How to report an AI Act violation in the EU

The mechanism is real, and it is live

The EU AI Act did not just set rules. It built a way for anyone to trigger enforcement. As of August 2025, the Act's market surveillance provisions are active for general-purpose AI models, and the full complaint mechanism opens for high-risk systems as the rest of the Act phases in. A citizen, a competitor, or a civil society group can file a complaint with a national market surveillance authority. That authority must assess it, and if it finds a plausible violation, it can demand documents, run tests, and order a system off the market.

This is not a theoretical process. The Help Net Security report on how to report an AI Act violation walks through the practical steps: who to contact, what evidence to include, and what happens after you file. The mechanism is deliberately low-friction. You do not need a lawyer. You do not need to prove harm. You need to describe the system, explain why you believe it breaches the Act, and name the provider.

What actually goes wrong for providers

The failure mode here is not that the complaint is unfair. It is that most organisations cannot produce a coherent answer when one lands.

A complaint arrives on a Tuesday. The authority asks for the system's risk assessment, the training data documentation, the human oversight logs, and the fundamental rights impact assessment. You have 30 days, maybe fewer. The people who built the system have moved to another project. The documents live in three different sharepoint folders. The version in production does not match the version in the documentation. Nobody can say who signed off on the last model update.

This is the governance gap. The EU AI Act does not just require you to do the right thing. It requires you to prove you did the right thing, on demand, in a format a regulator accepts. The Act's Article 27 requires a fundamental rights impact assessment for high-risk systems, and Article 14 requires human oversight. Both of those are paper exercises unless you can show the regulator the actual assessment and the actual oversight logs, tied to the exact model version that is live.

The precedent that should scare you

The Netherlands in 2021 is the cautionary tale. An automated fraud-risk system wrongly flagged thousands of families as likely fraudsters, based on a flawed model and no meaningful human review. The cabinet resigned. That was before the EU AI Act existed. Under the Act, that system would be high-risk, and the fundamental rights impact assessment would have forced the government to confront the bias before deployment.

Clearview AI is the other warning. It scraped billions of facial images without consent, drew multiple GDPR fines across the EU and UK, and was ordered to delete data. The GDPR gave regulators the hook. The AI Act's Annex III adds biometric identification to the high-risk list, which means a company building that system now needs a lawful basis, a data protection impact assessment, and biometric-use restrictions from day one.

Neither of those companies failed because the technology was exotic. They failed because the governance was absent. The lesson for every provider is that the regulator does not need to understand your model. They need to see your paperwork.

The key risk indicators

If you want to know whether you are ready for a complaint, track three numbers.

First, the percentage of your AI systems that have a current applicability mapping. If you cannot say which of your systems fall under the high-risk rules, you cannot know which ones are exposed.

Second, the days of buffer you have to each statutory deadline. The high-risk rules apply from December 2027, but the complaint mechanism is already live for general-purpose models. Your buffer is shrinking.

Third, the time it takes you to produce a complete evidence pack on request. If the answer is more than a week, you are not ready. A regulator will not wait while you find the right spreadsheet.

What to do

• Run a complaint drill. Pick one high-risk system, write a mock complaint, and give your team 30 days to produce the full evidence pack. Measure how long it actually takes.
• Assign an owner. One person must be responsible for responding to regulatory inquiries, with authority to pull documents from any team.
• Build the evidence pack now. For each high-risk system, assemble the risk assessment, the fundamental rights impact assessment, the human oversight logs, and the version history. Update it every time the model changes.
• Check your oversight logs. Article 14 requires human oversight. If your logs show that a human reviewed every high-risk decision, you are in good shape. If they show nothing, fix that before a regulator asks.
• Map your applicability today. Do not wait for the 2027 deadline. The complaint mechanism is already open, and the first complaints are already being filed. ThreatClaw (www.threatclaw.ai) tracks the threat side of these same systems across 22 live intelligence feeds, so you can see what an adversary might exploit before a regulator does. A control is only as good as the threat it is sized against.

The bottom line

The EU AI Act turned governance from a best practice into a legal obligation. The complaint mechanism makes that obligation enforceable by anyone with an internet connection. The companies that survive the first wave of complaints will be the ones that treated the evidence pack as a living document, not a one-time exercise.

More from our platforms

These sister platforms cover the parts of this problem that sit outside governance.

• Argus (argus.threatclaw.ai) records every trace an AI application produces and scans it for prompt injection, jailbreaks and data leaks, including the attacks hidden inside retrieved documents and tool results rather than in what the user typed. Governance decides what an AI agent is allowed to do. Argus shows what it actually did.
• ThreatClaw (www.threatclaw.ai) tracks the threat side of the same systems: 22 live intelligence feeds, exploitation predicted before it is officially confirmed, threat actor profiles, and detection rules you can deploy straight away. A control is only as good as the threat it is sized against.