AI Accuracy Requirements
Follow topic LLM context A cited markdown file you can paste into your AI assistant (ChatGPT, Claude, a RAG or project knowledge base) to ground it in this topic. Contains: the overview, key law text, case law, enforcement and guidance for this topic. Everything links back to its source on overview.legal — legal information, not advice.Accuracy is a critical AI Act requirement deserving dedicated topic coverage for measurement, validation, monitoring, and maintenance of AI system performance standards.
Overview
Legal Framework
The accuracy requirement for AI systems is established by the EU AI Act, specifically Recital 74. This recital mandates that high-risk AI systems must perform consistently throughout their lifecycle and meet an appropriate level of accuracy, robustness, and cybersecurity. The required level of performance is not a fixed standard but is assessed in light of the system's intended purpose and the generally acknowledged state of the art. This legal framework obliges providers to define and declare the expected level of performance using measurable metrics, which must be documented in the system's accompanying technical documentation.
Practical Application
The requirement for "appropriate" accuracy is purpose-driven and contextual. As emphasized in Recital 122, compliance with data governance requirements—a foundation for accuracy—is presumed when an AI system is trained and tested on data reflecting its specific intended operational setting, including geographical, behavioural, and functional contexts. This principle discourages the use of generic, non-representative datasets. While case law on AI-specific accuracy is nascent, established data protection principles provide guidance. The Minister v. M ruling underscores that data subjects must be able to check the accuracy of data processed about them, a right that logically extends to outputs from AI systems processing personal data. Furthermore, the WORTEN case reinforces that any data processing, including that which trains or feeds an AI system, must be necessary and proportionate to its stated purpose, which directly informs the scope and benchmarks for required accuracy.
Key Considerations
- Define and Document Context-Specific Metrics: Organizations must explicitly define quantifiable accuracy metrics (e.g., precision, recall, error rates) that are appropriate for the AI system's specific intended use and documented operational environment.
- Implement Lifecycle Monitoring: Accuracy is not a one-time validation. Providers must establish procedures for continuous post-market monitoring to ensure performance remains consistent and appropriate as the system operates in the real world and as the state of the art evolves.
- Validate Data Representativeness: To leverage the compliance presumption under Recital 122, rigorously document how your training, validation, and testing data reflect the actual geographical, behavioural, and contextual setting where the high-risk AI system will be deployed.