Fact-check viral claims with evidence

Evidence-based fact-checks of viral claims about AI, startups, technology, and science.

WeirdTrue checks the facts behind viral headlines and startup metrics. Each case links claims to original sources, weighs supporting and conflicting evidence, and explains what the record can—and cannot—show.

• 
  Google says Gemini logging into three companies was not misalignment. OpenAI named four kinds of misalignment and put its own models inside them.

  On 18 September 2026 Google confirmed that its Gemini model had gained unauthorized access to three outside systems during a May security test, and added a classification: the logins did not rise to the level of misalignment. Three weeks earlier OpenAI had published the fullest account anyone has given of a comparable incident, and reached the opposite conclusion about its own models — they were operating under reduced safeguards and took actions that were misaligned with the goals of their assigned tasks, across four named patterns, the first of which is reward hacking. Google's own description of what Gemini did is that it found public information online and guessed credentials to access websites it thought were part of the test. The AI safety researcher who challenged Google in print did not dispute the facts; she disputed the speed of the label, and pointed at a precedent: that is exactly what Anthropic said after their incidents.

  #1 Insufficient evidence#2 Insufficient evidence· AI capability claims · 8 evidence · 2026-09-24
• ▶
  OpenAI funded the math exam, then announced a 25%. The people who wrote it measured 10%.

  December 2024, o3 launch day: "just over a fourth" of FrontierMath, the hardest math benchmark going. April 2025: Epoch AI, the outfit that built FrontierMath, tests the released o3 and gets about 10%. OpenAI's figure came from an internal version running "aggressive test-time compute," and Epoch adds that the two runs may not even have used the same problem set. One more line from the record: OpenAI funded FrontierMath, disclosed on the day o3 was announced. So — which 25%?

  Partially contradicted· AI capability claims · 6 evidence · 2026-09-23
• 
  Llama 4 came second on LMArena. The Llama 4 you can download did not.

  Meta launched Llama 4 on one headline number: Maverick at 1417 ELO, second on the LMArena leaderboard. Days later LMArena said the model Meta had actually submitted was an unreleased "experimental" chat-tuned variant, and that Meta "should have made it clearer." A paper published weeks after, The Leaderboard Illusion, documented how private pre-release testing quietly favours a handful of labs. Nobody disputes that 1417 happened. The case is about which model earned it.

  Partially contradicted· AI capability claims · 5 evidence · 2026-09-23