234 incidents on record · 2026HeadlightsIncident reports by Ellie Harris · Melbourne
5 new this weekLibrary last updated 14 October 2026
Findings

What the record shows

Counts from all 234 incidents in the library, as of 14 October 2026. They are worked out from the entries each time the site is built, so they are always current.

Headlines
  1. Overconfidence is the behaviour Headlights finds most often: it is the main behaviour in 58 of 234 incidents (25%), ahead of rules-behaviour gap (38) and automation bias (28).
  2. Counting contributing behaviours too, the three most common are overconfidence (74), rules-behaviour gap (58), automation bias (33).
  3. In 48 incidents (21%) the public record does not support a reading of any behaviour, and Headlights says so rather than guess.
  4. Technology & software is the industry with the most incidents (38, 16%).
  5. The United States accounts for 115 incidents (49%) and Australia for 34, across 32 countries and regions in all.
Main behaviour in the record · 186 of 234 incidents
  1. Overconfidence58 · 25%
  2. Rules-behaviour gap38 · 16%
  3. Automation bias28 · 12%
  4. Judgement surrendered12 · 5%
  5. Concealment10 · 4%
  6. Normalisation of deviance9 · 4%
  7. Delayed intervention8 · 3%
  8. Authority ambiguity5 · 2%
  9. Defensive denial5 · 2%
  10. Diffusion of responsibility5 · 2%
  11. Group conformity4 · 2%
  12. Incentive pressure4 · 2%

Not established: 48 (21%). A behaviour tag is Headlights' reading of conduct in the public record, against a fixed list of thirteen, never a finding by a court, regulator or company.

The behaviour most often seen, by industry
IndustryIncidentsMost common main behaviour
Technology & software38Overconfidence (12)
Consumer AI36Overconfidence (15)
Government & public administration32Rules-behaviour gap (10)
Policing, justice & courts16Automation bias (6)
Retail, hospitality & travel16Overconfidence (5)
Professional & legal services15Automation bias (9)
Financial services & insurance14Overconfidence (3)
Health & medicine14Overconfidence (6)
Media & platforms14Automation bias (3)
Education12Overconfidence (3)
Employment & hiring12Overconfidence (3)
Industry, transport & logistics12Defensive denial (2)
Housing & property3Overconfidence (2)
By industry
  1. Technology & software38 · 16%
  2. Consumer AI36 · 15%
  3. Government & public administration32 · 14%
  4. Policing, justice & courts16 · 7%
  5. Retail, hospitality & travel16 · 7%
  6. Professional & legal services15 · 6%
  7. Financial services & insurance14 · 6%
  8. Health & medicine14 · 6%
  9. Media & platforms14 · 6%
  10. Education12 · 5%
  11. Employment & hiring12 · 5%
  12. Industry, transport & logistics12 · 5%
  13. Housing & property3 · 1%
By country, top ten
  1. United States115 · 49%
  2. Australia34 · 15%
  3. United Kingdom16 · 7%
  4. Italy7 · 3%
  5. Netherlands7 · 3%
  6. Brazil6 · 3%
  7. Canada5 · 2%
  8. South Korea5 · 2%
  9. Spain4 · 2%
  10. Austria3 · 1%
By year of the defining event, last four years
  1. 202334
  2. 202441
  3. 202577
  4. 202639

Earlier: 2015 (1), 2016 (3), 2017 (2), 2018 (4), 2019 (2), 2020 (9), 2021 (13), 2022 (9).

Read these numbers with care

Headlights is a curated library, not a census. It documents incidents with a public record strong enough to write from: regulators, courts, ombudsmen and reputable press. The counts show what that record contains, not how often AI fails in each industry or country. How Headlights works. Countries with open courts, active regulators and an English-language press are easier to document, which is one reason the United States and Australia lead.

Cite these findings: Harris, E. (2026). What the record shows: findings from the Headlights AI Incident Library. Headlights. https://useheadlights.com/findings/ · Data: doi.org/10.5281/zenodo.23312145