AppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

AI incident

Also called ai incident database

An AI incident is an event where an AI system actually causes harm, such as an injury, a wrongful arrest, a financial loss or a rights violation.

Safety and societyCurious

In plain words

When a plane has a problem, even a near-miss, it gets written down and studied so the same mistake is not made twice. AI is starting to get the same habit. An AI incident is a real case where an AI system caused or helped cause harm.

Examples include a face match that led police to arrest the wrong person, a chatbot giving a customer false information that cost them money, or a self-driving car hitting a pedestrian. A near-miss, where harm could have happened but did not, is usually called an AI hazard instead.

Public databases now collect these cases from news reports and submissions. The point is to learn, not to shame: builders can see how similar systems failed before and design around it.

Everyday example

The complaints register at a railway station

A railway station keeps a complaints book. One entry about a broken tap is minor. But if the same complaint appears at fifty stations, someone can spot the pattern and fix the design. AI incident databases are that register, shared across companies and countries, so one failure can warn everyone.

Where the example breaks: A complaints book records what passengers notice. Many AI harms are invisible to the people affected, such as a quietly rejected job application, so a lot of incidents are never reported at all.

Go deeper

How it actually works

Incident tracking borrows from aviation safety and computer security, where shared records of failures became standard practice.

  1. A working definition

    In May 2024 the OECD defined an AI incident as an event where the development, use or malfunction of AI systems directly or indirectly leads to harm: injury to health, disruption of critical infrastructure, violations of human rights or legal obligations, or harm to property, communities or the environment. A hazard is an event that could plausibly lead to one.

  2. Collecting the cases

    The AI Incident Database, launched in November 2020 and now run by the Responsible AI Collaborative, takes public submissions. The OECD's AI Incidents and Hazards Monitor scans large volumes of international news daily to spot new cases. Both rely mostly on what gets reported.

  3. Counting the trend

    Stanford's AI Index 2025 found reported incidents in the AI Incident Database rose to 233 in 2024, a record and up 56.4% on 2023. More use of AI and more attention both push the number up, so counts show reporting as much as risk.

  4. Mandatory reporting

    Laws are adding duties. California's SB 53, in force since January 2026, requires frontier AI developers to report critical safety incidents to the state within 15 days, or to an appropriate authority within 24 hours if there is imminent risk of death or serious injury. The EU AI Act requires providers of high-risk systems to report serious incidents, once high-risk rules apply (from December 2027 at the earliest).

  5. Learning from them

    Incident reports feed red teaming, tests and model cards, and they help regulators see where rules are needed. Near-misses are worth logging too, because they show weak spots before anyone is hurt.

Example
Moffatt v. Air Canada (February 2024): a website chatbot gave wrong refund rules and the airline had to pay
AI Incident Database
Introduced 18 November 2020, modelled on aviation reporting and the CVE list of security flaws; run by the Responsible AI Collaborative
OECD monitor
AI Incidents and Hazards Monitor, which scans international news daily for new incidents and hazards
Reported incidents
233 in 2024 in the AI Incident Database, up 56.4% on 2023 (Stanford AI Index 2025)

Questions people ask

What is the difference between an AI incident and an AI hazard?

An incident is harm that actually happened. A hazard is an event or situation that could plausibly have led to harm but did not, a near-miss. The OECD uses this split, and both are worth recording.

What is the AI Incident Database?

A free public collection of real AI failures, at incidentdatabase.ai. It was introduced in November 2020 and is run by the Responsible AI Collaborative. Anyone can submit a case, and each entry links to news reports about it.

Are AI incidents increasing?

Reported ones are. Stanford's AI Index counted 233 in 2024, up 56.4% from 2023. That reflects more AI in use and more people watching for problems, so it is not a precise measure of risk.

Do companies have to report AI incidents?

Increasingly, yes. California has required frontier developers to report critical safety incidents since January 2026, New York will from January 2027, and the EU AI Act requires reporting for high-risk systems. Many other cases remain voluntary.

Is a chatbot hallucination an AI incident?

Only if it causes harm. A wrong answer nobody acts on is just a hallucination. When a false answer costs someone money, as with Air Canada's chatbot in 2024, or damages a reputation, it becomes an incident.

Was this clear?

1 is confusing, 5 is crystal clear. One tap, no account needed.

Spotted something wrong, or know a better example? Suggest a change

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked November 2026.