This is not a trend in AI incidents. A month’s height reflects how completely it has been searched, not how much went wrong — so read a step as a sourcing event until proven otherwise. Click a barto list that month’s events.
Showing 225 of 276 events; 51 excluded — anything undated or before January 2025.
Where clusters form. Reads two ways: a row that is hot across many columns is one failure mode doing broad harm; a hot column is one impact type driven by several modes. Colour is each cell’s share of the busiest cell, so the scale re-reads as the register grows. Click any cell to list the events behind it.
Financial | Reputational | Regulatory/Legal | Safety | Social/Environment | Operational | Intellectual Property | |
|---|---|---|---|---|---|---|---|
| Model Hallucination | 7 | 27 | 8 | 1 | |||
| System Failure | 2 | 13 | 1 | 2 | 21 | ||
| Privacy Breach | 7 | 7 | 1 | 4 | 1 | ||
| Data Bias | 1 | 15 | 3 | 1 | |||
| Adversarial Attack | 10 | 17 | 6 | 4 | 2 | 11 | 12 |
| Governance Failure | 10 | 15 | 57 | 4 | 1 | 3 |
How often each failure mode is recorded, and what share of the register that is. Click a barto list that mode’s events.
Root events by region, in the same taxonomy as the home-page event heat map. Country codes inside each bar show what it is made of. This is where incidents were reported, and English-language media coverage is not evenly distributed — read a short bar as thin coverage rather than as safety. Click a barto list that region’s events.
The 8 biggest published figures, of 66 rows carrying one at all. A rank of what got reported, not of what was lost. Click an event for its full record.