How this works

We count what people say happened. We don't grade their LLM, and we're not pretending to.

Voting

• One vote per hashed client per UTC day.
• Hashes are salted server-side and not reversible to you.
• Votes reset at 00:00 UTC.

Confessions

• Required: the model, what happened, the failure type.
• Optional: impact area, what evidence you have, severity, did it double down.
• We don't host files. "evidence" is a label, not an upload.

Most reported

• The default view — raw share of submitted reports for each model.
• Not adjusted for how many people use that model overall.
• A model with more users will likely appear more here, all else equal.
• This view powers the daily results, monthly rankings, and the heatmap.

Relative to usage

• A second view, only shown when there is enough first-party data to support it.
• Shows complaints per 100 participating users who said they used that model in the window.
• The denominator comes from an optional weekly survey: "Which AI did you use most this week?"
• This is self-reported. It reflects the people who use this site, not the general population.
• Each model is only shown when at least 10 respondents reported using it in the window. Below that threshold, the result is suppressed.
• This is not a market-wide claim. Do not treat it as one.

What's public

• The feed, the heatmap, monthly ranks, 30-day trends.
• Flagged records are excluded from all of those, pending review.
• Rankings are volume of votes only unless you switch to the Relative to usage tab.

Time

• Everything is UTC. The monthly view resets on the 1st at 00:00 UTC.
• Usage survey responses are bucketed by UTC week (Monday start).