AI Incident Database
A public, searchable collection of real-world harms and near harms caused by deployed AI systems, run by the Responsible AI Collaborative. The best place to learn how AI fails outside the lab.
Platform
AI & Machine Learning
Responsible AI Collaborative
Browse for an hour, return when you write risk assessments
All
Free
What it is
The AI Incident Database collects reports of harms or near harms that AI systems have caused once deployed in the real world. It is modelled on the incident databases used in aviation and computer security, where learning from past failures is how the whole field gets safer. It is run by the Responsible AI Collaborative, a non-profit.
What you will find
Thousands of indexed reports, grouped into incidents, covering cases such as chatbots giving dangerous advice, lawyers and courts submitting AI-invented citations, deepfakes used for fraud and extortion, biased automated decisions, and surveillance systems misused. You can browse with a discovery view, a table or a map, and anyone can submit a new incident.
Why it matters for governance
Risk assessments are only as good as the risks people can imagine. Real incidents make risk concrete: what went wrong, who was harmed, what control was missing, and how the organisation responded. Incident reporting is also a legal duty in some places now: the EU AI Act requires providers of high-risk systems to report serious incidents to the authorities.
How to use it
- For a risk register. Before you list risks for an AI system, search the database for similar systems (hiring tools, chatbots, credit scoring) and note the failure modes that actually happened.
- For training and portfolio work. Pick one incident and write a short post-incident review: what happened, the root cause, which control from a framework such as the NIST AI RMF would have prevented or detected it, and who should have owned it.
- For interviews and conversations. Being able to discuss three real incidents in detail shows you understand the field beyond the frameworks.
Limitations
The database depends on public reporting, so it over-represents incidents that made the news in English-speaking countries. Absence of incidents for a system type does not mean it is safe.
Free or paid?
Free.
Ready to start?
Opens on Responsible AI Collaborative in a new tab.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.