The rules of the AI Act, which came into effect as of August 1, 2024, will become applicable in several stages. Initially, from February 2, 2025, the provisions related to prohibited AI practices must be applied, and from August 2, 2025, the rules related to general purpose AI models will become applicable (see especially Chapter V of the AI Act), as well as the sections guiding the designation and establishment of authorities and governance systems playing a key role in the implementation of the AI Act (see e.g., Chapter III Section 4, Chapter VII), and the sanctions (Chapter XII). In this post, I will dive deep into the requirements related to general purpose AI models and general purpose AI systems.
1. What are general purpose AI models and general purpose AI systems?
According to the AI Act, a general purpose AI model (GPAI model) means
an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. (AI Act, Article 3, Point 63)
"Whereas the generality of a model could, inter alia, also be determined by a number of parameters, models with at least a billion of parameters and trained with a large amount of data using self-supervision at scale should be considered to display significant generality and to competently perform a wide range of distinctive tasks." (see Recital (98)).
A general purpose AI system (GPAI system) means
an AI system which is based on a general-purpose AI model and which has the capability to serve a variety of purposes, both for direct use as well as for integration in other AI systems. (see AI Act, Article 3, Point 66)
When a general-purpose AI model is integrated into or forms part of an AI system, this system should be considered to be general-purpose AI system when, due to this integration, this system has the capability to serve a variety of purposes. A general-purpose AI system can be used directly, or it may be integrated into other AI systems. (see Recital (100), emphasis added)
(As a reminder: according to the AI Act, an AI system means "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.", see AI Act, Article 3, Point 1. Please also see my blog post about the definition of AI systems here.)
An important question is the distinction between GPAI models and GPAAI systems: "[...] Although AI models are essential components of AI systems, they do not constitute AI systems on their own. AI models require the addition of further components, such as for example a user interface, to become AI systems. AI models are typically integrated into and form part of AI systems. [...]" (see Recital (97), emphasis added).
Another important practical question is when new models are created through the modification or fine-tuning of GPAI models, as this affects the fulfillment of obligations related to the GPAI models and the addressee of these obligations.
The AI Act does not apply to AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development. (see Article 2 (6) of the AI Act), so - according to Recital (97) - "[...] The definition should not cover AI models used before their placing on the market for the sole purpose of research, development and prototyping activities. This is without prejudice to the obligation to comply with this Regulation when, following such activities, a model is placed on the market." (emphasis added)
What could be typical examples of GPAI models?
As stated in the preamble of the AI Act (see Recital (99)), "large generative AI models are a typical example for a general-purpose AI model, given that they allow for flexible generation of content, such as in the form of text, audio, images or video, that can readily accommodate a wide range of distinctive tasks." (emphasis added)
2. How do the rules for general-purpose AI models fit into the AI Act's risk framework?
The AI Act establishes four risk categories for AI systems:
- unacceptable risk (prohibited AI practices),
- high risk (high-risk AI systems),
- limited risk, and
- minimal risk.
The requirements for general-purpose AI models somewhat "stand out" from the above risk-based classification, as they are treated separately by the AI Act. However, additional requirements for general-purpose AI models that pose systemic risks are also defined, meaning that a risk-based approach is also present in this context, and additional requirements are associated with additional risks.
What does systemic risk mean?
According to the AI Act, "systemic risk means a risk that is specific to the high-impact capabilities* of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain." (Article 3, point 65)
*"High-impact capabilities means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models." (Article 3, point 64)
The preamble of the AI Act provides some further guidance on systemic risks:
"General-purpose AI models could pose systemic risks which include, but are not limited to, any actual or reasonably foreseeable negative effects in relation to major accidents, disruptions of critical sectors and serious consequences to public health and safety; any actual or reasonably foreseeable negative effects on democratic processes, public and economic security; the dissemination of illegal, false, or discriminatory content. Systemic risks should be understood to increase with model capabilities and model reach, can arise along the entire lifecycle of the model, and are influenced by conditions of misuse, model reliability, model fairness and model security, the level of autonomy of the model, its access to tools, novel or combined modalities, release and distribution strategies, the potential to remove guardrails and other factors. In particular, international approaches have so far identified the need to pay attention to risks from potential intentional misuse or unintended issues of control relating to alignment with human intent; chemical, biological, radiological, and nuclear risks, such as the ways in which barriers to entry can be lowered, including for weapons development, design acquisition, or use; offensive cyber capabilities, such as the ways in vulnerability discovery, exploitation, or operational use can be enabled; the effects of interaction and tool use, including for example the capacity to control physical systems and interfere with critical infrastructure; risks from models of making copies of themselves or ‘self-replicating’ or training other models; the ways in which models can give rise to harmful bias and discrimination with risks to individuals, communities or societies; the facilitation of disinformation or harming privacy with threats to democratic values and human rights; risk that a particular event could lead to a chain reaction with considerable negative effects that could affect up to an entire city, an entire domain activity or an entire community." (Recital (110), emphasis added)
3. How is the classification of general-purpose AI models as general-purpose AI models posing systemic risk conducted?
The AI Act defines the procedure by which a GPAI model can be classified as a general-purpose AI model posing systemic risk. This classification occurs if the model meets any of the following conditions (see Article 51(1) of the AI Act):
- it has high impact capabilities evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks;
- based on a decision of the Commission, ex officio or following a qualified alert from the scientific panel, it has capabilities or an impact equivalent to those set out in point (a) having regard to the criteria set out in Annex XIII.
A general-purpose AI model shall be presumed to have high impact capabilities pursuant to point (a) above, when the cumulative amount of computation used for its training measured in floating point operations* is greater than 10^25 (see Article 51 (2) of the AI Act). [*"floating-point operation means any mathematical operation or assignment involving floating-point numbers, which are a subset of the real numbers typically represented on computers by an integer of fixed precision scaled by an integer exponent of a fixed base", see Article 3, point 67 of the AI Act] Currently, it is likely that only a few models, such as OpenAI's ChatGPT4 and Google's Gemini models, may reach or exceed the aggregated computational sum of 10^25 measured in floating-point operations (see Moloney - Browne: What’s a FLOP? - How the AI Act Regulates General Purpose AI Systems, CEDPO AI and Data Working Group, Micro-Insights Series, March 2024, p. 6).
According to Annex XIII of the AI Act, for the purpose of determining that a general-purpose AI model has capabilities or an impact equivalent to those set out in Article 51(1), point (a), the Commission shall take into account the following criteria:
- the number of parameters of the model;
- the quality or size of the data set, for example measured through tokens;
- the amount of computation used for training the model, measured in floating point operations or indicated by a combination of other variables such as estimated cost of training, estimated time required for the training, or estimated energy consumption for the training;
- the input and output modalities of the model, such as text to text (large language models), text to image, multi-modality, and the state of the art thresholds for determining high-impact capabilities for each modality, and the specific type of inputs and outputs (e.g. biological sequences);
- the benchmarks and evaluations of capabilities of the model, including considering the number of tasks without additional training, adaptability to learn new, distinct tasks, its level of autonomy and scalability, the tools it has access to;
- whether it has a high impact on the internal market due to its reach, which shall be presumed when it has been made available to at least 10 000 registered business users established in the Union;
- the number of registered end-users.
(The Commission is empowered under the AI Act to adopt legal acts to specify and update the criteria set out in Annex XIII.)
The procedure can be initiated from two directions (see AI Regulation Article 52):
- if a general-purpose AI model meets the relevant condition, the relevant provider shall notify the Commission, (The provider may also present, with its notification, sufficiently substantiated arguments to demonstrate that, exceptionally, although the requirements are met, the GPAI model does not present, due to its specific characteristics, systemic risks and therefore should not be classified as a general-purpose AI model with systemic risk. Where the Commission concludes that the arguments submitted are not sufficiently substantiated and the relevant provider was not able to demonstrate that the GPAI model does not present, due to its specific characteristics, systemic risks, it shall reject those arguments, and the GPAI model shall be considered to be a general-purpose AI model with systemic risk.)
- the Commission – ex officio or following a qualified alert from the scientific body – may designate a general-purpose AI model as a model posing a systemic risk based on the criteria set out in Annex XIII. (If the Commission becomes aware of a GPAI model posing a systemic risk that has not been notified, it may decide to designate it as a model posing a systemic risk.)
The Commission shall ensure that a list of general-purpose AI models with systemic risk is published and shall keep that list up to date.
4. What obligations apply to providers of general-purpose AI models?
According to the AI Act, a "provider means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge" (Article 3, point 3). (I have discussed the "roles" under the AI Act in more detail in this post.)
Providers of general-purpose AI models shall:
-
draw up and keep up-to-date the technical documentation of the model, including its training and testing process and the results of its evaluation, which shall contain, at a minimum, the information set out in Annex XI for the purpose of providing it, upon request, to the AI Office and the national competent authorities;*
-
draw up, keep up-to-date and make available information and documentation to providers of AI systems who intend to integrate the general-purpose AI model into their AI systems. (The information and documentation shall: (i) enable providers of AI systems to have a good understanding of the capabilities and limitations of the general-purpose AI model and to comply with their obligations pursuant to the AI Act; and (ii) contain, at a minimum, the elements set out in Annex XII);*
-
put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790;
-
draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office.
(*The first two obligations above "shall not apply to providers of AI models that are released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available. This exception shall not apply to general-purpose AI models with systemic risks." See Article 53(2) of the AI Act)
Providers of GPAI models may rely on codes of practice (see Article 56 of the AI Act) to demonstrate compliance with the obligations set out above, until a harmonised standard is published. Compliance with European harmonised standards grants providers the presumption of conformity to the extent that those standards cover those obligations. Providers of general-purpose AI models who do not adhere to an approved code of practice or do not comply with a European harmonised standard shall demonstrate alternative adequate means of compliance for assessment by the Commission. (See Article 53 (4) of the AI Act)
Article 56 of the AI Act addresses the role of codes of practice in promoting compliance with the regulation. Given that the rules of the AI Act concerning general-purpose AI models will become applicable on August 2, 2025, codes of practice shall be ready at the latest by 2 May 2025. If, by 2 August 2025, a code of practice cannot be finalised, or if the AI Office deems it is not adequate, the Commission may provide, by means of implementing acts, common rules for the implementation of the obligations provided for in Articles 53 and 55 of the AI Act.
The processes for creating codes of practice have already begun, and the first draft has been published. Update [07.01.2025]: The second draft of the codes of practice was published in December, 2024.
Prior to placing a general-purpose AI model on the Union market, providers established in third countries shall, by written mandate, appoint an authorised representative which is established in the Union (see Article 54 of the AI Act). The obligations of the authorized representative are also defined by the AI Act. Without detailing such obligations, it can be stated that the role of the authorized representative encompasses not only narrowly defined tasks of being a point of contact to the provider but also extends to monitoring and ensuring the fulfillment of the provider's obligations. Moreover, the regulation stipulates that "the authorised representative shall terminate the mandate if it considers or has reason to consider the provider to be acting contrary to its obligations pursuant to this Regulation. In such a case, it shall also immediately inform the AI Office about the termination of the mandate and the reasons therefor." (see Article 54(5) of the AI Act, emphasis added). The obligation to appoint an authorized representative shall not apply to providers of general-purpose AI models that are released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available, unless the general-purpose AI models present systemic risks. (see Article 54(6) of the AI Act)
The AI Act - in addition to the above - establishes further obligations for providers of GPAI models that pose a systemic risk (see Article 55 of the AI Act). Such providers shall:
- perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model with a view to identifying and mitigating systemic risks;
- assess and mitigate possible systemic risks at Union level, including their sources, that may stem from the development, the placing on the market, or the use of general-purpose AI models with systemic risk;
- keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them;
- ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model.
Until a harmonized standard is published, providers of general-purpose AI models that pose a systemic risk can also rely on practical codes of conduct.
5. When must the obligations for general-purpose AI models be fulfilled?
The obligations for general-purpose AI models apply when such models are placed on the market or put into service, meaning they become accessible to parties outside the provider of the model. Accordingly, "[...] the obligations laid down for models should in any case not apply when an own model is used for purely internal processes that are not essential for providing a product or a service to third parties and the rights of natural persons are not affected. Considering their potential significantly negative effects, the general-purpose AI models with systemic risk should always be subject to the relevant obligations under this Regulation. The definition should not cover AI models used before their placing on the market for the sole purpose of research, development and prototyping activities. This is without prejudice to the obligation to comply with this Regulation when, following such activities, a model is placed on the market." (Recital (97), emphasis added).
6. How should the AI Act be applied to general-purpose AI systems?
For general-purpose AI systems, a conceptual element (see point 1 above) is that they are based on a general-purpose AI model. The providers of general-purpose AI systems may be the same as the providers of the used general-purpose AI model or may be separate from them, so the requirements for the applied general-purpose AI model and the requirements for the general-purpose AI system based on it may have different addressees.
As stated in the preamble of the AI Act:
When a general-purpose AI model is integrated into or forms part of an AI system, this system should be considered to be general-purpose AI system when, due to this integration, this system has the capability to serve a variety of purposes. A general-purpose AI system can be used directly, or it may be integrated into other AI systems. (see Recital (100) of the AI Act, emphasis added)
Beyond the detailed requirements for general-purpose AI models, the AI Act also includes some specific provisions for general-purpose AI systems. However, in general, the requirements for AI systems, particularly those for high-risk AI systems, apply to these general-purpose AI systems.
In this context, it is important to highlight that
General-purpose AI systems may be used as high-risk AI systems by themselves or be components of other high-risk AI systems. Therefore, due to their particular nature and in order to ensure a fair sharing of responsibilities along the AI value chain*, the providers of such systems should, irrespective of whether they may be used as high-risk AI systems as such by other providers or as components of high-risk AI systems and unless provided otherwise under this Regulation, closely cooperate with the providers of the relevant high-risk AI systems to enable their compliance with the relevant obligations under this Regulation and with the competent authorities established under this Regulation. (see Recital (85) of the AI Act, emphasis added)
[*See in particular Article 25 of the AI Act regarding responsibilities along the AI value chain.]
"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated." (Article 50 (2) of the AI Act, emphasis added) ("This obligation shall not apply to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof, or where authorised by law to detect, prevent, investigate or prosecute criminal offences.", Article 50 (2) of the AI Act, emphasis added)
7. How is the enforcement of the obligations set out in the AI Act regarding GPAI models carried out?
Regarding providers of general-purpose AI models, the powers related to supervision, investigation, enforcement and monitoring have been fundamentally defined at the EU level. Accordingly, the Commission has exclusive authority over the supervision and enforcement of the rules concerning general-purpose AI models as set out in Chapter V of the AI Actg (see Article 88 of the AI Act). The Commission performs this task through the AI Office. If necessary for the performance of their duties, market surveillance authorities may request the Commission to exercise the relevant powers.
According to the AI Act, the "AI Office means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission." (Article 3, point 47 of the AI Act). Please see my previous post about the AI Office here.)
To ensure compliance with the obligations under the AI Act, the AI Office monitors whether providers of general-purpose AI models comply with the AI Act, including adherence to approved codes of practice. In this context, downstream providers must be allowed to file complaints regarding alleged violations of the AI Act.
According to the AI Act, a "downstream provider means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations." (Article 3, point 68 of the AI Act).
The scientific panel also plays an important role in the supervision of general-purpose AI models, as the panel may provide a qualified alert to the AI Office where it has reason to suspect that:
- a general-purpose AI model poses concrete identifiable risk at Union level; or
- a general-purpose AI model meets the conditions referred to in Article 51 (i.e. it poses systemic risk).
In the exercise of supervisory rights, it is important that the Commission has the authority to request documentation and information.
The AI Office may conduct evaluations of the general-purpose AI model concerned:
- to assess compliance of the provider with obligations under the AI Act (where the information gathered is insufficient); or
- to investigate systemic risks at Union level of general-purpose AI models with systemic risk (in particular following a qualified alert from the scientific panel).
Independent experts can also be involved in the evaluation, and the Commission may request access to the relevant general-purpose AI model concerned through APIs or further appropriate technical means and tools, including source code.
If necessary and appropriate, the Commission may request certain measures for providers to comply with their obligations, including risk mitigation measures that are necessary if the evaluation has raised serious and well-founded concerns about a systemic risk at the Union level. The Commission may also require providers to restrict the making available on the market, withdraw or recall the model. (A structured dialogue with the provider may take place before requesting such measures.)
8. What sanctions can be applied to providers of general-purpose AI models?
In addition to the measures mentioned above, the Commission may impose on providers of general-purpose AI models fines not exceeding 3 % of their annual total worldwide turnover in the preceding financial year or EUR 15 million, whichever is higher (see Article 101 of the AI Act), when the Commission finds that the provider intentionally or negligently:
- infringed the relevant provisions of the AI Act;
- failed to comply with a request for a document or for information pursuant to Article 91, or supplied incorrect, incomplete or misleading information;
- failed to comply with a measure requested under Article 93;
- failed to make available to the Commission access to the general-purpose AI model or general-purpose AI model with systemic risk with a view to conducting an evaluation pursuant to Article 92.
9. How is the AI Act to be applied to general-purpose AI models already placed on the market?
Providers of general-purpose AI models that have been placed on the market before 2 August 2025 shall take the necessary steps in order to comply with the obligations laid down in the AI Act by 2 August 2027. (see Article 111 (3) of the AI Act).
Other posts in the "Deep Dive into the AI Act" series:
