AI Risk
The latest real-world AI harms logged by the AI Incident Database, what they are, and how often new ones appear.
25 AI incidents with new reports in 14 days
What’s happening right now
Latest reports added to the AI Incident Database
Each report is a news article or document the AIID editors attached to an incident. Several reports can cover one incident.
The trend
Reports and distinct incidents per day (feed window)
AIID's public RSS feed carries only its 100 most recent reports, so this covers a few weeks. Longer history isn't available key-free (the AIID API is restricted), so none is shown.
Background: why this matters
Artificial intelligence sits on both sides of this site's needle. It's a powerful tool for science and the economy (see AI Research), and it's also a source of real, documented harm today. Many researchers also think it could become a catastrophic risk as systems grow more capable. In May 2023 hundreds of AI scientists and industry leaders signed a one-sentence statement saying that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.
The harms that are already here are more mundane but concrete. They include wrongful arrests based on faulty facial recognition; a 2018 fatal crash involving an Uber self-driving test car in Tempe, Arizona; deepfake fraud and harassment; biased decisions in hiring, lending and welfare; and chatbots giving dangerous advice. The AI Incident Database (AIID), run by the non-profit Responsible AI Collaborative since 2020, catalogues these cases. Each incident is a real-world event where an AI system caused or nearly caused harm, backed by one or more news reports.
Governments have started to respond. The EU AI Act entered into force in August 2024, the first broad AI law of its kind. The UK's 2023 AI Safety Summit produced the Bletchley Declaration, signed by 28 countries and the EU, and the independent International AI Safety Report, chaired by Yoshua Bengio, reviewed the evidence on risks from general-purpose AI in 2025. Frontier AI companies now publish safety frameworks and test their models for dangerous capabilities, such as helping with biological weapons or cyberattacks, before release.
This factor counts how many distinct AIID incidents received new reports in the last 14 days, using the database's public feed of its 100 most recent reports. It's an honest measure of documented, present-day harm, not a forecast of catastrophe. The count also reflects how much the media covers AI and how quickly AIID's volunteer editors process submissions. A quiet fortnight doesn't mean AI is safe, and a busy one doesn't mean disaster is near. We treat it as a signal of how often AI is going wrong in the world right now.
What experts watch: AI agents acting autonomously on the internet, misuse for cyberattacks, fraud and disinformation, capability evaluations for biological and chemical risk, incidents involving critical infrastructure, and whether safety practices keep pace as models become more capable.
What would move the score
N = distinct AIID incident IDs cited by reports published in the last 14 days (from the 100-report RSS feed). Score = 100 at N ≤ 5, 0 at N ≥ 50, linear. Each additional incident costs about 2.2 points.
Right now: 56/100 · Distinct AI incidents reported, last 14 days: 25 · Reports in window: 42
↓ Toward extinction
- More distinct AI incidents reported in a fortnight: new kinds of harm, or harms spreading across many systems
- Waves of coverage about AI failures, which bring more reports into the database
↑ Toward prospering
- Fewer documented incidents as safety practices, testing and regulation improve
- Incidents being prevented before deployment rather than reported after harm
Exact scoring rule and calibration
N = distinct AIID incident IDs cited by reports published in the last 14 days (from AIID's live RSS feed of its 100 most recent reports). Score = 100 at N ≤ 5, 0 at N ≥ 50, linear between. If the feed doesn't reach back a full 14 days the count is shown as a minimum (≥ N). Caveat: counts reflect media coverage and AIID editorial capacity as well as underlying harm; it is an editorial threshold, not a probability of AI catastrophe.
Sources
Score data as of Sep 30, 2026, 00:00 UTC, fetched Oct 2, 2026, 18:04 UTC. Deep-dive data fetched Oct 2, 2026, 18:31 UTC. Pages refresh every 15 minutes; each source is cached for 15 minutes (GDELT) to 24 hours (long histories) to respect free-API limits. Nothing is estimated or filled in by hand.