AI-incidens
in English: AI incident
A case where an artificial intelligence system's operation or error directly or indirectly causes actual harm to people, organizations, or the environment, or violates fundamental rights.
In plain terms
An AI incident is the 'accident' of the artificial intelligence world. Just as in traffic everything from a fender bender to a mass casualty event counts as an accident, the scale here is wide: from a false bank rejection to fraud committed with a fake video to injuries caused by a self-driving car.
The concept is useful so that problems are not treated as individual bad luck but are collected and learned from. Aviation became safe largely because accidents and dangerous events are systematically investigated and recorded, and the same approach is being introduced in AI.
Unlike a simple software bug, here the damage often does not stem from "malfunction." The system runs technically flawlessly but makes poor decisions, judges with bias, confidently states falsehoods, or an agent independently performs actions no one requested.
How it works
- Detection: someone notices that something has gone wrong, for example, a customer files a complaint, a journalist reports the incident, or the company's own monitoring system triggers an alert.
- Recording: they document what happened, which system was affected, what damage occurred to whom, and preserve logs so the incident can be traced later.
- Classification: they decide how serious the case is and whether actual harm occurred or there was only a realistic risk of it (the latter is called an AI hazard).
- Investigation and remediation: the cause is identified (poor training data, lack of human oversight, abuse, excessive autonomy), then the system is fixed or shut down.
- Reporting and learning: serious cases must be reported to the authorities within the scope prescribed by law, and lessons learned may be entered into public databases so others do not repeat the same mistakes.
An everyday example
A customer service chatbot of an online store promises you a refund that does not exist in the company's policy, and you make a purchase based on it. When the company refuses to pay and you suffer a loss, this is already an AI incident, not a simple misunderstanding.
Who is affected
A common misconception
Limits and risks
- There is no single universally accepted definition: the OECD interprets it broadly, EU law defines only the 'serious incident', and company internal regulations mean something else again.
- Public databases mainly work from press reports, so they only show the tip of the iceberg, and the rise in case numbers partly reflects greater attention.
- It is often hard to prove that the damage was truly caused by AI and not by a human who used it poorly or trusted it blindly.
- Companies are reluctant to discuss their own cases, so many lessons never reach the public.
Similar, but not the same
- AI emergency – There, the harm had not yet occurred, only a realistic chance of it; in an incident, it has already happened.
- Cybersecurity incident – This is any breach of an IT system's security, typically due to an attack or data leak. An AI incident does not require an attacker; a system's own poor decision is sufficient. The two can overlap, for example if an AI system is hacked or coerced into releasing data.
- hallucination – A hallucination is the model's erroneous output, which in itself is not yet an incident; it only becomes one if someone suffers harm from it.
Related concepts
AI safety, AI risk management, AI governance
Serious incident, Incident reporting
AI emergency, hallucination, algorithmic bias, Human oversight, Deepfake, Red teaming
A brief history
The term was adopted from aviation and cybersecurity incident management practices. The public AI Incident Database launched in 2020, the OECD launched its own incident monitor in 2023, and the EU AI Act, adopted in 2024, established the 'serious incident' as a legal concept.
Where things stand
By the end of 2026, the term is included in legislation and internal regulations of major companies, while public databases record thousands of cases. In the EU, developers of the largest general-purpose models have been required to report serious incidents since August 2025, but reporting rules for high-risk systems will only become applicable later. Therefore, a unified reporting practice is still taking shape, and autonomous agents are introducing new types of cases.
In the news
- Anthropic says no single insurer may have capacity to cover a catastrophic AI incidentOct 26, 2026 · Money
- Hassabis calls AI hacking "early warning shots"; major incidents could come within two yearsOct 13, 2026 · Agent Age
- Microsoft's Brad Smith, Richard Hermer and 400 scientists call for tougher AI regulationOct 12, 2026 · Jog
- US and China agree to an AI incident channelOct 8, 2026 · Agent Age
- IAB Hungary: agentic ad buying could become routine within a yearSep 23, 2026 · Marketing
- US proposes AI incident alert system to China ahead of Trump–Xi summitSep 14, 2026 · Agent Age
Also known as
AI-incidens · MI-incidens · AI incident · AI incident · serious incident · serious incident · AI-baleset · AI event
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.