EU AI Act Art. 15 — Accuracy, robustness and cybersecurity
Article 15 sets the performance bar: high-risk AI must achieve appropriate accuracy, robustness, and cybersecurity for its purpose, consistent with the state of the art — and be resilient against manipulation and attacks.
At a glance
What this article requires
- Accuracy levels must be appropriate to the intended purpose and reflected in declared metrics.
- Systems must be robust against errors, inconsistencies, and incorrect inputs.
- Robustness covers the whole system — including surrounding software and hardware interactions.
- Cybersecurity measures must prevent attacks that exploit system vulnerabilities to manipulate outputs.
- Providers must disclose known limitations and the metrics behind their accuracy claims.
Scope
Who this applies to
Providers of high-risk AI must engineer and evidence performance; deployers should verify the declared metrics match their real deployment conditions before relying on the system.
Obligations
What you must actually do
Declare honest metrics
Define and document accuracy metrics per task (e.g. precision/recall, error rates across subgroups). Publish the measurement methodology so claims are checkable.
Harden against attacks
Address adversarial inputs, prompt or data poisoning, and model extraction. Mitigations must be documented and proportionate to the threat model.
Design for the real environment
Robustness must account for the system interacting with other software, hardware, and noisy real-world data — test in conditions that match deployment.
Action plan
Practical first steps
- 1
Write a performance-and-robustness test plan with clear metrics before launch.
- 2
Run subgroup accuracy analysis — overall accuracy hides demographic failures.
- 3
Threat-model the system: who would attack it, how, and what is the blast radius?
- 4
Set monitoring thresholds so drift and degradation trigger alerts, not surprises.
Penalty exposure
Accuracy/robustness/cybersecurity failures are in the general tier: up to €15 million or 3% of global annual turnover.
FAQ
Questions about Art. 15
How accurate is 'accurate enough'?
There is no universal number — accuracy must be appropriate to the intended purpose, consistent with state of the art, and disclosed. Medical triage needs a far higher bar than content tagging, and the metrics must be published.
Does Article 15 require me to prevent all attacks?
No — it requires cybersecurity measures proportionate to the risks and vulnerabilities, following accepted practice. Perfect security is not achievable; documented, tested defence-in-depth is the standard.
Sources
Citations & further reading
Related
More article explainers
Wondering which articles apply to your AI?
Describe your system in the free Risk Scanner and get a preliminary risk read with the obligations that likely apply — in seconds.
Check my use casePreliminary EU AI Act clarity summary. Not legal advice.