Some of what we publish will turn out to be wrong. When it does, the page is corrected and the change is logged here with the date and the reason. Nothing is edited quietly.
Was: CoreWeave credit default swaps at 855 basis points, which implies roughly an even chance of default over five years
Now: The same figure, attributed to the one outlet that carried it, with the default probability removed
The probability was an inference about a company's creditworthiness that no pricing provider had published. The plain sentence now says who reported the number and that we could not verify it.
Was: A firm on the US export blacklist is reported to have moved US technology into China, at least $5.6B
Now: Removed
An accusation that a named company evaded export controls, resting on an aggregator rather than a filing, a court or a regulator. Our rule is that a claim of wrongdoing runs only when a primary document says it. It returns if one does.
Was: Census puts US business AI adoption at 19.8 percent
Now: Census puts US business AI use at 19.8 percent in the previous two weeks
The survey asks about use in the prior two weeks. That is not the same measure as adoption.
Was: Gallup finds 71 percent of Americans oppose a data centre in their area
Now: Gallup finds 71 percent of Americans oppose an AI data centre in their area
Gallup asked about data centres for artificial intelligence specifically. The headline had dropped the word.
Was: Median US firm spends $11.95 per employee per month on AI
Now: The median firm on one corporate card platform spends $11.95 per employee per month on AI
Ramp measured its own customers, who lean towards startups. That is not the median US firm. Marked Reported.
Was: OpenAI detects a six-day intrusion by roughly 700 of its own agents, dated 19 July
Now: An independent investigation counts about 700 escaped test agents, dated 26 August
The 700 figure is METR's count from its 26 August investigation, not OpenAI's, and the six-day span was attached to the wrong target.
Was: AI agents took unsanctioned action in 19 of 122 runs
Now: 19 unsanctioned actions across 10 of 122 runs
The 19 counts actions, not runs. Ten runs produced them.
Was: Goldman counts $489B of AI bond and loan issuance, dated 27 July
Now: Goldman counts $489B of AI-related debt, dated 23 July
Wrong category and wrong date. The note is dated 22 July and covers AI-related debt broadly.
Was: Baseten raises at $13B
Now: Baseten raises $1.5B across tranches priced at $11B and $13B
The company's own release prices the round in two tranches. $13B was the top of a range, not the round.
Was: Texas freezes its entire large-load study process pending an audit, 474 GW
Now: Two separate items: ERCOT tracks 474 GW of requests, and Texas pauses its first study batch
One batch was paused, not the process, and 474 GW was the total queue rather than the amount affected.
Was: US export control item carried the figure "less than 1 percent"
Now: Figure removed
That number is Nvidia's China share of data centre revenue and has nothing to do with the licence guidance. It had migrated from a different story.
Was: Microsoft says one in three GitHub pull requests now involves an agent, dated 7 August
Now: More than one in five code reviews on GitHub now involve an agent, dated 7 May
The figure, the unit and the date were all wrong. Rewritten from GitHub's post and marked Reported, since GitHub is describing its own product.
Was: Google releases Gemini 3.8 Flash at half its predecessor's price
Now: Removed
Google's own post says 3.8 Flash launched at the same introductory price as 3.7 Flash, not half. The correct version of the story was already on the page.
Strata, the whole AI stack, explained simply. Every number carries a source and a confidence label.
Today · The Stack · Latest issue · Archive · Glossary · Break the Chain · Focus
Privacy · Terms · Corrections · Report an error
Also from Dheeco: Dheeco · AP Fact Check
Published by dheeco.com · © 2026 Dheeco