Chapter 5 GENERAL-PURPOSE AI MODELS Flashcards
Q: How is a general-purpose AI model classified as having systemic risk?
A: A general-purpose AI model has systemic risk if it has high impact capabilities based on technical tools/methodologies or if the Commission decides it has equivalent capabilities/impact based on criteria in Annex XIII. There is a presumption of high impact if the computation used for training exceeds 10^25 FLOPs.
Q: What is the procedure for classifying a general-purpose AI model as having systemic risk?
A: The provider must notify the Commission within 2 weeks of meeting the high impact threshold, including information demonstrating this. The provider can present arguments that the model should not be classified as having systemic risk despite meeting the threshold. The Commission can also designate models as having systemic risk ex officio or following an alert from the scientific panel.
Q: What are the obligations for providers of general-purpose AI models?
A: Providers must keep technical documentation including training, testing and evaluation; provide information to downstream AI providers to understand capabilities/limitations; put in place a policy to comply with copyright law; and publish a summary of training data. Cooperation with authorities is required. Some exceptions apply for open source models.
Q: What additional obligations apply to providers of general-purpose AI models with systemic risk?
A: Providers must also perform and document model evaluation/testing to identify and mitigate systemic risks; assess and mitigate possible EU-level systemic risks; report serious incidents and corrective measures to authorities; and ensure adequate cybersecurity protection for the model and infrastructure.
Q: What are the requirements for authorised representatives of providers of general-purpose AI models?
A: Non-EU providers must appoint an EU-based authorised representative by written mandate. The representative verifies technical documentation is complete, keeps a copy for 10 years, provides information to demonstrate compliance to authorities, and cooperates with authorities on any actions regarding the model. Some exceptions apply for open source models.
Q: How can codes of practice be used by providers of general-purpose AI models to demonstrate compliance?
A: The AI Office shall facilitate development of EU-level codes of practice. Providers can rely on approved codes of practice to demonstrate compliance with their obligations until harmonised standards are published. Providers in compliance with harmonised standards are presumed compliant. Alternative means of compliance require Commission approval.
Q: What should codes of practice for general purpose AI models cover?
A: Codes should cover obligations in Articles 53 and 55, means to keep information up to date, level of detail for training data summaries, identification of systemic risks, and proportionate measures to assess and manage systemic risks. The AI Office shall monitor achievement of code objectives and regularly evaluate the codes.
Q: Who can participate in developing codes of practice and what is the process?
A: The AI Office may invite all general purpose AI providers and relevant authorities to participate, with support from civil society, industry, academia and other stakeholders. Participants should regularly report on implementation and outcomes. The Commission may approve a code by implementing act to give it general validity. Codes should be ready in 9-12 months or the Commission may provide common implementing rules.
Q: What confidentiality provisions apply to information obtained under the requirements for general purpose AI models?
A: Any information or documentation obtained, including trade secrets, must be treated in compliance with the confidentiality obligations in Article 78.
Q: How can the Commission amend the Annexes and thresholds relating to general purpose AI models?
A: The Commission can adopt delegated acts to amend the computation thresholds, supplement benchmarks and indicators, and amend Annexes XI (technical documentation) and XII (information for downstream providers) in light of technological developments.
