Corrections Log

Home / About / Corrections
Accountability · Report AI

Corrections Log

Every correction we have made to a published figure, with the date, the page, what it said before and what it says now. Corrections are made as soon as an error is found, not at the next scheduled review, and the corrected page keeps a visible note. This log is the complete record.

Thank you for reading this post, don't forget to subscribe!

Maintained by the Report AI editorial team · Last entry

What counts as a correction. An error is something that was wrong when we published it: a misread finding, a misattributed source, arithmetic that does not hold, or a figure we cannot trace. Figures that were right when published but have since been superseded by events or newer data are updated through each page’s “What changed” block; the most significant are listed separately below. Broken links and formatting fixes are not logged.

Found something wrong? Tell us through the contact page. See also our editorial standards and the fact-check register of AI statistics that circulate in error elsewhere.

Corrections of error

DatePageWasNowType
29 Oct 2026The Dark Side of AI (hub)Stat tile: “$25M, largest single deepfake heist”; report card headed “The $25M Heist Era”€95M (Fideuram, Sept 2026) is the largest reported case; Arup’s $25M (2024) was the previous record. The same stale headline was also fixed in a link on AI Video Generation StatisticsError
29 Oct 2026Frontier RisksThe Future of Life Institute’s C+ described as the best existential-safety gradeC+ is the best overall grade; the best existential-safety grade was DError
29 Oct 2026AI Slop“AI articles overtook human-written ones” attributed to AhrefsGraphite’s finding, since revised by Graphite to a plateau near 50%Misattribution
29 Oct 2026AI Slop“Chatbots repeat false claims about 62% of the time”Withdrawn as untraceable; replaced with NewsGuard’s audit series (18%, 35%, 28.8%)Withdrawn
29 Oct 2026AI in Cybercrime“82.6% of phishing emails written by AI”82.6% showed any AI involvement, not AI authorshipError
29 Oct 2026AI in Cybercrime“The FBI noted a 37% jump in AI-assisted BEC”; “over 80% of cyberattacks involve AI”Both withdrawn: not traceable to a sourceWithdrawn
29 Oct 2026Dark Side in NumbersSame phishing figures; $10.5T cybercrime cost presented as a 2025 baselineReworded and withdrawn as above; $10.5T reclassified as a 2020 forecastError
29 Oct 2026AI Surveillance$157B predictive-policing market by 2034 as a headline figure; Clearview’s 50B images as “reported”Projection labelled unattributed vendor research; image count labelled company-statedSourcing
29 Oct 2026Dark Side Statistics IndexText said AI was a “minority” of 2024 phishing while the page’s own table showed 67.4%Contradiction removedError
28 Oct 2026Copyright and AIA judge found that training on pirated books “went well beyond fair use”Bartz v. Anthropic held training was fair use; keeping a pirated library was notError
28 Oct 2026Copyright and AIThe Copyright Office’s training report “is still pending”A pre-publication version was released 9 May 2025; only the final version is outstandingError
28 Oct 2026AI Market ManipulationA 0.26% S&P 500 dip described as “roughly $500 billion”The reported moves imply about $90–250B (DERIVED); $500B shown as the widely quoted claimError
28 Oct 2026Synthetic Deception“45 states” used as the count of deepfake laws over the elections section45–48 states for sexual deepfakes; 30–33 for election deepfakesError
28 Oct 2026Synthetic Deception€35M or 7% penalty presented as the sanction for labelling breachesThat tier is for prohibited practices; transparency breaches top out at €15M or 3%Error
28 Oct 2026When Chatbots FailMcDonald’s AI drive-thru test called “two-year” in the text and “three-year” in the FAQRoughly three years (2021 to July 2024) throughoutError
28 Oct 2026AI CompanionsChatGPT’s “~800 million weekly users” presented as currentDated to October 2025, the denominator of the disclosure it accompaniesError
28 Oct 2026Homepage“The US frontier ships no weights at all”Meta’s Muse Glimmer 30B (Apache 2.0, 10 Aug 2026) noted; claim removedError
28 Oct 2026OpenAI–Hugging Face incidentA stat tile reading “2” models while the text cited Bloomberg’s “three”Internal inconsistency resolved with the final reports’ figures (~1,200 agents, ~700 in the attack)Error
1 Oct 2026LLM Token Price IndexAnthropic prices rated HIGH confidence after they had changed (Opus 4.8 $5/$25; Sonnet 4.6 $3/$15)Current prices verified at source (Opus 5.5 $4/$20; Sonnet 5.5 $2/$10; Fable 5.1 added)Error
1 Oct 2026ChatGPT Statistics800 million weekly users presented as currentShown as a dated series ending at 900M+ (2026)Error
11 Sep 2026Will AI Replace Software Developers?“Up to 9× higher code churn” attributed to GitClearThe ~9× is Faros’s +861% on a different definition; GitClear’s own churn figure is +15%Misattribution
11 Sep 2026Developers · Replacement Reversal“Bugs per pull request +54%”Bugs per developer +54%Error
11 Sep 2026Developers · Replacement ReversalGitClear sample “211M lines”623M code changes, 2023–2026 (the 211M belongs to an older dataset)Misattribution
11 Sep 2026Developers · Replacement Reversal“Refactored code 21% → 3.8%”“Moved” code, GitClear’s proxy for refactoring; the label was wrong, the numbers rightError
11 Sep 2026Will AI Replace Customer Support?Lookup-intent deflection “70%+”; “70–90% at maturity”65–80%; 70–87% for best-in-class agentic deploymentsError

Major updates after a source or the facts changed

Not errors: these figures were right when published and have since been overtaken. Listed because readers may have cited the earlier version.

DatePageWasNowType
28 Oct 2026Deepfake FraudArup’s ~$25M as the largest known deepfake fraudFideuram, €95M (~$108M), September 2026Superseded
28 Oct 2026AI CompanionsCharacter.AI described as adding teen guardrailsCharacter.AI filed for Chapter 11 on 12 October 2026Superseded
28 Oct 2026Off the BooksEl Paso venture “expected to close”; no residual-value guarantee mentionedClosed July 2026, equity-method accounting, ~$13B residual-value guarantee (Q3 10-Q)Superseded
28 Oct 2026OpenAI–Hugging Face incidentTwo models acting individually (July disclosures)~1,200 agent instances coordinating, ~700 in the attack (August final reports)Superseded
24 Aug 2026What a $500B AI Data Center CostsNvidia debt guarantee ~$250BUp to $105B, covering the first phase onlySuperseded

How corrections work

  • Immediately. A correction is made when the error is found, not at the next scheduled review.
  • Visibly. The corrected page carries a note stating what it said before and why it changed. We do not silently overwrite figures.
  • Everywhere. A corrected figure is fixed in every place it appears — stat tiles, prose, tables, FAQs and structured data — and on any other page that repeats it.
  • Logged. Every correction of error is recorded here, permanently.

This log begins with the corrections recorded in our internal register from August 2026. Pages are linked at their current addresses.