It’s easy to assume every EU AI Act obligation is on the same deferred timeline as high-risk systems. General-purpose AI (GPAI) model obligations are not. They’ve applied since 2 August 2025, they weren’t touched by the Digital Omnibus, and they sit on a completely separate legal track from the Annex III high-risk classification most compliance teams focus on. If your company builds, fine-tunes, or provides access to a foundation model, this is the obligation set that already applies to you, right now — regardless of what any downstream application does with it.
What Counts as a GPAI Model
The Act’s GPAI obligations attach to the model, not the application built on top of it — this is what makes it a distinct track from Annex III risk classification, which attaches to systems and their use cases. A general-purpose AI model is one trained on broad data at scale, capable of competently performing a wide range of distinct tasks, and that can be integrated into a variety of downstream systems and applications. If you provide a foundation model — whether proprietary, open-weight, or fine-tuned from a base model you release — you are likely a GPAI provider under the Act, even if you have no visibility into how a given customer eventually deploys it.
Baseline Obligations for All GPAI Providers
Article 53 sets four baseline duties:
1. Maintain technical documentation. Keep up-to-date documentation of the model, including its training and testing process and evaluation results, containing at minimum the information specified in Annex XI — for provision to the AI Office and national authorities on request.
2. Provide downstream documentation to integrators. Make available information and documentation to providers who intend to integrate your model into their own AI systems, sufficient for them to understand the model’s capabilities and limitations and meet their own obligations under the Act — while still respecting IP rights and trade secrets. This is the documentation pack your customers will ask for when they classify their own downstream systems.
3. Maintain a copyright policy. Put in place a policy to comply with EU copyright law, including identifying and complying with rights reservations.
4. Publish a training content summary. Make publicly available a sufficiently detailed summary of the content used to train the model, according to a template provided by the AI Office.
The Higher Bar: Systemic Risk
A subset of GPAI models are classified as having systemic risk — broadly, models with high-impact capabilities, assessed against criteria including computational scale used in training. Article 55 adds obligations on top of the baseline Article 53 duties for these models:
- Standardised model evaluation, including documented adversarial testing aimed at identifying and mitigating systemic risks
- Assessing and mitigating systemic risks at EU level — not just risks to the provider’s own users, but broader risks the model’s availability creates across the Union
- Serious incident tracking and reporting to the AI Office and relevant national authorities without undue delay
- Cybersecurity protection for both the model itself and the physical infrastructure it runs on
Most mid-market companies are not training models at the scale that triggers systemic-risk classification — that threshold is aimed at the handful of frontier-scale labs. But if you fine-tune or otherwise modify a systemic-risk base model in ways that materially change its capabilities, it’s worth confirming with counsel whether that modification transfers any of the systemic-risk provider obligations to you.
Where This Intersects With Your Own Compliance Work
Two practical implications for teams focused on the high-risk/Annex III track:
If you’re a GPAI provider, you likely also have downstream Annex III exposure. A foundation model provider that also ships a specific application built on that model (a hiring-assistant product, a credit-risk tool) is a GPAI provider and potentially a high-risk system provider for that application, evaluated separately under Annex III. The two obligation sets stack — meeting Article 53 doesn’t satisfy Annex III requirements for the downstream product.
If you integrate someone else’s GPAI model, you need their Article 53(1)(b) documentation pack — not their marketing page. When you classify a downstream system built on a third-party foundation model, the model provider’s technical documentation is a direct input into your own Annex IV documentation and provider/deployer analysis. Treat “ask the model vendor for their GPAI documentation pack” as a standing line item in vendor onboarding, not a one-time question.
What to Do Now
- Determine if you’re a GPAI provider — if you release or provide access to a foundation model under your own name, you are, regardless of whether you consider yourself an “AI company.”
- Check systemic-risk status against the Commission’s published criteria and notification thresholds, and flag it for legal review if you’re near the boundary.
- Confirm your documentation and training-summary obligations are current — these have been enforceable for a year already.
- Separately track your Annex III exposure for any application built on top of the model — GPAI compliance and high-risk system compliance are two different checklists that both apply if you’re doing both.
Aikraft’s system inventory tracks both tracks side by side, so a GPAI model and the applications built on it don’t get conflated into a single, incomplete classification. See how classification works or take the free risk quiz to get a first read on your own systems.