Direct answer

What evidence does a provider of a GPAI model have to be able to show under the AI Act?

6 obligations under the AI Act bear on this, of which 6 apply today.

First step: Record per requirement which standard or specification you rely on, and justify every departure.

To show that a provider of a GPAI model complies with the AI Act, 3 dossiers are required. Below is what belongs in each dossier and which obligation it follows from. Likely role: provider of a gpai model.

The conclusion and your first steps

Preliminary reading · Articles 40 to 42: standards, common specifications and presumption of conformity

What is open is the anchoring of the duty holder. Of the three Articles only Article 41(5) contains a rule of conduct addressed to a role this graph knows: providers of high-risk AI systems or general-purpose AI models must duly justify that they have adopted at least equivalent technical solutions where they do not apply a common specification. The object is anchored on that paragraph. A defensible alternative reading treats Articles 40 to 42 as entirely institutional, with an empty duty holder and `out_of_scope`, and hangs the justification on Article 11 and Annex IV, where the technical documentation is described. On that reading the recommended action and the file below move to the object on Article 11; the substance does not change, the basis does. Also open is where the justification must be recorded. Article 41(5) prescribes no form, no place and no recipient. We read it as belonging in the technical documentation, because that is the only file a market surveillance authority can request under Article 21; that is an inference and not text. Not open is the dating: Articles 40, 41 and 42 sit in Chapter III, Section 5, and that Section is not excepted in the third paragraph of Article 113, so the general application date of the second paragraph governs, 2 August 2026.

The text of the provision is settled. How it plays out has not been decided in case law or guidance, so this is our reasoned reading and not established law. Weigh it yourself, or put it to your own lawyer.

Preliminary reading · Article 4a: legal basis for bias testing with special categories of personal data

Two things are open here, and the risk runs the other way round than with a right such as Article 86: a broad reading here benefits the controller and not the data subject, because this concerns data on ethnicity, health, religion, trade union membership and sexual orientation. Where in doubt the narrow reading is therefore the safe one. First, the reach of "other AI systems and models" in paragraph 2, which on its face covers any AI system and any model and for which no delimitation exists. We read it on its face, but with the threshold in paragraph 2, point (a), as the real boundary: without a consequence for health and safety, fundamental rights or prohibited discrimination there is no basis. A defensible alternative reading is that paragraph 2 is confined to systems comparable to the examples in recital 9, such as scoring tools for permits and public services. Second, the relationship with Article 9 GDPR. Recital 9 states that the extension is subject to the same limitations, conditions and safeguards and thereby ensures compliance with Article 9(2), point (g), of Regulation (EU) 2016/679, so we read Article 4a as the Union law measure that point requires, with the safeguards carried by the six conditions themselves. The counterargument stands against that and has not gone away, but it has narrowed since 27 July 2026: the anchoring sits in a recital and not in the article, and the article itself designates no ground from Article 9(2). What no longer supports that counterargument is Article 2(7). Until 27 July 2026 that paragraph left the GDPR unaffected without reservation, but it was replaced by Article 1, point 2(b), of Regulation (EU) 2026/1744 and now reads: "Union law on the protection of personal data, privacy and the confidentiality of communications applies to personal data processed in connection with the rights and obligations laid down in this Regulation. Without prejudice to Articles 4a and 59 of this Regulation, this Regulation shall not affect Regulation (EU) 2016/679 or (EU) 2018/1725, or Directive 2002/58/EC or (EU) 2016/680." The Union legislature therefore carved the reservation out for Article 4a precisely, which points towards reading Article 4a as the Union law measure itself rather than a mere cross-reference to the GDPR. Anyone citing this object while quoting the former wording of Article 2(7) is quoting a replaced provision. A defensible alternative reading remains that a national or Union measure with specific safeguards is still needed alongside it, but it now rests only on the absence of an express designation in the article itself. On the date from which the basis operates, part is settled and part is not. What is settled is what recital 9 says, namely that the basis should apply from the date of entry into application of Regulation (EU) 2024/1689; that has been read and is not a house reading. What remains open is which date this object therefore carries. We hold to 27 July 2026, the day Article 4a entered the text, because a basis that was not yet there was in fact not available. The alternative reading follows recital 9 literally and lets the basis reach back to the date of application of the base Regulation. That difference is not academic for anyone who has to justify processing from that period. The editorial statement below marks that choice as our inference.

This applies now

Articles 40 to 42: standards, common specifications and presumption of conformityApplicable
Article 4a: legal basis for bias testing with special categories of personal dataApplicable
Article 52: notification of a GPAI model with systemic riskApplicable
Article 53: GPAI model providersApplicable
Article 54: authorised representative of a provider of a GPAI modelApplicable
Article 55: GPAI models with systemic riskApplicable

What you have to be able to show

Coverage matrix and justification for standards and specifications
Per requirement of Section 2: the harmonised standard or common specification applied with its version, the publication status of the reference in the Official Journal, what the standard or specification does and does not cover, and on departure the Article 41(5) justification with the technical solution chosen and the test showing equivalence.
Necessity file for bias testing
Per processing operation: the system or model, the paragraph of Article 4a relied on, the justification why synthetic or anonymised data do not suffice, the technical and organisational safeguards applied, the access list, the confirmation that no other party can reach the data, and the deletion date. This is also the text that paragraph 1, point (f), requires in the record of processing activities.
Systemic-risk notification file
Per model version: the measured and planned training compute with the scope of recital 111, so including pre-training, synthetic data generation and fine-tuning, the moment the threshold was reached or foreseen, the notification sent with its supporting information, any arguments under paragraph 2, any reassessment request under paragraph 5, and the response or designation decision of the Commission.

A dossier is not a document but a demonstrable whole: who maintains it, where it lives and when it was last updated.

Heavy fine-tuning makes you the model provider
A European scale-up fine-tunes an existing general-purpose AI model for its own product, using more compute than one third of the compute used to train the original model.
Provenance: The Commission guidelines on the scope of the GPAI obligations address this case when determining when someone becomes a model provider themselves. The document is non-binding.
Estimate your fine-tuning compute against the original model before you start, because exceeding one third of it makes you the provider, with documentation, copyright and training-content duties limited to your modification.
Commission Guidelines C(2025) 5045 final, 18.7.2025, scope of the obligations for GPAI model providers
financial services
Light fine-tuning does not make you a model provider
A bank fine-tunes an existing general-purpose AI model on its own product documentation and customer questions, using a fraction of the original compute, and builds a customer chatbot around it that it offers under its own name.
Provenance: The Commission guidelines on the scope of the GPAI obligations address this case when determining when someone becomes a model provider themselves. The document is non-binding.
If your fine-tuning stays well below one third of the original compute you do not become the model provider, but the obligations for the AI system you offer under your own name still apply.
Commission Guidelines C(2025) 5045 final, 18.7.2025, scope of the obligations for GPAI model providers
Modifying a systemic-risk model pulls the heaviest duties to you
A downstream actor modifies an existing systemic-risk model so substantially that the change exceeds the threshold, and publishes the result as its own model.
Provenance: The Commission guidelines on the scope of the GPAI obligations address this case when determining when someone becomes a model provider themselves. The document is non-binding.
Where a modification of a systemic-risk model crosses the threshold, the result is presumed to have high-impact capabilities, so estimate the compute in advance and notify the Commission within two weeks.
Commission Guidelines C(2025) 5045 final, 18.7.2025, scope of the obligations for GPAI model providers
technology and software
Open source in name, but not within the meaning of the Regulation
Three providers call their model open source. The first licence allows non-commercial research only. The second requires a separate commercial licence once monthly active users pass a threshold. The third gives the model away for free but hosts it exclusively on its own platform where visitors are served paid advertisements.
Provenance: The Commission guidelines on the scope of the GPAI obligations address this case when determining when someone becomes a model provider themselves. The document is non-binding.
A usage restriction is not automatically fatal: you may include specific, proportionate and non-discriminatory safety terms, whereas a monthly active user threshold or a separate commercial licence disqualifies the licence.
Commission Guidelines C(2025) 5045 final, 18.7.2025, scope of the obligations for GPAI model providers

When a downstream party that fine-tunes becomes a GPAI provider itself
The Commission guidelines of 18 July 2025 (C(2025) 5045 final) on the scope of the obligations for providers of general-purpose AI models state in point (61) that it is not necessary for every modification of such a model to lead to the downstream modifier being considered the provider of the modified model, in line with the Blue Guide, which states that a product subject to important changes or overhauls aiming to modify its original performance, purpose or type may be considered a new product. Point (62) states that the Commission considers a downstream modifier to become the provider of the modified model only if the modification leads to a significant change in the model's generality, capabilities or systemic risk. Point (63) sets the indicative criterion: a downstream modifier is considered to be the provider where the training compute used for the modification is greater than a third of the training compute of the original model. Point (64) states that where the downstream modifier cannot be expected to know that value, for example because it has not been communicated by the provider of the original model, and cannot estimate it, the threshold is replaced by a third of 10 to the power of 25 FLOP where the original model is a model with systemic risk, and otherwise by a third of 10 to the power of 23 FLOP. Point (65) explains that a modification of that size is expected to display a significant change justifying the transparency obligations of Article 53(1)(a) and (b), that such a modification can be expected to have used a significant amount of data relevant to the copyright policy and the public summary of training content under Article 53(1)(c) and (d), and that where the original model has systemic risk the modified model can be expected to present significantly different systemic risk. Point (67) notes that currently few modifications meet this criterion, that the number of downstream modifiers becoming providers may increase over time, and that the criterion is thus primarily forward-looking. Point (68) states that in the case of a modification the obligations are limited to that modification: the documentation under Article 53(1)(a) and (b) is limited to information on the modification, and the copyright policy under point (c) and the summary of training content under point (d) are limited to the data used as part of the modification. Point (69) states that a downstream modifier who becomes a provider must also comply with Article 54, which means appointing an authorised representative established in the Union to the extent that the modifier is itself established outside the Union. Point (70) states that where a downstream actor modifies a model classified as having systemic risk in such a way that they become the provider of the modified model, the resulting model is presumed to have high-impact capabilities and is therefore considered a model with systemic risk, and point (71) states that the modifier must then comply with the obligations for providers of models with systemic risk and notify the Commission in line with Article 52(1).
Commission Guidelines C(2025) 5045 final, 18.7.2025, Section 3.2 points (60) to (67) and Sections 3.2.1 and 3.2.2, points (68) to (71)

Read the official sourceChecked on 8 August 2026

General interpretation, not legal advice. Checked against Regulation (EU) 2024/1689 and the Digital Omnibus (EU) 2026/1744; the official source remains authoritative.

Full map for your situation

What evidence does a provider of a GPAI model have to be able to show under the AI Act?

This applies now

What you have to be able to show

Follow-up questions

Source and locator

Implement the GPAI provider obligations for each model

Zahed Ashkara

What evidence does a provider of a GPAI model have to be able to show under the AI Act?

This applies now

What you have to be able to show

How this looks in practice4 practice examples

What the Commission says1 guidance statement

Follow-up questions

Source and locator

Implement the GPAI provider obligations for each model

Zahed Ashkara