Why GPAI gets its own rulebook
The risk-tier logic from session b1 classifies AI systems by what they do. But a foundation model does not do any one thing - it can be pointed at a thousand tasks downstream. So the Act treats general-purpose AI models as their own category with their own obligations, sitting parallel to the system tiers. If you train, fine-tune, or rebrand such a model, this part of the Act is yours. If you only call one through an API, most of it belongs to your supplier - but you still need to know what to ask them for.
Baseline GPAI obligations 9 min live
Every provider of a general-purpose AI model owes four baseline duties under Article 53 - two documentation duties, a copyright policy, and a public training-content summary. Two of them relax for open-source models, but never for systemic-risk ones.
LiveDuty 1 - technical documentation (Annex XI)2 min▶
The provider keeps up-to-date technical documentation of the model and makes it available to the AI Office and national authorities on request. Annex XI sets out what goes in it: the training and testing process, evaluation results, design choices, and the model's capabilities and limitations.
- Who reads it: the regulator, not the public. It is your evidence that you built and tested the model responsibly.
- Keep it current: "up-to-date" is doing real work here - it must track material changes to the model.
LiveDuty 2 - documentation for downstream providers (Annex XII)2 min▶
Separately, the provider gives documentation to the companies that will build AI systems on top of the model. Annex XII covers what these downstream providers need: the model's capabilities and limitations, how to integrate it, and what it can and cannot safely be used for.
- Who reads it: the developers downstream - so they can comply with their own obligations.
- The point: obligations flow down the chain. A downstream team cannot govern what it was never told.
If you are the deployer calling the model, this is the document you should be asking your supplier for.
LiveDuty 3 - a copyright policy2 min▶
The provider puts in place a policy to comply with EU copyright law. In particular, it must honour the text-and-data-mining (TDM) opt-outs that rights-holders can exercise under Article 4(3) of the Copyright Directive (Directive 2019/790).
- What it means in practice: if a rights-holder has reserved their content from data mining, your training pipeline must respect that reservation.
- Why it is here: foundation models are trained on vast scraped corpora - this is the Act's hook to keep that lawful.
LiveDuty 4 - a public training-content summary2 min▶
The provider publishes a "sufficiently detailed summary" of the content used to train the model, following a template provided by the AI Office. This one is public - it is the transparency counterweight to models trained on data nobody can see.
- Not a full dataset dump: a summary, at the level of detail the template asks for - enough for rights-holders and the public to understand what went in.
- The template matters: the AI Office sets the shape, so this is not a free-form disclosure.
LiveThe open-source relaxation - and its hard limit2 min▶
- What relaxes: for a free and open-source model released with public parameters and architecture, the first two documentation duties (Annex XI and XII) are relaxed.
- What never relaxes: the copyright policy and the public training-content summary stay - open or closed.
- The hard limit: the open-source relaxation does NOT apply to models with systemic risk. Cross that threshold and every duty is back on, plus the Article 55 tier.
The systemic-risk tier 9 min live
The largest models get extra scrutiny. A model is presumed to carry systemic risk once its training compute crosses 10^25 FLOP - and then a heavier set of duties under Article 55 attaches on top of the baseline.
LiveThe 10^25 FLOP threshold (Article 51)2 min▶
- The presumption: a GPAI model is presumed to have high-impact capabilities - and therefore systemic risk - when the cumulative compute used for its training exceeds 10^25 floating-point operations (FLOP).
- The other route: the Commission can also designate a model as systemic based on other criteria, and it can update the threshold as the field moves.
- Scale: this catches a handful of frontier models. Almost nothing a typical company trains gets near it.
LiveThe four extra obligations (Article 55)3 min▶
On top of the baseline duties, a systemic-risk model provider must:
- Evaluate the model - including adversarial testing and red-teaming to find and document risks.
- Assess and mitigate systemic risks at Union level, across the model's lifecycle.
- Track and report serious incidents to the AI Office (and relevant authorities), with corrective measures.
- Ensure cybersecurity of the model and its physical infrastructure.
These are ongoing, operational duties - not a one-time filing. They mirror the kind of frontier-safety work the largest labs already publish.
LiveProvider or deployer? The role that decides everything2 min▶
- Most companies are deployers of GPAI, not providers - you call a model through an API or embed a licensed one, and the provider's obligations sit with the lab.
- But you can inherit them: if you materially fine-tune a model, or put a model on the market under your own name or brand, you can become a provider - with the Article 53 (and possibly 55) duties.
- The honest test: "am I presenting this as my own model, or meaningfully changing what it is?" If yes, assume you have picked up provider obligations.
Vantage builds on a third-party LLM. Vantage's hiring tool calls a commercial foundation model through an API - it does not train or materially fine-tune it. So the LLM vendor is the GPAI provider and carries the Article 53/55 duties; Vantage is a downstream deployer. Vantage's job is to demand the Annex XII documentation from the vendor and fold it into its own high-risk system file. If Vantage later fine-tuned the model on its own candidate data and shipped it as "Vantage AI", that calculus would change - it could become a provider too.
Which hat are you wearing? ★ 10 min
Decide, for one real use of a foundation model, whether it makes you a GPAI provider or a downstream deployer - and what you owe as a result.
Name the model + your use. Which foundation model do you rely on, and how - API call, embedded license, or your own training/fine-tuning?
Apply the two flip tests. Are you putting it out under your own name? Are you materially fine-tuning it? A yes to either points to provider; two no's point to deployer.
List what you owe. If deployer: the documentation you must get from your provider (Annex XII) and fold into your own system file. If provider: the four baseline duties, plus the Article 55 tier if you are anywhere near systemic scale.
Reflect. In one line: are you certain your vendor actually meets its GPAI duties - and have you asked them for the evidence?
Before session b5 ◐ 40 min total
- For every foundation model your team uses, record whether you are a provider or a deployer, and the reasoning (name, fine-tuning, rebranding).
- For each model you deploy, list the documentation you should obtain from the provider and note whether you actually have it.
- If any model you touch could be a provider situation, sketch the four baseline duties against it and flag gaps.
- Read: EU AI Act Articles 51, 53, and 55 - the GPAI definition, baseline duties, and systemic-risk obligations.
Official sources covered
Taught from the EU AI Act text. This page covers ~80% of the working content on the GPAI regime.
Three questions before you go 🎯 ◐ 90 seconds
1 · Which baseline GPAI duty is public (not just filed with regulators)?
The training-content summary is published following the AI Office template. The Annex XI docs go to the AI Office/authorities on request, and Annex XII docs go to downstream providers - neither is public.
2 · A GPAI model is presumed to carry systemic risk when...
Article 51 sets a presumption of systemic risk above 10^25 FLOP of training compute; the Commission can also designate a model and update the threshold. Open/closed and user counts are not the test.
3 · Your team calls a commercial LLM through an API without fine-tuning it. You are...
Calling a model as-is makes you a deployer; the vendor is the GPAI provider. You would only inherit provider duties by materially fine-tuning the model or putting it out under your own name.