How we count AI incidents
What the dashboard shows, where the numbers come from, how current they are, and what they can’t tell you.
The short answer
The AI Incident Database records cases where AI systems caused or nearly caused harm, from news reports and public submissions: a self-driving car crash, a chatbot giving dangerous advice, a deepfake used for fraud. This dashboard counts those incidents by year, kind of harm and the companies named.
How we know
The data. The database is run by the Responsible AI Collaborative. Its editors give each incident the date it happened, a title, and the developers, deployers and people allegedly harmed. Many incidents are also classified by the MIT AI Risk Repository into seven kinds of harm. The database publishes a full snapshot every week, and we download the newest one when it appears.
Our calculations. We count incidents by the year they happened and by kind of harm, and count well-known companies named as developers in the last 12 months. We list only organisations by name, because tags can name private people. We use titles, dates and tags only, not the text of the reports, which belongs to their publishers.
How current each part is
- Incidents: a new snapshot every week; incidents are added after they are reported, so recent months fill in over time.
- Kinds of harm: classified some time after an incident is added.
What the dashboard can’t tell you
How much harm AI causes overall. The database records what is reported and noticed; most harms are never reported, and attention to AI has grown.
How serious each incident was, or whether the allegations were proven.
Which systems are safest. Widely used systems appear more often because more people use them.
Further reading
- AI Incident Database: search and read every incident and its reports. incidentdatabase.ai
- MIT AI Risk Repository: the classification of AI risks. airisk.mit.edu
Credits
Data: AI Incident Database, Responsible AI Collaborative, shared under the Creative Commons Attribution-ShareAlike 4.0 licence; kinds of harm from the MIT AI Risk Repository classifications in the database. Our data files for this page are shared under the same licence (CC BY-SA 4.0). Neither has endorsed this site.
← Back to the dashboard