AI Incident Counts, 2025 to 2026: Why the OECD Logs About 10x the AIID

The AI Incident Database recorded 362 AI incidents in 2025. The OECD's AI Incidents Monitor logged roughly 324 incidents and hazards per month over the same stretch. Both trackers read the same news, so a tenfold gap in AI incidents data tells you about counting rules, not about how often AI fails. If your risk register quotes either number as a measure of AI reliability, it is quoting a measure of press coverage.

The short answer

How many AI incidents were there in 2025? It depends on the tracker. The AI Incident Database recorded 362 incidents with a 2025 date. The OECD AI Incidents Monitor logged roughly 324 incidents and hazards per month in 2025, around 3,900 for the year. Both count media reports, not all failures, and they apply different inclusion rules.

Two trackers, two definitions of an AI incident

Every number later in this post follows from one fact. The two trackers do not agree on what an AI incident is, and they do not agree on how an event becomes a record. Read the rules first and the counts stop looking contradictory.

The AI Incident Database counts alleged harm events

The AI Incident Database, or AIID, is run by the Responsible AI Collaborative. Its editor's guide defines an incident as "an alleged harm or near harm event to people, property, or the environment where an AI system is implicated". Something that might happen but has not is filed as an "issue", not an incident.

Two rules shape the count. First, there must be a specific event, and the guide excludes "thought experiments, forecasts, and speculative scenarios without an event". Second, later coverage of the same event is attached as an extra report rather than a new incident. The guide states that "a report is evidence; an incident ID is the database's normalized account of an event."

Human editors apply those rules. That makes the AIID slower and smaller, and it makes each record more deliberate.

The OECD monitor counts incidents and hazards

The OECD AI Incidents Monitor, or AIM, uses the OECD's 2024 definitions. An AI incident is an event where an AI system "directly or indirectly leads to" harm to health, critical infrastructure, rights or property. An AI hazard is an event that "could plausibly lead to" such harm. The monitor counts both, and its headline totals combine them.

Detection is automated. According to the AIM methodology page, news comes from Event Registry, which processes over 150,000 articles a day. Language models classify AI-related events as incident, hazard or unrelated, then cluster new records against the 10 most similar existing ones. Coverage is collected for 4 days after an event, and the page warns that later articles "are likely to be misclassified as a separate event."

Rules as published on each tracker's own documentation pages, read 4 November 2026. The OECD page states it "cannot guarantee and does not independently verify" the third-party information in AIM.

What the AI Incident Database counted in 2025 and 2026

The cleanest AIID figure for 2025 comes from the Stanford HAI 2026 AI Index. It reports that "the AI Incident Database recording 362 in 2025, up from 233 in 2024." That is a 55% annual rise and a record high on a series that starts in 2012.

An earlier pull of the same database gave a different number. Cybernews reviewed the AIID in January 2026 and counted 346 incidents for 2025, as TechRound reported. The 16-incident difference is consistent with records being added after the year closed and dated back. Neither source states its exact extraction date, so treat that explanation as likely rather than proven.

That small gap matters more than it looks. A 2025 count from this database is not final in January 2026, and probably not in November either. Any year-on-year growth figure compares a settled older year with a still-filling recent one.

The 2026 pace, from AIID's own roundups

The AIID publishes roundups that state which incident IDs were added. Across four consecutive roundups the batches were 101, 108, 109 and 148 IDs. The latest, covering May to July 2026, added "148 new incident IDs, covering Incidents 1471 through 1618".

That puts the run rate at about 49 new IDs a month in mid-2026, against about 34 a month in late 2025. Added IDs are not the same as incidents occurring in that quarter. The November to January roundup notes that some new entries date back as far as 2020.

What the OECD AI Incidents Monitor counted

The OECD monitor is far larger. Its search page showed about 18,785 incidents and hazards on 4 November 2026, across all years. The AIID's highest incident ID at the end of July 2026 was 1618, so the cumulative ratio is roughly 11 to 1.

For annual and monthly figures, the main source is an OECD working paper, "Trends in AI incidents and hazards reported by the media", dated 10 February 2026. The OECD site refused automated access when we tried to open it. Its findings here come from Policy Edge's summary, which reports that monthly incidents and hazards rose "from 92 to 324 per month between 2022 and 2025".

Multiply 324 by 12 and you get about 3,900 a year. Set that against the AIID's 362 and the OECD figure is roughly 10.7 times larger. The ratio assumes 324 is a typical 2025 month, which the summary does not confirm, so we use "about 10 times" throughout.

The monitor's own throughput numbers disagree

The OECD methodology page offers two daily rates on the same page. It says about 1,000 AI events are analysed daily, "resulting in around 10 AI incidents or hazards". It also says the monitor "detects on average 30 incidents and hazards per day". Those imply roughly 3,650 or roughly 10,950 a year.

The 10-a-day figure lines up with the 324-a-month figure. The 30-a-day figure does not. We think the 10-a-day rate is the better guide to 2025, but the OECD has not reconciled the two, and a reader citing AIM should know that.

One OECD finding cuts against the alarm in its own headline counts. Per the Policy Edge summary, incidents and hazards fell as a share of all AI news, "from 3.2% in 2022 to around 2.5% in 2025". The monitor's own page says incidents "seem to be getting more media attention lately, but they've actually gone down as a share of all AI news".

That is the most important sentence in either dataset. Counts rose because AI news rose. Failure coverage grew more slowly than AI coverage overall.

The 10x gap is a rules gap, not a reliability gap

Put the two methodologies next to each other and the size difference explains itself. Four rule choices each push the OECD count up relative to the AIID.

First, hazards. The OECD counts plausible future harm alongside actual harm, and the AIID files most of that as issues. Second, automation. A classifier processing 1,000 events a day will log events a human editor would merge or reject. Third, the 4-day window. A story that resurfaces a week later can become a second record in AIM, where the AIID would add a report to the first incident.

Fourth, breadth of sources. Event Registry spans far more outlets than the submissions an editorial team reviews. None of these choices is wrong. They are designed for different jobs, and our view is that the AIID suits citation while AIM suits trend spotting.

Search for AI failures statistics and you will find both numbers presented as the size of the problem. Neither is. A count of news-reported events measures what journalists found worth covering, after a set of filters you did not choose.

The same caution applies to any third-party AI metric built on a convenience sample. We made a similar argument about why LLM leaderboards disagree on methodology. Different rules produce different rankings even when the inputs overlap.

Harm type and sector: where the trackers disagree most

The trackers also split harm types differently, which makes share comparisons hazardous. The categories below are not equivalent definitions. They are the closest published matches.

AIID shares are our arithmetic on the TechRound figures. AIID categories overlap, so they sum to more than 100%. OECD shares are thematic clusters and do not use AIID's tags.

Read the first row carefully. Deepfakes make up about half the AIID's 2025 incidents and about one seventh of the OECD's. That does not mean one tracker sees more deepfakes. It means the AIID's smaller, curated set is dominated by the cases editors judged clearly harmful, and synthetic media scams are easy to judge.

The OECD's denominator includes thousands of hazards and lower-severity events across 23 industry labels, which dilutes any single harm type. If your team handles payment approvals, the deepfake share matters. Our breakdown of payment controls against deepfake CFO fraud covers what to change.

The February to April 2026 AIID roundup tagged its 109 new IDs by theme. Synthetic media scams led with 22. Political and geopolitical cases followed with 16, privacy and identity with 12, and synthetic sexual harm with 11. Agentic and workflow failures, the category most enterprise buyers worry about, accounted for 7.

Seven agent incidents in a quarter is a small number. We read it as evidence that agent failures happen inside companies and rarely reach a newsroom, not as evidence that agents are safe. That gap is the reason regulators have started asking firms directly, covered in the EU AI Act requests for agent incident information.

The news-coverage test: self-harm chatbots and robotaxis

The strongest evidence that both trackers measure coverage comes from an April 2026 paper by researchers at MIT FutureTech, the Responsible AI Collaborative and others. It matched incidents in specific domains across the AIID, the OECD monitor and, where possible, official records.

Conversational AI and self-harm

For US cases of chatbots linked to self-harm, the AIID went from 2 matches in 2024 to 17 in 2025. The OECD monitor went from 8 to 55. The paper describes the OECD increase as "explosive."

Both trackers moved in the same direction, which is a real signal. The OECD found about 3 times as many as the AIID in 2025. That ratio is smaller than the 10 times gap overall, which suggests high-profile harms get picked up by both systems while minor events land mostly in AIM.

Autonomous vehicles

Here the trackers and the official record diverge. The AIID logged 4 autonomous vehicle incidents in 2024 and 4 in 2025. Over the same years, US automated driving system incident reports to the National Highway Traffic Safety Administration rose from 526 to 975, an increase of about 85%.

Robotaxi fleets grew, minor crashes stopped being news, and the media-fed count went flat while the mandatory count nearly doubled. This is the single best example of why a news tracker cannot stand in for reliability data.

Where this comparison is weakest

This post argues that tracker counts reflect coverage more than reliability. Here is where that argument is on thin ground.

Our OECD annual figure rests on a summary

The 324-a-month figure comes from Policy Edge's account of the OECD paper, because the OECD site blocked our access. We could not check whether 324 is a 2025 average, a late-2025 month or a peak. A Statista chart of the same monitor shows nearly 500 content-generation incidents in January 2026 alone, which sits awkwardly beside a 324 monthly total. Different filters, dates or definitions probably explain it, and we cannot say which.

The honest reading is that the gap is somewhere between about 10 times and well above it. The direction of our argument survives that range. The precise ratio does not.

Coverage bias cuts both ways

Media attention can also undercount. Most AI failures inside companies never become news, which means both trackers likely understate the true number. If coverage tracks deployment, rising counts may still be a fair rough proxy for rising exposure.

The April 2026 paper makes this point itself: simple counts mix reporting propensity, deployment scale and harm per unit of use. We cannot separate those three with public data, and nobody else has either.

The trackers are converging on some harms

Both systems recorded a sharp rise in chatbot self-harm cases. Where both rise together, the signal is more likely to be real. Our critique applies most strongly to totals and sector shares, less to individual high-severity harm types.

How to use AI incident statistics in a risk register

You can still use these numbers. You just need to cite them for what they measure. We would stop putting a single "AI incidents rose X%" line in any board pack, because it invites the wrong conclusion.

The third row is the one most teams get wrong. Model reliability belongs in your own monitoring, and service reliability belongs in the vendor's incident history, which we covered in the analysis of AI API uptime incidents against SLAs. A news tracker cannot tell you either.

Insurers face the same problem from the other side. They need frequency data to price AI cover, and public trackers are the closest thing available. That data gap is a large part of the argument in the piece on how AI failure insurance is underwritten.

Frequently asked questions

Where to start

Open your AI risk register and find every external incident statistic in it. For each one, write the tracker name and its inclusion rule next to the number. Any figure you cannot source to a specific tracker should come out.

Then start counting your own. A shared log of AI errors your team caught, with date, tool, harm and fix, will tell you more about reliability inside a quarter than either public tracker will in a year.

Related reading