The MIT AI Incident Tracker, a project of the MIT AI Risk Initiative, has classified roughly 1,600 real world incident reports drawn from the AI Incident Database, sorting each one by risk domain, cause, harm type and severity. The breakdown gives a picture of where reported AI failures are concentrated.

Malicious actors and misuse account for 38.8 percent of classified incidents, the largest single category. AI system safety, failures and limitations follow at 20.1 percent. Discrimination and toxicity account for 15.8 percent, misinformation for 13.7 percent, and privacy and security for 7.4 percent. Human computer interaction cases make up 2.7 percent and socioeconomic and environmental cases 1.4 percent.

The tracker reports that about half of all logged incidents were tagged as intentionally caused. Almost all classified incidents occurred after deployment, with only 2 percent traced to a pre deployment stage. Applying European Union AI Act risk levels to the same dataset, 5.7 percent of reported incidents would fall into the unacceptable tier and 33 percent into the high risk tier.

Raw reports from the AI Incident Database are processed by an automated classification pipeline that applies the MIT Risk Repository taxonomy and a harm severity rating based on the CSET AI harm taxonomy. The harm view separates 10 categories, including physical harm, property damage, financial loss and human rights.

MIT flags limits on the dataset. Reporting to the AI Incident Database is voluntary and depends on public and expert submissions, so depth and reliability vary across records and sampling bias is present. A systematic validation study of the classification tool has not been completed.

Source: MIT AI Risk Initiative - https://airisk.mit.edu/ai-incident-tracker