Serious incident reporting under the EU AI Act

Serious incident reporting under the EU AI Act 

Beginning of this month, the Commission published its Reporting Template for Serious Incidents involving General-Purpose AI Models with Systemic Risk. Providers now have a concrete view of what ‘accountability’ looks like under the EU AI Act.

Article 55(1)(c) of the EU AI Act imposes a clear legal obligation on providers:

“keep track of, document and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them.”

The Commission makes it clear that reporting a serious incident does not constitute admission of wrongdoing. Whether this principle will hold in practice remains to be seen, but it mirrors the data-breach notification approach under the GDPR, where reporting does not automatically lead to sanctions. The underlying rationale is that information that may directly or indirectly signal the lead-up to a serious incident is often dispersed, ephemeral, and easily lost, overwritten, or fragmented by the time a provider becomes aware of the incident itself. Therefore, the Commission considers it essential that providers establish proactive processes and safeguards to continuously track and document relevant information before a serious incident occurs.

Under Commitment 9 of the Safety and Security Chapter of the General-Purpose AI Code of Practice, providers of systemic-risk GPAI models accept a set of binding obligations that go well beyond ad-hoc reporting. They must establish processes and resourcing to (i) keep track of relevant information, (ii) document it, (iii) report it “without undue delay,” and (iv) address serious incidents through corrective actions. These obligations apply across the entire model lifecycle, not only at deployment. Keep in mind that the Code of practice is only voluntary tool for providers of GPAI models, but seen as adequate by the Commission and the AI Board to demonstrate compliance with the AI Act.

Information to be reported

The following information must be reported to the best of the provider’s knowledge:

1. the start and end dates of the serious incident, or best approximations thereof if the precise dates are unclear;
2. the resulting harm and the victim or affected group of the serious incident;
3. the chain of events that (directly or indirectly) led to the serious incident;
4. the model involved in the serious incident;
5. the available evidence: a description of material available setting out the model’s involvement in the serious incident;
6. the response: what, if anything, the provider intends to do or has done in response to the serious incident;
7. the recommendation: what, if anything, the provider recommends the AI Office and, as applicable, national competent authorities to do in response to the serious incident;
8. a root cause analysis with a description of the model’s outputs that (directly or indirectly) led to the serious incident and the factors that contributed to their generation, including the inputs used and any failures or circumventions of systemic risk mitigations; and
9. any patterns detected during post-market monitoring (pursuant to Measure 3.5) that can reasonably be assumed to be connected to the serious incident, such as individual or aggregate data on near misses.

Providers will provide the information in points (1) to (7) in an initial report If information is missing at first, providers must record that fact: the duty is to report what is known, not to know the unknowable immediately. The detail required must match the severity of the incident.

Template Report for Serious Incidents under the AI Act (General-Purpose AI Models with Systemic Risk)

Reporting timelines 

Within 2 days: serious and irreversible disruption of critical infrastructure or reasonable suspicion of such a causal link
Within 5 days: serious cybersecurity breach (including exfiltration of model weights, cyberattacks)
Within 10 days: death of a person, or reasonable suspicion of causality
Within 15 days: serious harm to health (mental or physical), infringement of EU fundamental-rights, serious harm to property or the environment

Intermediate reports: every 4 weeks, with updated information until the incident is resolved.

Final report: Within 60 days after resolution, containing all information required.

Given the staged reporting obligations under the AI Act, one would expect the template to contain a clear field identifying the report’s status (initial, intermediate, or final).

Multiple similar incidents may be grouped if deadlines for the earliest event are respected.

Providers must retain all documentation for at least 5 years from the date of (i) the documentation itself, or (ii) the serious incident, whichever is later. This interacts directly with the principle that information must be tracked before harm occurs, otherwise key evidence will be lost.

While we understand that this reporting obligation may feel like yet another compliance requirement to an already long list, the reality is more nuanced. Serious-incident reporting will undoubtedly require significant effort, new processes, and cultural change within organisations. Yet in the long run, these mechanisms can strengthen internal governance, improve visibility over model behaviour, and help organisations detect risks earlier and respond more effectively. What begins as a regulatory duty can ultimately become a driver of better AI oversight, greater organisational maturity, and more trustworthy AI practices overall.