AI INCIDENT TRACKER updated —

When AI fails,
it lands here.

FAILCON — — in 30 days

A daily tracker of documented AI incidents and hazards — from biased models and deepfake scams to autonomous-system accidents. Counted, charted, and sourced from the three most serious public databases. No doom, no hype; just the score.

TL;DR

Loading the latest numbers…

The score

stats.json →
Incidents & hazards tracked OECD AIM, 2020 → today
This year documented so far this year
Last 30 days vs previous 30 days: —
Curated in depth incidents in the AI Incident Database

FAILCON — how loud is it right now?

Computing the trailing 30-day incident count…

The long fail

Incidents per yearOECD AIM · hatched = year to date
Hover a bar for details.
Last 24 monthsOECD AIM · monthly
Hover a bar for details.
What gets harmedOECD AIM · harm types, all time
One incident can carry several harm types.
Where AI failsMIT taxonomy · risk domains
Classified incidents from the AI Incident Database.

Fresh off the wire

Full wire →

Hall of Fail

All entries →

Methodology & sources

Full methodology →

Nobody agrees on what exactly counts as an "AI incident", so we don't pretend there is one true number. This tracker aggregates three public databases with different editorial bars and shows you which number comes from where:

Known gaps: all three sources skew towards English-language media; the OECD feed is LLM-classified news, so single events occasionally split or merge; and "documented" lags "happened" by days. Counts here are floors, not ceilings.

FAQ

What counts as an AI incident?

We follow the OECD definition: an event where an AI system's development or use caused actual harm — to people, property, rights, or infrastructure. A hazard is the near-miss variant: plausibly could have caused harm, didn't (yet). Both are tracked and labelled separately on the wire.

Why do your numbers differ from other trackers?

Because editorial bars differ. The OECD monitor mines global news automatically (~16k events), while the AI Incident Database curates by hand (~1.6k incidents). We show both and label which is which — treat every count as a floor, not a ceiling.

How fresh is the data?

A scheduled pipeline refreshes the OECD and MIT feeds daily and the AIID snapshot weekly. The "updated" stamp at the top of the page is the last successful run.

Is this site anti-AI?

No — it's built by people who ship software with AI every day. Ignoring failure modes is how they repeat. Counting them honestly is the least an industry this fast should do.