Preamble 111 to 120.
(111) It is appropriate to establish a methodology for the classification of general-purpose AI models as general-purpose AI model with systemic risks. Since systemic risks result from particularly high capabilities, a general-purpose AI model should be considered to present systemic risks if it has high-impact capabilities, evaluated on the basis of appropriate technical tools and methodologies, or significant impact on the internal market due to its reach.
High-impact capabilities in general-purpose AI models means capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models. The full range of capabilities in a model could be better understood after its placing on the market or when deployers interact with the model.
According to the state of the art at the time of entry into force of this Regulation, the cumulative amount of computation used for the training of the general-purpose AI model measured in floating point operations is one of the relevant approximations for model capabilities.
The cumulative amount of computation used for training includes the computation used across the activities and methods that are intended to enhance the capabilities of the model prior to deployment, such as pre-training, synthetic data generation and fine-tuning. Therefore, an initial threshold of floating point operations should be set, which, if met by a general-purpose AI model, leads to a presumption that the model is a general-purpose AI model with systemic risks.
This threshold should be adjusted over time to reflect technological and industrial changes, such as algorithmic improvements or increased hardware efficiency, and should be supplemented with benchmarks and indicators for model capability. To inform this, the AI Office should engage with the scientific community, industry, civil society and other experts.
Thresholds, as well as tools and benchmarks for the assessment of high-impact capabilities, should be strong predictors of generality, its capabilities and associated systemic risk of general-purpose AI models, and could take into account the way the model will be placed on the market or the number of users it may affect. To complement this system, there should be a possibility for the Commission to take individual decisions designating a general-purpose AI model as a general-purpose AI model with systemic risk if it is found that such model has capabilities or an impact equivalent to those captured by the set threshold.
That decision should be taken on the basis of an overall assessment of the criteria for the designation of a general-purpose AI model with systemic risk set out in an annex to this Regulation, such as quality or size of the training data set, number of business and end users, its input and output modalities, its level of autonomy and scalability, or the tools it has access to. Upon a reasoned request of a provider whose model has been designated as a general-purpose AI model with systemic risk, the Commission should take the request into account and may decide to reassess whether the general-purpose AI model can still be considered to present systemic risks.
(112) It is also necessary to clarify a procedure for the classification of a general-purpose AI model with systemic risks. A general-purpose AI model that meets the applicable threshold for high-impact capabilities should be presumed to be a general-purpose AI models with systemic risk. The provider should notify the AI Office at the latest two weeks after the requirements are met or it becomes known that a general-purpose AI model will meet the requirements that lead to the presumption.
This is especially relevant in relation to the threshold of floating point operations because training of general-purpose AI models takes considerable planning which includes the upfront allocation of compute resources and, therefore, providers of general-purpose AI models are able to know if their model would meet the threshold before the training is completed. In the context of that notification, the provider should be able to demonstrate that, because of its specific characteristics, a general-purpose AI model exceptionally does not present systemic risks, and that it thus should not be classified as a general-purpose AI model with systemic risks.
That information is valuable for the AI Office to anticipate the placing on the market of general-purpose AI models with systemic risks and the providers can start to engage with the AI Office early on. That information is especially important with regard to general-purpose AI models that are planned to be released as open-source, given that, after the open-source model release, necessary measures to ensure compliance with the obligations under this Regulation may be more difficult to implement.
(113) If the Commission becomes aware of the fact that a general-purpose AI model meets the requirements to classify as a general-purpose AI model with systemic risk, which previously had either not been known or of which the relevant provider has failed to notify the Commission, the Commission should be empowered to designate it so.
A system of qualified alerts should ensure that the AI Office is made aware by the scientific panel of general-purpose AI models that should possibly be classified as general-purpose AI models with systemic risk, in addition to the monitoring activities of the AI Office.
(114) The providers of general-purpose AI models presenting systemic risks should be subject, in addition to the obligations provided for providers of general-purpose AI models, to obligations aimed at identifying and mitigating those risks and ensuring an adequate level of cybersecurity protection, regardless of whether it is provided as a standalone model or embedded in an AI system or a product.
To achieve those objectives, this Regulation should require providers to perform the necessary model evaluations, in particular prior to its first placing on the market, including conducting and documenting adversarial testing of models, also, as appropriate, through internal or independent external testing.
In addition, providers of general-purpose AI models with systemic risks should continuously assess and mitigate systemic risks, including for example by putting in place risk-management policies, such as accountability and governance processes, implementing post-market monitoring, taking appropriate measures along the entire model’s lifecycle and cooperating with relevant actors along the AI value chain.
(115) Providers of general-purpose AI models with systemic risks should assess and mitigate possible systemic risks. If, despite efforts to identify and prevent risks related to a general-purpose AI model that may present systemic risks, the development or use of the model causes a serious incident, the general-purpose AI model provider should without undue delay keep track of the incident and report any relevant information and possible corrective measures to the Commission and national competent authorities. Furthermore, providers should ensure an adequate level of cybersecurity protection for the model and its physical infrastructure, if appropriate, along the entire model lifecycle.
Cybersecurity protection related to systemic risks associated with malicious use or attacks should duly consider accidental model leakage, unauthorised releases, circumvention of safety measures, and defence against cyberattacks, unauthorised access or model theft. That protection could be facilitated by securing model weights, algorithms, servers, and data sets, such as through operational security measures for information security, specific cybersecurity policies, adequate technical and established solutions, and cyber and physical access controls, appropriate to the relevant circumstances and the risks involved.
(116) The AI Office should encourage and facilitate the drawing up, review and adaptation of codes of practice, taking into account international approaches. All providers of general-purpose AI models could be invited to participate. To ensure that the codes of practice reflect the state of the art and duly take into account a diverse set of perspectives, the AI Office should collaborate with relevant national competent authorities, and could, where appropriate, consult with civil society organisations and other relevant stakeholders and experts, including the Scientific Panel, for the drawing up of such codes.
Codes of practice should cover obligations for providers of general-purpose AI models and of general-purpose AI models presenting systemic risks. In addition, as regards systemic risks, codes of practice should help to establish a risk taxonomy of the type and nature of the systemic risks at Union level, including their sources. Codes of practice should also be focused on specific risk assessment and mitigation measures.
(117) The codes of practice should represent a central tool for the proper compliance with the obligations provided for under this Regulation for providers of general-purpose AI models. Providers should be able to rely on codes of practice to demonstrate compliance with the obligations. By means of implementing acts, the Commission may decide to approve a code of practice and give it a general validity within the Union, or, alternatively, to provide common rules for the implementation of the relevant obligations, if, by the time this Regulation becomes applicable, a code of practice cannot be finalised or is not deemed adequate by the AI Office.
Once a harmonised standard is published and assessed as suitable to cover the relevant obligations by the AI Office, compliance with a European harmonised standard should grant providers the presumption of conformity. Providers of general-purpose AI models should furthermore be able to demonstrate compliance using alternative adequate means, if codes of practice or harmonised standards are not available, or they choose not to rely on those.
(118) This Regulation regulates AI systems and AI models by imposing certain requirements and obligations for relevant market actors that are placing them on the market, putting into service or use in the Union, thereby complementing obligations for providers of intermediary services that embed such systems or models into their services regulated by Regulation (EU) 2022/2065. To the extent that such systems or models are embedded into designated very large online platforms or very large online search engines, they are subject to the risk-management framework provided for in Regulation (EU) 2022/2065.
Consequently, the corresponding obligations of this Regulation should be presumed to be fulfilled, unless significant systemic risks not covered by Regulation (EU) 2022/2065 emerge and are identified in such models.
Within this framework, providers of very large online platforms and very large online search engines are obliged to assess potential systemic risks stemming from the design, functioning and use of their services, including how the design of algorithmic systems used in the service may contribute to such risks, as well as systemic risks stemming from potential misuses. Those providers are also obliged to take appropriate mitigating measures in observance of fundamental rights.
(119) Considering the quick pace of innovation and the technological evolution of digital services in scope of different instruments of Union law in particular having in mind the usage and the perception of their recipients, the AI systems subject to this Regulation may be provided as intermediary services or parts thereof within the meaning of Regulation (EU) 2022/2065, which should be interpreted in a technology-neutral manner.
For example, AI systems may be used to provide online search engines, in particular, to the extent that an AI system such as an online chatbot performs searches of, in principle, all websites, then incorporates the results into its existing knowledge and uses the updated knowledge to generate a single output that combines different sources of information.
(120) Furthermore, obligations placed on providers and deployers of certain AI systems in this Regulation to enable the detection and disclosure that the outputs of those systems are artificially generated or manipulated are particularly relevant to facilitate the effective implementation of Regulation (EU) 2022/2065.
This applies in particular as regards the obligations of providers of very large online platforms or very large online search engines to identify and mitigate systemic risks that may arise from the dissemination of content that has been artificially generated or manipulated, in particular risk of the actual or foreseeable negative effects on democratic processes, civic discourse and electoral processes, including through disinformation.
Note: This is the final text of the Artificial Intelligence Act (Regulation (EU) 2024/1689), as it was published in the Official Journal of the European Union in July, 2024.