeting the threshold, and then carry extra duties under Article 55: model evaluation including adversarial testing, assessment and mitigation of systemic risks at Union level, tracking and reporting serious incidents to the AI Office without undue delay, and adequate cybersecurity for the model and its infrastructure.
The Code of Practice
The General-Purpose AI Code of Practice, published on 10 July 2025, has chapters on transparency, copyright, and safety and security. Signing it is voluntary and is a recognised way to demonstrate compliance with Articles 53 and 55. Whether your model provider has signed is a useful due-diligence question in its own right.
Where most companies actually sit
If you build a product on a model accessed through an API, you are not a GPAI model provider. You are the provider of an AI system, and possibly also its deployer, and your obligations depend on how that system is classified — high-risk or not — plus the Article 50 transparency duties.
Modification is where it changes. According to the Commission's guidelines, a downstream organisation that modifies a model using more than a third of the compute originally used to train it is treated as the provider of the modified model, with obligations limited to the modification. Ordinary fine-tuning sits far below that line, but it is worth recording the compute you used so you can show it.
Dates and enforcement
- 2 August 2025: Chapter V obligations apply to models placed on the market from that date.
- 2 August 2026: the Commission's enforcement powers apply, including fines of up to 3% of worldwide annual turnover or €15 million, whichever is higher.
- 2 August 2027: deadline for models already on the market before 2 August 2025.
What to ask your model provider
- The downstream documentation they owe you under Article 53, covering capabilities, limitations and intended uses.
- Their copyright policy and training-content summary.
- Whether they have signed the Code of Practice.
- How they will notify you of serious incidents or material changes to the model.
- Contractual terms on whether your prompts and outputs are used for training.