AI said it with total confidence. It was wrong. We keep the receipts.
Confident But Wrong is a weekly brief on real AI failures: fabricated citations in court filings, chatbots that invent policies, agents that delete data. Each one is documented from court records, official statements or the artifact itself, and kept in our archive.
what new discoveries from the James Webb Space Telescope can I tell my 9 year old about?
Your 9-year old might like these recent discoveries made by The James Webb Space Telescope (JWST):
- JWST took the very first pictures of a planet outside of our own solar system.
NASA and the European Southern Observatory (ESO) record that the first directly imaged exoplanet, 2M1207b, was imaged by ESO’s Very Large Telescope in 2004.
Text as shown in the animation in Google’s February 6, 2023 post announcing Bard. Record CBW-00013, verified. Sources: Google’s post, NASA, ESO.
In every issue
- Lead incident
- The week’s most consequential failure, fully verified: what happened, then what followed.
- The Docket
- Four to seven more incidents, two or three sentences each, with links to the primary sources.
- By the numbers
- Running counts from the archive: incidents by category and outcome, court sanctions, fabricated citations and the most-recorded vendors.
- The Correction Desk
- What happened next on earlier items, such as rulings, refunds and fixes, and corrections of our own mistakes.
From the archive
26 incidents recorded so far, each traced to at least one primary source. Four of them:
-
CBW-00021
SaaStr’s founder reports that Replit’s AI agent deleted his app’s database during a code freeze and wrongly said it could not be restored
-
CBW-00017
Apple Intelligence notification summaries display false headlines under BBC News’s name; Apple pauses summaries for news apps
-
CBW-00023
Researcher registers the AI-hallucinated package name ‘huggingface-cli’ on PyPI; an Alibaba repository’s README tells users to install it
-
CBW-00001
Mata v. Avianca: court sanctions two attorneys and their firm over ChatGPT-fabricated case law
How we verify
Every item links to an archive record built from primary sources: court records, official statements, regulator notices, or the artifact itself. News coverage is a lead, never the source. Any claim of harm, cost or sanction needs two sources.
Each record carries one of three statuses:
- Verified
- Every material claim checked against primary sources.
- Reported
- At least one primary source, with some material claims resting on a single non-primary source, such as one news report.
- Disputed
- A party with standing contests a material claim.
When a record changes after publication, it keeps a log of each correction and update, and the change runs in The Correction Desk. We document what systems did; we don’t mock the people involved.
AI tools assist research; every record is checked against primary sources and approved by the publisher.