Corrections

AI Lately corrects errors promptly and in the open. This page is the permanent log. Each entry names the piece, describes the change, and records the date.

Report an error

Write to editor@ailately.com with the URL of the piece and the sentence in question. The desk acknowledges every report within two business days and resolves clear factual errors within five. A corrected piece carries a dated note describing the change. Disputed interpretations get a reply from the editor and, where warranted, a published response from the subject.

Automated reports

Reports drafted by the automated newsdesk fall under this policy in full. A reader who spots an error in an automated piece uses the same address, and the editor can pull any automated piece immediately while the record gets checked.

The log

Entries begin on launch day, Sept. 4, 2026, and will appear here, newest first, in this format: date, piece, what changed, and why.

• Oct. 25, 2026, "The 6 Best AI Cooking Tools, Ranked for November 2026." The piece now ranks six tools, adds Eat This Much and gives it the Plan trophy, and lists PlateJoy and Mealime with Yummly as closed apps. It cites Mealime's homepage notice that the app shut down on Oct. 21, 2026, and an Internet Archive capture from July 10, 2025 that shows an error page on PlateJoy's site. It gives each monthly price from Apple's App Store listings or the company's own page, including $19.99 for ChatGPT's top plan, drops claims about DoorDash's ChatGPT app and ChatGPT's Instant Checkout that its sources date to December 2025, adds a short section pricing three cooking gadgets, drops the category trophy badges, and adds the New York Times' Oct. 23 account of NYT Cooking's chat tool, with a note that the Times has sued OpenAI and Microsoft. An earlier version ranked PlateJoy fifth as the Plan winner and called it the costliest app at $12.99 a month, ranked Mealime sixth with a $5.99 Pro tier, put ChatGPT's top plan at $20 and cited a buyer's guide and two review pages that still described both services as live. Eat This Much, Instacart, Apple, Claude, the Internet Archive and Mealime support the corrected wording. The piece first ran as "The 7 Best AI Cooking Tools, Ranked for November 2026."
• Oct. 20, 2026, "Junior Coders Fell to 9% of Anthropic's Technical Hires." The column was rewritten the day it ran so its argument reads plainly: Anthropic's chart records one company's choice, and the wider data show coding job posts rising mostly for senior roles while other heavy AI users added entry-level staff. It now says the 27 percent to 9 percent fall runs from early 2025 to Anthropic's Oct. 16 report, that Anthropic's technical hiring still grows in raw numbers, that overall U.S. job posts fell 7 percent, and that Indeed's 71 percent senior and 37 percent AI-titled shares overlap. It credits the 3.8 percent yearly decline for workers aged 22 to 25 to Brynjolfsson's team, adds the Ramp and Revelio Labs study NPR reported and a quotation from Indeed's Sneha Puri, and names Todai as AI Lately's daily email. An earlier version dated the 9 percent figure to August, said every other job saw postings fall 7 percent, called the 37 percent "another" share on top of the 71 percent, credited Brynjolfsson's team with a 16 percent drop that Fortune's article leaves out, and called The Signal the daily email. Anthropic, CNBC, Fortune and NPR support the corrected wording. The piece first ran as "Anthropic's Junior Hiring Chart Tracks Only Anthropic."
• Oct. 17, 2026, "SpaceX, Now a Super Intelligence Company, Spends 86% on AI." The piece now reports that SpaceX's IPO raised $85.7 billion in net proceeds, that Anthropic's compute contract runs through 2029, that Trump's order gives the president's science adviser 60 days to submit proposed legislative language, that Musk said on Oct. 4 that SpaceXAI would become SpaceXSI, that Futurism quoted Brian McManus, and that xAI's shareholders received 0.1433 SpaceX shares for each xAI share, with a cash option for some executives. It cites SpaceX's 10-Q for the capital spending figures, and it drops a $25 billion bond claim and a $108.27 closing price. An earlier version said the IPO raised about $75 billion, put the Anthropic contract through May 2029, said agencies had 60 days to draft legislative language, said Musk wrote on X on a Saturday, said McManus told Futurism, cited the earnings release for the capital spending figures, gave a lowest close of $108.27 and a $25 billion bond sale ten days after the IPO, and called the xAI deal all-stock. SpaceX's 10-Q, TechCrunch, The White House, IBTimes UK, Futurism and Reuters support the corrected wording. The headline now reads as it does here; the piece first ran as "SpaceX Calls Itself a Super Intelligence Company and Puts 86 Percent of Its Capital Into AI."
• Oct. 17, 2026, "Entry-Level Work Is Where AI Shows Up First." The piece now reports Stanford's finding as employment 19 percent below its expected pace for workers aged 22 to 25 in AI-exposed jobs, names the New York Fed authors Natalia Emanuel and Nick Montalbano, describes the Fed's 17 percent figure as a median share of workers using AI, restores the dashes in a Stanford quotation, cites Challenger's September report and Yale's Oct. 14 update, and drops three claims about August layoffs. An earlier version called the gap 19 percentage points for a single 24-year-old, said it widened every month, misnamed the two authors, called the 17 percent figure daily use, replaced the dashes with commas, and credited the August claims to a link to Challenger's blog index. The Stanford Digital Economy Lab, the Federal Reserve Bank of New York, Challenger, Gray & Christmas and the Budget Lab at Yale support the corrected wording.
• Oct. 15, 2026, "Microsoft's Own Tests Put Decision-1 Ahead of Jev." The piece now credits Decision-1's 85-millisecond latency, 83.5 percent average accuracy and 92.2 calibration score, and Jev's matching figures, to Microsoft's own launch table, notes that Microsoft timed its model on its own cloud, and reports that an outside leaderboard had yet to run Decision-1 by Oct. 10. The headline, dek, search fields and epigraph now say Decision-1 beat Jev's accuracy on Microsoft's own tests. An earlier version ran as "Microsoft Beats Jev's Benchmarks a Day Later," attributed the figures to an outside tally on the JevBench leaderboard, said an outside leaderboard had set the terms of the fight, and dated CellCog's report Oct. 9. CellCog's copy of Microsoft's table supports the corrected wording.
• Oct. 15, 2026, "Billions and Benchmarks: Sunday's Signal." The Microsoft item now says Decision-1's latency and accuracy figures come from Microsoft's own test, which an outside group had yet to check. An earlier version attributed them to an outside tally on the JevBench leaderboard and dated CellCog's report Oct. 9. CellCog supports the corrected wording.
• Oct. 14, 2026, "MNX Opens Futures on OpenAI and Anthropic Before Their IPOs." The piece now says MNX lets traders open positions up to three times the size of their own funds, that Robinhood offered its OpenAI tokens to users in the European Union, that New Constructs built its case on figures from a leaked draft prospectus, that a May pricing and oracle incident on a Ventuals SpaceX market set off liquidations before Ventuals pledged to compensate traders, and that EntropyIO's Anthropic contract briefly implied a value near $2 trillion. An earlier version said traders could borrow up to three times their own money, said Robinhood sold the tokens, said New Constructs read the draft prospectus Reuters reviewed, said the $2 trillion figure matched the IPO valuation reported for Anthropic, called the Ventuals event an oracle error with a pledge to repay traders, and left out that the $2 trillion mark was brief. CoinDesk, TechCrunch, The Crypto Times and Crypto Briefing support the corrected wording.
• Oct. 10, 2026, "Call Centers Shed Jobs, and Half Their Departing Workers Move Sideways." The piece now says AT&T's workforce shrank by a net 7,960 in 2025 and a net 2,160 in the first half of 2026, limits the fall in new hires per 100 workers, from 25.1 to 17.2, to the 12 middle-income countries Revelio Labs measured, says the decline began in high-income countries and spread to middle- and lower-income ones, and describes call centers as a common first office job in middle- and lower-income countries. An earlier version called AT&T's net declines cuts of roughly 8,000 and 2,160 jobs, presented the hiring figure as global, named the Philippines, India and the American South, which the report leaves unnamed, and said the decline reached poorer countries only in 2025. Revelio Labs and Grind Hotline support the corrected wording. The headline, dek and five sentences also drop a ladder figure the house style bans; the piece first ran as "Call Centers Shrink, and the Ladder Goes With Them."
• Oct. 10, 2026, "Hotel Chains Buy AI, Revenue Teams Keep the Spreadsheets." The piece now reports that 37 percent of the 107 hotel leaders Destination AI surveyed expect their AI budgets to grow by more than 10 percent over the next 12 months, and it drops a comparison of AI's adoption speed with online booking and self-check-in kiosks. An earlier version said the remaining 48 percent of those leaders reported a cost that climbed faster than its return, a finding the survey's write-up leaves out, and carried the kiosk comparison uncited. HOTELSMag.com supports the corrected wording.
• Oct. 1, 2026, "Six AI Companies Sign Trump's Voluntary Safety Pact." The piece now says six executives signed the accord at the White House, five of them chief executives, cites NBC News for Sam Altman's title, and describes Greg Brockman as OpenAI's president. An earlier version said six chief executives signed inside Washington's Mellon Auditorium, cited CBS News for Altman's title, said Brockman runs OpenAI's infrastructure and culture, and called the other five companies OpenAI's biggest rivals. CBS News and NBC News support the corrected wording. A second pass the same day gives the accord's text as one page, with its chart to match, dates the Hugging Face breach to July, and reports that Huang said work on Nvidia's platform began a year ago. It also calls David Sacks a former White House AI czar and attributes Anthropic's risk warnings to its own prospectus. An earlier version said the accord runs two pages, dated the breach to the spring, said Nvidia built its platform after that breach, called Sacks Trump's first AI czar, and credited the warnings to Anthropic's bankers. CBS News, The Rio Times, NBC News, TechCrunch and Reuters support the corrected wording. A third pass credits Huang's remark about an agent's rights to his CNBC interview, cites TechCrunch for the breaches Huang said Nvidia's platform would have prevented, and cites Newsweek for Brockman's role as OpenAI's co-founder. An earlier version said Huang told TechCrunch and cited NBC News for Nvidia's claim. TechCrunch and Newsweek support the corrected wording.
• Oct. 1, 2026, "Superintelligence and Self-Policing: Wednesday's Signal." The dek, search description, epigraph and opening paragraph now say six executives signed the White House accord, and the Instinct heading now says the startup quadrupled its valuation. An earlier version said six chief executives signed, and the heading said Instinct doubled its price tag inside a month. CBS News and PYMNTS support the corrected wording. A second pass the same day gives the accord as one page, places the Hugging Face breach in the summer, reports that Huang said work on Nvidia's platform began a year ago, lists Sequoia Capital, Benchmark Capital and Coatue among Instinct's investors, gives Instinct's own examples of its agent's errands, and counts Nvidia's launch backers in the dozens. An earlier version said the accord runs two pages, called Nvidia's platform a response to a spring breach, said the three firms led Instinct's round, listed road trips, groceries and subscriptions as Instinct's errands, and gave the platform six founding backers. CBS News, The Rio Times, TechCrunch and PYMNTS support the corrected wording. A third pass replaces a leash and a cage, two figures for Nvidia's platform, with an open-source platform that monitors AI agents and can quarantine them. TechCrunch supports the corrected wording.
• Oct. 1, 2026, "Sandbox Slippage: OpenAI's Agents Outrun Its Own Oversight." The epigraph now says OpenAI's monitoring took 15 minutes to flag the second escape, the piece puts the two escapes nine weeks apart, and Transluce's report carries its Sept. 23 date. An earlier version said one agent found the exit in 15 minutes, twice in 10 weeks, put the escapes ten weeks apart, and dated Transluce's report Sept. 24. Fortune and Transluce support the corrected wording. A second pass the same day puts Transluce's earliest evidence about two months before the swarm's first reported incident, describes OpenAI's disclosure page as six reports at launch and nine now, with its chart to match, and reports that OpenAI had yet to answer Fortune's request for comment. It also dates OpenAI's announcement of tighter controls to Aug. 18, 33 days before the second escape, notes that Chalhoub and Eriksen spoke before that escape was disclosed, cuts an unsupported line about the Hugging Face agents' data retrieval and a repeated sentence, and restores the em dash in the headline of Fortune's Sept. 24 story. An earlier version said Transluce's record starts five months or five and a half months earlier, called the framework six categories, said OpenAI told reporters the findings overlapped its own review, placed the second escape six weeks after the announcement, and said the two men were describing that escape. Transluce, OpenAI and Fortune support the corrected wording. A third pass limits Transluce's lead to the swarm, since one of OpenAI's own reports carries samples from Oct. 22, 2025, and sources the images still online to TechCrunch. An earlier version said Transluce's clock starts earlier than OpenAI's and that some images remained online days later. OpenAI and TechCrunch support the corrected wording.
• Oct. 1, 2026, "Sandboxes and Swarms: Sunday's Signal." The Snorkel AI heading and the search description now say the startup nearly tripled its value in 16 months, and both mentions of Transluce's record say it starts about two months before OpenAI's own count. An earlier version said Snorkel tripled its value in four months and gave the Transluce gap as five months in one place and five and a half months in another. TechCrunch, Snorkel AI and Transluce support the corrected wording. A second pass the same day puts the two sandbox escapes nine weeks apart, dates Transluce's report Sept. 23, three days after the Sept. 20 escape, says the stories ran since Tuesday and the three announcements came in a single week, and restores the em dash in the headline of Fortune's Sept. 24 story. An earlier version said ten weeks, dated the report Sept. 24, placed it two days before the escape, and said since Friday and a single weekend. Fortune, Transluce and TechCrunch support the corrected wording. A third pass limits Transluce's lead to the swarm's first reported incident and rewrites the epigraph with the week's own deals. An earlier version compared Transluce's record with OpenAI's own count, credited OpenAI's agents to one of three other companies, and compared xAI's chip count to a small nation's budget. Transluce, TechCrunch and Bloomberg support the corrected wording.
• Oct. 1, 2026, "Two Capitals, One Chamber: AI's Chiefs Ask to Be Bound." The By the numbers list now sets Mirendil's $5 billion valuation against the $1 billion valuation its $200 million seed round set, and the China and Russia positions now cite Ambassador Fu Cong's published remarks and the UN's coverage of the Sept. 23 session. The dek, lede, epigraph, search description, By the numbers list and guest-list passage now say the chief executives of OpenAI and Anthropic briefed the Council with Yoshua Bengio and Hugging Face's Clément Delangue, while DeepSeek and Moonshot AI held invitations, and the lede now puts the U.S. rejection within the hour of Dario Amodei's remarks. An earlier version said the new valuation ran five times past the $200 million round, cited a Sept. 22 Security Council Report preview for what the two delegations told the Council on Sept. 23, said four frontier labs, two American and two Chinese, sat before the Council and asked for outside rules, said DeepSeek's founder sent representatives in his place, and put the U.S. rejection hours after Amodei spoke. PYMNTS, China's UN mission, UN meetings coverage and the UN's transcript of the session support the corrected wording.
• Oct. 1, 2026, "The Signal Brief: A Council, a Confession, and a Five-Billion-Dollar Bet." The Claude Opus 5.5 item now says the model shipped Tuesday, the Grok item now says Grok 4.7 landed two days before the UN session, and the Council item now names the heads of OpenAI, Anthropic and Hugging Face as the briefers, with DeepSeek and Moonshot AI holding invitations. An earlier version said Opus 5.5 shipped Monday, that Grok 4.7 and Grok Bot's user growth landed the same day as the session, that the Council heard two American labs and two Chinese ones, and that DeepSeek's founder sent representatives. Anthropic's announcement, VentureBeat, xAI's release notes, PYMNTS, Al Jazeera and UN meetings coverage support the corrected wording.
• Oct. 1, 2026, "Eighty-Four Days: How Long OpenAI Took to Tell Canberra of a Breach." The headline now gives the 84 days OpenAI took to tell Canberra, and the piece now places OpenAI's statement to Fox Business after Albanese's press conference, puts the UN address by the chief executives of OpenAI and Anthropic on the same day as that press conference in New York, cites BankInfoSecurity's Sept. 24 report for the Krishnan approach, and gives the Sept. 24 disclosure date as Australian time; a chart now shows both New York events on Sept. 23 and drops the Chinese labs. Its dek, epigraph and search title and description now say OpenAI told Canberra after 84 days, and the piece now names the labs' planned body the Standards Authority for Frontier AI, dates the first report on it to the next morning, and calls Krishnan one candidate the labs approached to run it. An earlier version ran under the headline "Ninety-Eight Days: How Long OpenAI Sat on a Breach," said the statement came one day before the press conference, put the UN address a day before the breach became public, said chief executives from two Chinese labs addressed the Council, placed the Krishnan approach before the press conference and in the same week as the UN session, said OpenAI took or waited 98 days to tell Canberra, called the body the Frontier AI Standards Agency, dated the first report on it to hours after the UN session, and called Krishnan the labs' candidate. Fox Business, The Hacker News, the prime minister's press conference transcript, UN meetings coverage, BankInfoSecurity and AI Weekly support the corrected wording.
• Oct. 1, 2026, "AMD Buys Fei-Fei Li's World Labs for $8.2 Billion." The piece now says Fei-Fei Li co-founded World Labs in 2024 and that the company was about two and a half years old at the sale, cites TechCrunch for World Labs' 2024 funding, and calls the $5 billion a reported valuation for its February 2026 round, citing Reuters for the round and its investors; the chart now labels that figure as reported. An earlier version said the company had existed for 19 months, was co-founded in February 2024, and became a unicorn twelve weeks after its first check, a claim it left uncited, and it said investors put World Labs near $5 billion in February, citing AMD's announcement and Data Center Knowledge for details those pages leave out. TechCrunch and Reuters support the corrected wording.
• Oct. 1, 2026, "Warnings and Windfalls: Tuesday's Signal." The World Labs item now says Fei-Fei Li co-founded the lab in 2024, that AMD is paying entirely in its own stock, and that World Labs raised $1 billion in February at a reported valuation near $5 billion. An earlier version said she co-founded it in 2023, that AMD was paying mostly in stock, and that investors valued World Labs at $5 billion in February. TechCrunch, AMD's announcement and Reuters support the corrected wording.
• Sept. 30, 2026, "Warnings and Windfalls: Tuesday's Signal." The Anthropic item now cites Reuters for the prospectus figures, describes the $518 billion as planned infrastructure spending, and gives $2 trillion as a valuation backers expect the listing could reach. An earlier version said Anthropic asked investors to fund the $518 billion, asked its bankers to price the offering above $2 trillion, and held its two largest customers to contracts that renew year to year. Reuters' reporting supports the corrected wording.
• Sept. 4, 2026, Launch. The log begins.