The corrections log

Every correction this record has made, newest first. The log is permanent and citable: a correction, once published, stays published.

79 corrections across 68 entries

When an entry is wrong, the text is fixed and a dated line is added at the foot of that entry saying what changed. The original claim stays visible in the correction. This page gathers every one of those lines, so the record’s repairs are as public as the entries themselves. Entries are never quietly rewritten and never deleted.

To report an error, write to the editor at [email protected] with the entry’s address and the source that contradicts it. Corrections backed by a primary document are made the same week.

5 October 2026

Jensen Huang defends AI distillation as competition

Restored the quotes per the editor’s hand-verification of the CNBC article: “That’s called competition,” “You’re allowed to test somebody else’s products all you want,” “down to bones,” “competition makes everything better,” and “know your customers, and disable the service.” Set the CNBC video as primary with its real title and the CNBC article as secondary with its real headline. Added the Bessent July “theft” remark and sanctions threat to “why” per the editor.

5 October 2026

Anthropic ships Claude Code as a research preview

Replaced the ScriptByAI and Taskade sources with Anthropic’s official launch announcement (primary) and TechCrunch’s launch coverage (secondary), ahead of the publisher-list cleanup removing scriptbyai.com and taskade.com.

5 October 2026

An OpenAI bot beats a professional Dota 2 player

Replaced the Engadget 2018 secondary, which covered the later OpenAI Five exhibition rather than the 2017 Dendi match, with TechCrunch’s 12 August 2017 coverage of the bot beating Dendi at The International.

5 October 2026

The Asilomar AI Principles are published

Corrected the publication date from 11 August 2017 to January 2017 (month precision). Contemporary coverage shows the principles were already published on FLI’s website in January 2017: GeekWire reported them on 30 January 2017 and Robohub’s February 2017 digest described them as just published, following the 5-8 January Beneficial AI conference.

26 September 2026

OpenAI announces a $110 billion funding round

Verified the OpenAI “Scaling AI for everyone” primary (live URL resolves; archived copy confirms the $110B raise and Amazon/Nvidia/SoftBank backing) alongside the two TechCrunch reports; entry published.

26 September 2026

Pentagon moves to designate Anthropic a supply-chain risk

Added the Fast Company report of 25 September 2026 as an independent second source; it confirms that Trump and Hegseth designated Anthropic a supply-chain risk in February 2026 and that Anthropic would challenge the designation in court. Entry published.

26 September 2026

Anthropic and OpenAI ship rival coding models minutes apart

Verified the Anthropic “Introducing Claude Opus 4.6” primary (agent teams, 1M-token context window in beta) against an archived copy; the OpenAI GPT-5.3-Codex primary URL resolves but its content is bot-walled, so the Codex launch details rest on the two TechCrunch reports. Entry published.

26 September 2026

xAI restricts Grok image editing after manipulated images spread

Replaced the unverifiable Yahoo Finance link with the Al Jazeera report of 15 January 2026 as the second source; the Reuters (via Globe and Mail) and Al Jazeera reports independently confirm the paid-subscriber restriction, the block on editing images of real people in revealing clothing, and the Indonesia and Malaysia bans. Entry published.

26 September 2026

Anthropic agrees to pay $1.5 billion to settle authors’ claims

Removed “and destroying the pirated copies” from the summary. Neither cited source could be re-verified for that term (the Post page is unreachable) and the Authors Guild explainer does not mention it; the $1.5 billion figure, the per-work math, and the preliminary approval are all confirmed. The “largest disclosed copyright recovery” characterization is corroborated independently (the Authors Guild called it the largest recovery in US copyright litigation history).

26 September 2026

Tesla starts a limited robotaxi service in Austin

Added the SRN News (Reuters) 2025-06-25 report as a third source; the cited CNBC and NBC pieces predate the launch and could not support the post-launch driving-incident sentence, which the Reuters reporting confirms (wrong-lane entries, drop-offs in multi-lane roads and at intersections).

26 September 2026

An AI-generated image wins a Colorado State Fair prize

The Washington Post source was unreachable on two fetch attempts, so the summary was cut back to what the Smithsonian source supports: removed the emerging-artist category detail, the “hundreds of prompts” figure, and the edit/upscale description.

26 September 2026

Google fires the engineer who called LaMDA sentient

The Washington Post source was unreachable on two fetch attempts (HTTP timeout), so the leave detail was cut back to what the CNN source supports: “paid administrative leave in June”, without the 11 June date or the confidential-material reason. Lemoine worked on Google’s Responsible AI organization and spent months in conversation with LaMDA, the company’s conversational language model. In the spring of 2022 he told executives that the system had become sentient. Google reviewed the claim and rejected it, holding that LaMDA, like its other language models, assembles plausible text without awareness or understanding of what it says. He was placed on paid administrative leave in June. Google fired him on 22 July. The dispute was reported worldwide, and the reporting turned on a question that outlived the case. A model trained to produce the most plausible next words will produce statements about its own inner life if the conversation invites them, and the fluency of those statements is not evidence that anything is behind them.

26 September 2026

Nvidia agrees to buy Arm from SoftBank for $40 billion

Added the Nvidia/SoftBank termination announcement (7 February 2022) as a source for the abandoned-deal claim. Announcement facts verified against the cited SEC 8-K (via an archived copy; sec.gov blocks automated access) and the Stack Overflow secondary.

14 September 2026

Altered Pelosi video draws Congress to manipulated media

The 2026-09-07 correction removed facts that the sources cited then did not cover. CNN’s report of 23 May 2019, now cited, confirms that the video was slowed, that YouTube removed it and that Facebook showed it to fewer people, so those facts are restored. The same report shows that President Trump tweeted a different edited clip rather than the slowed video, and the entry now says so. YouTube is listed again among the organizations.

14 September 2026

Amazon opens a cashier-less Go store to its employees

The 2026-09-07 correction dropped the date of the public opening because the sources cited then, all from December 2016, predate it. The store did open to the public on 22 January 2018. That date is restored and now cites NPR’s report from the opening day.

13 September 2026

OpenAI sells shares at $500 billion and converts to a PBC

This supersedes the amount claim in the 2026-09-01 correction. About $6.6 billion was also reported for the completed October 2025 share sale; that amount did not belong only to the 2024 funding round. Contemporary reporting distinguishes the completed sale from the larger amount authorized. The summary now identifies 2 October as the reporting date, not an established completion date. Replaced the Wikipedia overview with OpenAI’s primary announcement published on 2025-10-28. It describes talks with the attorneys general, not formal approval. The consequence is narrowed to retained nonprofit control and ownership at recapitalization, as documented in the company’s structure account.

13 September 2026

ProPublica reports racial bias in COMPAS risk scores

Added an independent reanalysis published by The Washington Post. The two ProPublica articles were from one publisher and did not meet the requirement for two independent secondary sources.

7 September 2026

Altered Pelosi video draws Congress to manipulated media

This entry originally said the video was “slowed to make her sound impaired,” spread “on Facebook and Twitter after a 22 May 2019 speech,” was “shared by President Trump,” and that “YouTube took it down; Facebook limited its reach and left it up.” Neither cited source supports the slowing mechanism, the exact speech date, Trump sharing the video, or any YouTube action; both describe only Facebook’s decision to leave an edited video up. The summary now states only what the cited sources confirm, and YouTube and Twitter have been removed from orgs.

7 September 2026

Google will not renew its Project Maven contract

This entry originally said “more than 3,000 employees” signed the letter opposing Project Maven. The cited NBC News source, published the same day as the event, says more than 4,600 employees had signed by 1 June 2018; corrected.

7 September 2026

Amazon opens a cashier-less Go store to its employees

This entry originally said the public opening “did not arrive until January 2018,” a fact neither cited source (both from December 2016) supports. Replaced with the sources’ own contemporaneous claim that Amazon planned a public opening in early 2017.

7 September 2026

Google puts neural machine translation into production

This entry originally said the switch handled “tens of millions of translations a day.” The cited Google Research post gives an exact figure, “about 18 million translations per day” for the Chinese-to-English pair; corrected.

7 September 2026

ProPublica reports racial bias in COMPAS risk scores

The “Machine Bias” article was labeled a primary source. It is a news investigation, not an official document from Northpointe (the maker of COMPAS), so it is relabeled secondary; both cited sources are now secondary, consistent with the rule that a news article is never primary.

7 September 2026

Microsoft takes the Tay chatbot offline after 16 hours

This entry originally said Microsoft withdrew Tay “within about 16 hours and 96,000 tweets.” Neither cited source gives a tweet count, so that figure is removed; the 16-hour figure is confirmed by the cited TechCrunch report.

7 September 2026

Nature publishes the paper behind AlphaGo

The Google Research listing of this paper was labeled a secondary source. It is Google’s own index of its researchers’ paper, not independent reporting, so it is relabeled primary.

1 September 2026

Genome language models design working bacteriophages

The cited source title originally read “Generative design of novel bacteriophages...”; Crossref’s record for the published paper gives the title without “novel,” corrected to match. Verified against the publisher’s own bibliographic metadata (science.org itself refused direct fetches); date, authorship and the phage-resistant-bacteria result all check out.

1 September 2026

US lifts export restrictions on Anthropic’s top Claude models

The Al Jazeera source URL had a typo (missing “-on-”) that made it a dead link; corrected. The summary’s description of who the restrictions covered was also loosened to “some foreign-national staff”; both cited sources say the order covered all foreign nationals, including Anthropic’s own employees, and Anthropic shut the models off entirely because it could not filter by nationality on short notice; corrected. Both sources now confirmed by direct fetch; moved out of draft.

1 September 2026

OpenAI shuts down the Sora video app

This entry’s “why” originally said the CFO’s compute-shortage remark came “days earlier”; she made it in a 24 March 2026 interview, about a month before this 26 April closure (the cited TechCrunch piece is close in time to the CFO’s remark, not to this later event); corrected.

1 September 2026

TSMC introduces the A13 chip process node

This entry originally dated the A13 announcement to 23 April 2026. TSMC’s press release carries an Apr. 22, 2026 dateline; corrected, and the entry’s filename date corrected to match.

1 September 2026

Father sues Google over Gemini’s role in son’s death

This entry originally said “the parents” of Jonathan Gavalas sued Google. Reporting on the complaint (Fortune, 2026-03-05) identifies the plaintiff as his father, Joel Gavalas, alone; corrected to “father” and Joel Gavalas added to the entry’s people.

1 September 2026

Pentagon designates Anthropic a supply-chain risk

This entry’s summary named “5 March 2026” as the designation date, more precise than the entry’s own datePrecision: month. Contemporary reporting is split between 5 and 6 March for the Pentagon’s formal notification; loosened to “early March 2026” to match what the sources actually support.

1 September 2026

OpenAI sells shares at $500 billion and converts to a PBC

This entry originally said the October 2025 secondary share sale was worth $6.6 billion. That figure belongs to OpenAI’s October 2024 funding round; the cited Wikipedia source states the October 2025 sale was “up to $10 billion,” corrected accordingly.

1 September 2026

Waymo’s robotaxi rides roughly triple to about 14 million

This entry originally said Waymo’s paid trips reached “roughly 250,000 a week in mid-2025,” a figure neither cited source supports. Corrected to the eWeek-reported average of roughly 269,000 paid trips a week over the year; the eWeek source’s publication date was also added.

1 September 2026

Newsom vetoes California’s SB 1047 frontier AI bill

This entry originally said Newsom “directed state agencies to draft rules instead.” The cited press release says he asked outside GenAI experts to help California draft guardrails, and separately ordered state agencies only to expand their assessment of catastrophic risks; corrected.

1 September 2026

AI-cloned Biden robocall targets New Hampshire primary voters

This entry originally dated the robocalls to 23 January 2024, the date of the New Hampshire primary. The FCC’s forfeiture order states the calls went out two days before the primary, on 21 January; corrected, and the entry’s filename updated to match.

1 September 2026

Microsoft announces multibillion-dollar investment in OpenAI

This entry originally attributed the roughly $10 billion investment figure to “Bloomberg and other outlets.” The cited TechCrunch source attributes that figure to Semafor’s reporting on a proposed deal, and Bloomberg is not mentioned in either cited source; corrected.

1 September 2026

OpenAI releases ChatGPT as a free research preview

This entry originally attributed a description of ChatGPT writing “essays, scripts” to unnamed researchers, and folded OpenAI’s own “plausible-sounding but wrong” caveat into that same attribution. The cited Guardian article describes it as capable of writing code and solving problems, and the “plausible but wrong” phrasing is OpenAI’s own; corrected to attribute each claim to its actual source.

1 September 2026

Frances Haugen testifies to Congress on Facebook’s algorithms

This entry originally described Frances Haugen as a “former Facebook data scientist” and cited a “downstream MSI” effect of the algorithms she testified about. Both cited sources describe her as a former Facebook product manager, and neither supports the MSI claim; the title is corrected and the MSI claim removed.

1 September 2026

Anthropic emerges with $124 million and a safety focus

This entry originally described Daniela Amodei as “previously vice president of safety and policy” at OpenAI. Neither cited source supports that title; corrected to identify her only as Dario Amodei’s sister, per the cited reporting.

1 September 2026

OpenAI opens GPT-3 to developers through a private API beta

This entry originally said the private beta’s dozen partner companies worked on “search, sentiment analysis and moderation.” Neither cited source supports that description; corrected to name three of the actual partners (Algolia, Quizlet, and Reddit) per OpenAI’s own post.

1 September 2026

Reporting reveals Cambridge Analytica’s use of Facebook data

This entry originally said Cambridge Analytica obtained data from up to 87 million Facebook profiles. The cited 17 March 2018 reporting said 50 million; 87 million was a Facebook estimate published in April 2018, after the events this entry describes. Corrected to 50 million throughout.