> ## Content Index > Fetch the complete content index at: https://aiadoption.org/llms.txt > Use this file to discover other available public pages before exploring further. # AI incidents are now a monthly category - URL: https://aiadoption.org/ai-analysis/ai-incidents-are-now-a-monthly-category/ - Published: 2026-09-28T06:04:05.000Z - Updated: 2026-09-28T06:04:05.000Z - Description: The AI Incident Database recorded 362 incidents in 2025, up from 233 in 2024. OECD's automated tracker recorded a monthly peak of 435 in January 2026. - Author: Jassie - Tags: AI Analysis, AI Cybersecurity, AI Strategy The AI Incident Database recorded 362 incidents in 2025, up from 233 in 2024\. The OECD's automated AI Incidents and Hazards Monitor recorded a monthly peak of 435 in January 2026\. AI incident frequency has moved from "occasional newsworthy events" to "monthly category." The deployment risk landscape that procurement and risk teams plan against has changed shape. AI incidents reported annually, 2012–25 (AIID): The annual numbers from the AI Incident Database (the longest-running curated tracker) show a steep curve. Annual incidents stayed under 100 through 2021\. The 2022–2023 period saw the count cross into the hundreds. 2024 closed at 233\. 2025 closed at 362\. The year-on-year increase from 2024 to 2025 alone is roughly 130 additional documented incidents. AIID curates manually. Human editors review submissions against a defined threshold of AI involvement, drawing from academic and investigative journalism sources. The manual process produces higher-quality records but skews toward English-language media and high-visibility incidents. Less-accessible regions and lower-profile incidents are likely underrepresented. The OECD AI Incidents and Hazards Monitor is automated and multilingual. Its absolute numbers run higher than AIID's because the automated pipeline catches incidents the curated process does not. OECD AIM recorded a monthly peak of 435 in January 2026 and a six-month moving average of 326\. Both trackers show consistent and sharp increase, though the absolute scales differ. Monthly AI incidents reported in news, 2020–2026 (OECD AIM) The McKinsey survey provides a third angle. The share of organisations reporting AI incidents stayed steady at 8% across 2024 and 2025, meaning the population of incident-experiencing organisations did not grow. But within that population, the share that experienced 3–5 incidents rose from 30 to 50%. The share experiencing just 1–2 incidents dropped from 42 to 29%. Organisations that hit one incident are more likely to hit several in the same year. The data converges on a consistent picture: AI incidents are becoming more frequent in absolute terms, more concentrated within incident-experiencing organisations, and more visible in news coverage. The trajectory through 2025–2026 is steeper than 2022–2024. What an incident looks like in practice is illustrated by selected 2025 examples. In July 2025, Grok generated antisemitic content and hate speech after a system update relaxed safety filters; xAI removed the content within hours and issued a public statement. In March 2025, AI-generated deepfake videos impersonating Chinese actor Jin Dong were used in romance scams targeting older women, with one widely-reported case involving a woman who nearly divorced her husband to meet the scammer. In August 2025, after Joann Fabrics filed for bankruptcy, AI tools enabled scammers to clone the retailer's website and deploy fraudulent storefronts in dozens of variations within minutes. These are not edge cases. They are the kinds of incidents the trackers are now logging at hundreds per year. ## Three observations follow from the trajectory. **The first: AI risk planning frameworks built around "AI incidents are rare events" are now operating against data that contradicts the assumption.** Incidents are no longer rare. They are a regular operational category. Risk frameworks that treat them as low-frequency-high-impact events under-allocate to the monitoring, response, and mitigation infrastructure organisations now need. **The second: incident frequency is rising alongside deployment intensity, which makes the year-over-year trend hard to disambiguate.** Some of the increase reflects more AI being deployed and more incidents being possible. Some reflects better detection and reporting. Some reflects genuinely worse incident rates per deployment. The available data does not cleanly separate these three drivers. Risk planning that depends on knowing which of the three dominates should treat the question as open and plan against the upper bound of the three interpretations. **The third: the concentration effect inside incident-experiencing organisations matters for risk pricing.** An organisation that hits one AI incident in a year is, in 2025, substantially more likely to hit several in that same year than the same organisation was in 2024\. The implication: the marginal cost of the second, third, fourth incident is higher than the marginal cost of the first, because each new incident in the same year compounds reputational, operational, and regulatory exposure. Risk planning that prices incidents linearly will under-allocate to the concentration risk. > The trajectory: AI incident frequency will continue rising in 2026–2027 across all three measurement methods (AIID curated, OECD AIM automated, McKinsey self-reported). The slope may flatten or steepen depending on deployment growth, detection improvements, and mitigation maturity, but the trajectory in the data does not suggest a reversal. Risk planning that builds for "incident frequency continues to rise; the concentration effect within incident-experiencing organisations intensifies; reporting infrastructure becomes more visible to regulators and customers" will produce better outcomes than risk planning that treats 2025 numbers as a ceiling. For practitioners doing AI deployment in 2026, the planning anchor needs to shift. Incident response infrastructure (monitoring, escalation paths, communication protocols, post-incident review) is no longer optional for organisations with non-trivial AI deployment. It is now baseline operational infrastructure. Plans that defer building it produce the McKinsey data point that incident-experiencing organisations rated their incident response as "excellent" 28% of the time in 2024 and only 18% in 2025\. As incidents become more frequent, the share of organisations confident in their response capacity is dropping. By 2027, that gap is the procurement-relevant signal between organisations that have adapted to the new operating reality and those that have not. --- ### Sources - **Primary**: Stanford AI Index 2026, Chapter 3 (Responsible AI) 3.2 — [hai.stanford.edu/ai-index/2026](http://hai.stanford.edu/ai-index/2026?ref=aiadoption.org) - **Curated incident tracker**: [AI Incident Database (AIID)](https://incidentdatabase.ai/?ref=aiadoption.org), 2025 — annual incident counts 2012–2025 - **Automated incident tracker**: [OECD AI Incidents and Hazards Monitor (AIM)](https://oecd.ai/en/incidents?ref=aiadoption.org), 2026 — monthly news-based incident pipeline - **Organisation-level data**: McKinsey & Company "State of AI" Survey, 2025 — incident frequency and response quality by organisation