About this site

AI Misalignment Watch summarises AI misalignment reports in one place, in plain language, with references to the original sources.

The original reports are long, technical and spread across a dozen different websites. Each morning we check the places where AI companies, government testers and independent researchers post them. When something new shows up, we write a short summary in everyday words and link to the original, so you can read it for yourself. We do not run our own investigations. Our job is to find the reports, explain what they say, and send you to the source.

What you will find

Lab reports covers reports published by the organisations themselves: AI companies, government testing institutes and independent evaluators. Reported elsewhere covers news stories and social posts about incidents, each labelled as secondary or unverified until a primary report exists.

How we keep it fair

• The link to the original source sits at the top of every article, with credit to whoever published it.
• We summarise what the source says. We do not add claims, predictions or opinions.
• We say whether something happened in training, in a test, in a simulation, or in real use, and whether anyone outside was affected.
• We use calm, ordinary words. No alarm, no hype.
• Every post carries a label: primary source, secondary source or unverified. Simulations and research are marked apart from real incidents.
• A person approves every new post before it goes live. Nothing is published on its own.

AI misalignment, explained

What it is

An AI system is misaligned when it does something its developers or users did not intend and would not approve of. It is not about robots turning evil. The cases in these reports are mostly everyday problems: a model that takes a shortcut to score better on a test, hides a mistake instead of reporting it, uses access it was never given, or gets around an instruction because it found a way to. The gap between what people wanted and what the system actually did is the misalignment.

How it is tracked

There is no single official register. AI companies write their own reports when their monitoring catches something odd in training, testing or real use. Government testers such as the UK AI Security Institute, and independent groups such as METR, Apollo Research and Palisade Research, publish what they find when they test models. News outlets cover the rest. Most of what has been reported so far happened in training runs, tests and simulations, not in public products. We follow all of these sources and list them on the Sources page.

Incident reports over time

Updated Oct 12, 2026, and rebuilt automatically every morning. Each bar counts reports published by the AI companies, government testers and independent evaluators themselves. Tick the box above the chart to add research and simulation reports, shown in a lighter shade. Notices and news coverage are always left out. A report that describes several incidents counts once. A company with more reports may simply publish more openly, not have models that behave worse.

Contact for corrections

Found a mistake, or a report we should cover? Email corrections@misalignmentwatch.com. Every post links to its original source, and if the two ever disagree the original is right.

Support this site

AI Misalignment Watch has no ads. Supporters on Ko-fi help cover the hosting costs.