No single number settles whether AI is a blessing or a curse, but incident databases give a measurable record of harm. The AI Incident Database logged 233 reported AI-related incidents in 2024, up 56.4 percent from 2023, according to Stanford's 2025 AI Index, and the OECD runs a separate monitor with its own definitions. A rising count can reflect more AI use and more press coverage as much as more harm per system.
Key takeaways
- The AI Incident Database, run by the Responsible AI Collaborative, indexes reported harms and near harms from deployed AI systems.
- The OECD AI Incidents Monitor separates incidents, where harm occurred, from hazards, where harm is plausible.
- Both databases depend on public reporting, so they hold a subset of what happens.
- A count of incidents has no denominator unless the amount of AI use is known as well.
The AI Incident Database
The AI Incident Database describes its purpose as indexing the collective history of harms or near harms realized in the real world. The Responsible AI Collaborative, an organization chartered to advance the database, runs it, and anyone can submit a report through the site. Entries link to news articles and other published accounts, so each record rests on documents rather than on a laboratory test.
Stanford's 2025 AI Index uses this database for its incident figure. The report states that 233 reported incidents were recorded in 2024, a 56.4 percent increase over 2023 and the highest annual count it cites. The report page does not define what qualifies as an incident, so the underlying criteria have to be read in the database's own documentation.
The OECD AI Incidents Monitor
The OECD built a second system, the AI Incidents and Hazards Monitor, with a formal definition. An AI incident is an event where the development, use, or malfunction of an AI system directly or indirectly causes actual harm, including injury to health, disruption of critical infrastructure, violations of rights protected by law, and harm to property, communities, or the environment. An AI hazard is an event that could plausibly lead to such harm without having done so yet.
The monitor collects news articles through the Event Registry aggregation service. Language models sort events into incident, hazard, or unrelated, and a pipeline runs daily. The OECD methodology page lists its own caveats: only publicly reported events appear, the OECD does not verify third-party information, incidents before November 2024 were classified with older methods and are being re-annotated, and an open submission process is still being developed.
Why a rising count is hard to interpret
Three effects can raise a count without raising the harm per system. More people and companies use AI, so more events occur. Journalists and researchers pay more attention, so more events are reported. And classification methods change, as the OECD caveat about re-annotation shows. The reverse also holds: a flat count during rapid growth in use could mean each system became safer.
The figure in the AI Index responsible AI chapter therefore supports a modest claim, that reported harms are growing in number, and not a larger one about risk per use.
Comparing the two records
| Feature | AI Incident Database | OECD AI Incidents Monitor |
|---|---|---|
| Operator | Responsible AI Collaborative | OECD |
| Main input | Reports submitted and indexed from published accounts | News articles collected through an aggregation service |
| Includes near misses | Yes, harms or near harms | Yes, as hazards |
| Classification | Indexing by the database's own process | Language-model pipeline with checks |
| Stated limit | Reflects what is reported | Subset of events; third-party facts not verified |
What incident data cannot say about benefit
Harm counts are one half of the blessing-or-curse question. The other half is benefit, which is measured with different tools, mainly controlled studies of tasks, and the two do not net out against each other. Ten documented harms do not cancel ten documented gains, because the harms fall on different people than the gains. As an illustration, a missed alert that a hospital avoids and a person misidentified by a facial recognition system do not sit on the same ledger.
Checklist for reading an AI incident count
- Which database or monitor produced the number, and under what definition?
- Does the count include hazards and near misses, or only realized harm?
- Is the year-over-year change compared with growth in AI use?
- Did the classification method change between the years compared?
- Who was harmed, and how severe was the harm in each case?
FAQ
How many AI incidents happened in 2024?
The AI Incident Database recorded 233 reported incidents in 2024 according to the 2025 AI Index. The true number is unknown because many incidents are never reported.
What is the difference between an AI incident and an AI hazard?
Under the OECD definitions, an incident involves actual harm and a hazard involves plausible future harm. Both are tracked by the monitor.
Does a rising incident count mean AI is getting more dangerous?
Not by itself. More use, more reporting, and changes in classification can all raise the count.
Can incident databases show whether AI helps more than it harms?
No. They record harm only. Benefit has to be assessed through separate studies, and the comparison depends on who gains and who is harmed.
Related reading
- How AI systems fail and how to check them
- The limits of AI autonomy
- When emotion recognition gets people wrong
- Survey data on AI sentience timelines
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Responsible AI chapter, AI Index Report 2025, Stanford Institute for Human-Centered AI, 2025
- AI Incident Database home page, Responsible AI Collaborative, 2026
- AI Incidents and Hazards Monitor methodology, OECD AI, 2026