Correction: this post has been updated. See what actually changed and what didn't for the current EU AI Act timeline.
The EU AI Act is the world's first comprehensive legal framework for artificial intelligence. Its provisions for high-risk AI systems apply from 2 December 2027 for Annex III stand-alone systems and 2 August 2028 for Annex I product-embedded systems. Article 50 transparency obligations, covering chatbot disclosure, synthetic content marking, and deepfake labelling, still apply from 2 August 2026. If you are deploying AI in Europe - or deploying AI that affects people in Europe - you need to understand what it requires and how to demonstrate compliance.
What the EU AI Act requires
The Act establishes mandatory requirements for AI systems based on their potential to cause harm. For high-risk systems, these requirements are substantial: documented risk management processes, technical robustness standards, human oversight mechanisms, transparency obligations, and ongoing monitoring after deployment.
The compliance burden is not just administrative. The Act requires that you can demonstrate, with evidence, that your system meets the technical standards. Saying "we tested it" is not sufficient. You need documented test procedures, results, and the ability to show that your testing methodology was appropriate for the risk level of your system.
Enforcement is handled by national market surveillance authorities, with penalties for non-compliance reaching up to 30 million euros or 6% of global annual turnover - whichever is higher. The Act also includes a right of redress for individuals affected by high-risk AI decisions.
Risk classification
The Act classifies AI systems into four risk tiers. Understanding where your system sits determines what compliance obligations apply.
- Unacceptable risk - Systems that are prohibited outright. Includes social scoring systems, real-time biometric surveillance in public spaces, and systems that exploit psychological vulnerabilities.
- High risk - Systems in regulated domains (healthcare, employment, credit, education, critical infrastructure, law enforcement) that must meet full compliance requirements including conformity assessments before deployment.
- Limited risk - Systems with specific transparency obligations, such as chatbots that must disclose they are AI, or deepfake content that must be labelled.
- Minimal risk - Systems with no mandatory requirements beyond good practice. Most AI applications fall here, but this does not mean testing is optional - it means the legal floor is lower.
If you are uncertain about your classification, err on the side of caution. The cost of treating a high-risk system as minimal risk is far higher than the cost of applying appropriate evaluation standards to a minimal-risk system.
What to test
For high-risk systems, the Act mandates testing across several specific areas. Each of these maps directly to evaluation activities you can structure and document.
- Accuracy across intended use cases - Performance must be declared and demonstrated. You need test results showing accuracy across the range of inputs your system is expected to handle in production.
- Robustness to errors and faults - Systems must be resilient to errors in inputs, unexpected conditions, and attempts to manipulate outputs. Adversarial testing is not optional.
- Bias and fairness - Systems must be tested for discriminatory outputs across protected characteristics. This includes testing on demographically diverse test sets and documenting findings.
- Human oversight mechanisms - You must be able to demonstrate that humans can intervene in, override, or shut down the system. The evaluation should include testing these mechanisms.
- Data governance - Training and test data must be documented. Gaps, biases, and limitations in the data must be identified and addressed.
- Transparency and explainability - Users and oversight authorities must be able to understand what the system does and why it produces the outputs it does.
How EvalSpec maps to Article requirements
EvalSpec's evaluation framework was designed with regulatory alignment in mind. The five dimensions map directly to the technical requirements in the Act's key Articles.
Article 9 (Risk Management System) requires documented risk identification, analysis, and mitigation. EvalSpec's Safety and Compliance dimension structures exactly this - forcing you to document identified risks, test for them explicitly, and record your mitigation approach with evidence.
Article 15 (Accuracy, Robustness, and Cybersecurity) sets technical performance standards. EvalSpec's Accuracy and Groundedness and Robustness and Edge Cases dimensions are designed to produce the documented test evidence Article 15 requires.
Article 13 (Transparency and Provision of Information) requires that systems are designed to be sufficiently transparent. EvalSpec's Alignment and Usefulness dimension includes transparency scoring - whether the system behaves in ways that are understandable and predictable to users.
Article 14 (Human Oversight) requires that human oversight is technically feasible and has been tested. EvalSpec includes human oversight as a structured evaluation checkpoint within the compliance dimension.
Using a structured evaluation tool does not replace legal review, but it produces the documented evidence base that legal review and regulatory inspection will look for.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation