OpenAI, Google, Microsoft and Anthropic will obey one shared robots.txt signal that lets websites refuse AI training

The stories

37 stories on this beat today.

Every story told the same way: what happened, who is involved, why it matters, how it works, the truth and the good news.

1. 01

   OpenAI, Google, Microsoft and Anthropic will obey one shared robots.txt signal that lets websites refuse AI training

   4

   of the biggest AI companies promise to obey one shared line

   OpenAI, Google, Microsoft and Anthropic; Meta backed it without promising [1]

   1 Feb 2027

   the day Microsoft starts obeying the signal

   Google follows in early 2027, and Anthropic already obeys it [2]

   What happened

   • OpenAI, Google, Microsoft and Anthropic say they will obey one shared signal against AI training. [1]
   • The signal is one line in robots.txt, a website's file of rules for automatic visitors. [1]
   • Those visitors are crawlers: programs that copy pages for search, archives or AI training. [1]
   • With one line, a site can now refuse all 4 companies at once. [2]
   • A site adds a section for a new name, __usage_ai-training, and lists what is off limits. [2]
   • It can block training site-wide but still allow chosen pages, such as a press kit. [2]
   • The line does not take a site out of Google's search results. [2]
   • Google reads it as a stop for Google-Extended, the crawler that gathers Gemini's training material. [2]
   • Anthropic already obeys the signal. [1]
   • Microsoft will start on 1 February 2027, and Google in early 2027. [2]
   • OpenAI says it will follow in the coming months. [1]
   • Meta, owner of Facebook and Instagram, backed the idea but did not promise to obey. [1][3]
   • Andy Stone, Meta's head of communications, called a common signal an important step. [1]
   • Microsoft will also let sites refuse grounding, where an AI checks a trusted source before answering. [1]
   • Sites keep their place in Bing, Microsoft's search engine, whatever they choose. [1]

   Who is involved

   Kevin Kelley, a standards lead at Anthropic, wrote the proposal. [1]

   It is before the IETF, the open body that sets the internet's technical standards. [2]

   • OpenAI AI company
   • Google AI and search company
   • Microsoft AI and software company
   • Anthropic AI company
   • Kevin Kelley a standards lead at Anthropic; he wrote the proposal
   • IETF sets internet standards
   • Meta the owner of Facebook and Instagram; it backed the idea but has not promised to obey
   • OpenAI, Google, Microsoft and Anthropic four of the biggest AI companies; they promised to obey the signal
   • The IETF the open body that sets the internet's technical standards; it is weighing the proposal

   How it unfolded

   Until now, a site had to find each AI company's crawler and write a rule. [2]

   Publishers face a steady stream of AI crawlers copying their sites into AI models. [1]

   Cloudflare, a company that carries traffic for websites, already blocks AI crawlers by default. [1]

   Another standard, called RSL, lets sites charge the crawlers that visit them. [1]

   1. 1 Feb 2027 Microsoft starts
   2. Early 2027 Google starts
   3. Until now sites write one rule for each AI crawler [2]
   4. This week four AI companies back one shared signal [1]
   5. Now Anthropic already obeys it [1]
   6. 1 Feb 2027 Microsoft starts obeying it [2]

   Why it matters

   For a blog or news site, one line now covers four of the biggest AI makers. [2]

   But it cannot undo past copying: pages already taken stay with the AI companies. [2]

   How it works

   Internet engineers built robots.txt as a request, honoured only by crawlers that choose to. [2]

   So the signal reaches only as far as the companies that promise to read it. [1]

   The truth

   • What is really going on At the same time, the websites AI learns from are getting new ways to refuse, from one shared robots.txt line to a UK investigation of Meta's terms. [1][3]
   • Anthropic, whose staff wrote the rule, is the only one of the four obeying it. [1]
   • A site that adds nothing stays open to AI training by all of them. [2]
   • Nobody knows yet whether Meta and smaller AI firms will obey it. [1]
   • Who gains Anthropic — Its own standards lead wrote the shared signal, and it is the only one of the four firms already obeying it. [1]
   • Who pays Websites that do nothing — A site that never adds the new line stays open to AI training, and pages already copied stay copied. [2]
   • Nobody knows yet

     Whether Meta and smaller AI companies will obey the new robots.txt signal.

     Meta backed the idea but did not say it would honour it, and the signal only works for crawlers that choose to. [1][2]

   What comes next

   The next fixed date is Microsoft's start, on 1 February 2027. [2]

1. 02

   Amazon borrows against its AI machines

   What happened

   • Amazon set up a loan of up to $80 billion backed by AI hardware it buys. [4]
   • Lenders get rights over some of that equipment, though Amazon's filing does not say which. [4]
   • Amazon can draw the money from the start of 2027 to the third quarter of 2028. [4]
   • In the year to September, Amazon spent $42.4 billion more cash than it brought in. [4]
   • A year earlier, it had $14.8 billion left over. [5]
   • Its spending promises reached $883 billion, more than double a year before. [4]

   Who is involved

   Amazon's boss Andy Jassy says AWS installs a new AI chip about every 14 seconds. [14]

   How it unfolded

   Until now Amazon paid with cash from AWS, its cloud arm, and record bond sales. [4]

   1. Start of 2027 the loan opens
   2. Third quarter of 2028 the last draw

   Why it matters

   Amazon borrows instead of selling shares; Alphabet is selling shares to raise $80 billion. [4]

   How it works

   Bankers call this secured lending: lenders hold a claim on the machines, not just a promise. [4]

   Why lenders to AI builders want a claim on the machines

   Bankers call it secured lending: a lender that can take the chips or the building if payments stop risks less, so it will lend more.

   The twist

   A lender to a weak borrower is really lending against the machines, so the loan is only as safe as the price those machines would fetch.

   The picture

   Microsoftits own name

   vsvs

   NscaleNvidia's backing

   Step by step

   1. 1 A firm spends more cash on AI than its business brings in
   2. 2 It borrows to cover the gap
   3. 3 Lenders ask what they get back if the payments stop
   4. 4 A claim on the machines, or a stronger firm's guarantee, answers that
   5. 5 So the weaker the borrower, the more the loan rests on the machines

   The same force, elsewhere today

   Where this chain is also running, in today's other stories.

   • Amazon borrows against its AI machines

     Amazon's lenders get rights over some of the equipment the $80 billion buys, so they can fall back on the machines and not only on Amazon's promise.

   • Nvidia may back Nscale's loans

     Nscale is young and short of chips, so Nvidia offers to cover the building's value and take it over if the payments stop.

   • Lenders sort AI borrowers by strength

     Microsoft, rated AAA, borrows $25 billion on its own name, while Anthropic, expected to be rated junk, leans on stronger partners to back its debt.

   Where you've seen this

   Home loans

   the bank lends against the house and can sell it if the owner stops paying

   Car finance

   the lender keeps a claim on the car until the last payment is made

   Pawn shops

   a watch or a ring is held until the loan is paid back

   The catch

   A claim on machines protects a lender only if someone will buy them, and if many AI builders fail at once they all try to sell the same chips together.

   And the whole of it

   The lenders, the chip makers, the cloud firms and the people who use AI each see one part of these deals. If AI demand slows, the lenders would be left holding machines fewer people want, and the chip makers and cloud firms would feel it in the same months.

   The truth

   • What is really going on Alphabet, Amazon and Meta are spending more cash than their businesses bring in, so Amazon is borrowing against its AI machines and Nvidia may stand behind Nscale's loans. [5][4][6]
   • Amazon's debt had passed $225 billion by June, before this loan. [4]
   • Nobody knows yet

     What happens to Amazon's lenders if AI demand cools.

     Amazon has not said which machines back the $80 billion, so nobody outside can judge what the lenders could recover. [4]

1. 03

   Big Tech's AI promises pass $3 trillion

   What happened

   • Five tech giants have signed up for more than $3 trillion of future AI spending. [15]
   • Bloomberg counted Alphabet, Meta, Microsoft, Amazon and Oracle. [15]
   • The money is for data centres, leases, equipment and energy. [15]
   • Alphabet, Google's owner, has the most: more than $1 trillion, nearly 8 times a year ago. [15]
   • Meta, Facebook's owner, has more than $800 billion, a bit under half for data-centre leases. [15]
   • Some of the leases run for 30 years. [15]
   • Alphabet, Amazon and Meta now spend more cash than their businesses bring in. [16]
   • Microsoft is the only one of the four still bringing in more than it spends. [16]

   Who is involved

   • Alphabet cash out beats in
   • Amazon cash out beats in
   • Meta cash out beats in

   How it unfolded

   In July a similar count of four of them came to about $2.4 trillion. [15]

   Why it matters

   The giants are paying for AI with their older businesses: adverts, business software and online shops. [16]

   How it works

   Accountants call cash left after spending free cash flow; below zero, firms borrow or sell shares. [5]

   The truth

   • Why it works on people A figure like $3 trillion is too big to picture, so it can sound like strength even when much of it is promised or borrowed. [15][5]
   • Gartner, a research firm, has warned that much of this spending may never pay off. [15]
   • Each company counts these promises differently, so the total mixes short orders with decades-long leases. [15]
2. 04

   Nvidia may back Nscale's loans

   What happened

   • Nvidia, which makes the chips AI runs on, may guarantee billions of dollars of Nscale's loans. [6]
   • Nscale, a British AI cloud company, is building a West Virginia data centre for Anthropic. [17]
   • Anthropic signed a $44.6 billion contract for the first building alone. [6]
   • If Nscale missed payments, Nvidia would take over the site and help find a buyer. [6]
   • The talks are not signed, and the terms could still change. [17]

   Who is involved

   • Nvidia AI chip maker
   • Nscale AI cloud company
   • Anthropic the tenant

   How it unfolded

   Nscale had about 25,000 AI chips running in August, against nearly 500,000 planned. [6]

   This month Microsoft pulled out of a planned Nscale site in England. [6]

   1. A set deadline Nscale's money must be ready
   2. If not Anthropic can walk away

   Why it matters

   Anthropic can walk away if Nscale's money is not in place by a set deadline. [17]

   How it works

   Bankers call it a residual value guarantee: Nvidia promises the building's worth to lenders. [17]

   Nscale is young and short of chips, so Nvidia offers to cover the building's value and take it over if the payments stop.

   The idea, in story 02: Why lenders to AI builders want a claim on the machines →

   The truth

   • Nvidia keeps a big chip buyer, and would have to step in if Nscale fails. [6]
   • Nvidia has given AI cloud firms and big customers about $170 billion in investment and credit. [6]
   • Who gains Nvidia — Backing Nscale's loans would keep a project on track that is due to fill a building with its chips. [6]
   • Nobody knows yet

     Whether Nvidia will actually sign the guarantee for Nscale.

     The talks are not signed and the terms could still change; Nvidia and other backers asked for stricter terms after markets turned. [17]

3. 05

   Lenders sort AI borrowers by strength

   What happened

   • Microsoft borrowed $25 billion from investors on Thursday, selling its first bonds since 2017. [18]
   • Its 10-year bonds cost 5.6% a year, barely more than the US government pays. [18]
   • Microsoft holds a AAA credit rating, the top grade; only Johnson & Johnson shares it. [18]
   • SpaceX, rated BBB, 3 steps lower, will pay up to 8.75% on $40 billion. [18]
   • Bonds of CoreWeave, an AI cloud firm deep in the junk grade, have traded above 15%. [18]
   • Blackstone, a big investor, now expects Anthropic to get a junk rating after its share sale. [19]
   • Junk is the grade for riskier borrowers; Blackstone expects the top band of it, BB. [19]

   Who is involved

   • Blackstone big investor
   • Anthropic junk rating expected

   How it unfolded

   Blackstone had earlier said Anthropic would win a safer grade before 2027. [19]

   Why it matters

   A junk rating would make it harder and dearer for Anthropic to borrow for chips. [19]

   How it works

   Credit analysts grade how likely a borrower is to repay, and lower grades pay more interest. [18]

   Microsoft, rated AAA, borrows $25 billion on its own name, while Anthropic, expected to be rated junk, leans on stronger partners to back its debt.

   The idea, in story 02: Why lenders to AI builders want a claim on the machines →

   The truth

   • Microsoft gets cheap money for AI; weaker firms like CoreWeave pay far more. [18]
   • Who gains Microsoft — Its AAA rating let it sell 10-year bonds at 5.6%, barely above what the US government pays. [18]
   • Who pays Weaker AI borrowers — CoreWeave's bonds trade above 15%, and a junk rating would make Anthropic's borrowing dearer. [18][19]
4. 06

   Trump calls imported robots a security threat

   What happened

   • President Trump declared imported robots and factory machines a threat to US national security. [8]
   • He signed it on Wednesday, after a Commerce Department study begun in September 2025. [8]
   • It sets no tariffs, which are taxes on imports, for now. [8]
   • If talks with trading partners fail, the options include tariffs, quotas and fees. [8]
   • The US installed almost 38,500 industrial robots in 2025, up 12%. [8]
   • US firms often buy heavy robot arms from FANUC, ABB, Yaskawa and KUKA, all based abroad. [8]
   • No US company makes comparable heavy-duty arms, Ars Technica reported. [8]

   Who is involved

   • FANUC robot arm maker
   • ABB robot arm maker
   • Yaskawa robot arm maker
   • KUKA robot arm maker
   • General Motors, Ford, Tesla carmakers
   • Boeing, Lockheed Martin plane makers

   How it unfolded

   1. Sept 2025 Commerce study begins
   2. Wednesday Trump signs the order

   Why it matters

   General Motors, Ford, Tesla, Boeing and Lockheed Martin all rely on such robots. [8]

   How it works

   Trade lawyers call this Section 232: a 1962 law letting presidents limit imports for security. [8]

   The truth

   • The order names no country, and none of the foreign firms it warns about. [8]
   • Nobody knows yet

     Whether the robot tariffs come at all.

     Officials have 180 days of talks and the order names no country. [8]

   What comes next

   Officials must report back to President Trump within 180 days. [8]

5. 07

   Britain investigates Meta's terms for publishers

   What happened

   • The CMA, Britain's competition watchdog, is investigating Meta's new terms for publishers. [3]
   • The terms could let Meta use news sites' content for its own products, including AI. [3]
   • The CMA will judge whether Meta is abusing a dominant position with exploitative terms. [3]
   • Meta had planned the terms for 20 November, then put them on hold. [3]
   • A Norwegian court this week banned Meta from enforcing the terms for now. [3]
   • In France, an appeals court told Meta to negotiate fairly over payments to the press. [3]
   • Meta says it has promised not to change its terms without consulting publishers. [3]

   Who is involved

   • The CMA Britain's competition watchdog
   • Meta owner of Facebook

   How it unfolded

   Since May, over 340 news sites have blocked archive crawlers to keep their work from AI. [3]

   1. Planned terms due 20 November
   2. Now on hold

   Why it matters

   Thousands of UK businesses rely on Meta for customers, so unfair terms could break the law. [3]

   How it works

   Competition law calls it abuse of dominance when a firm's size lets it impose unfair terms. [3]

   The truth

   • Publishers gain a regulator on their side; Meta may be made to change its terms. [3]
   • Who gains News publishers in Britain — The competition watchdog is now examining whether Meta's terms for their content are unfair. [3]
6. 08

   Wikipedia's operator takes a middle line

   What happened

   • The New York Times says OpenAI and Microsoft broke copyright by training AI on its articles. [20]
   • The Wikimedia Foundation, which runs Wikipedia, has filed a brief in that case. [21]
   • OpenAI and Microsoft say that training is fair use, which US law allows without permission. [20]
   • Wikimedia says training should count as fair use, but chatbots copying articles back should not. [21]
   • It warns that a sweeping win for either side could harm the open web. [20]
   • Wikimedia sites are among the most important sources of AI training data, the filing says. [21]

   Who is involved

   • The New York Times the newspaper suing
   • OpenAI and Microsoft the AI firms sued
   • Wikimedia Foundation runs Wikipedia

   Why it matters

   But AI answers draw visitors away from the volunteers who keep the sites going. [21]

   The case will decide whether training AI on news articles breaks US copyright law. [20]

   How it works

   US fair use asks whether a new use makes something new and different from the original. [21]

   The truth

   • Wikimedia wants both: AI tools for its editors, and readers who still come to Wikipedia. [21]
7. 09

   Australia's AI copyright plan meets its constitution

   What happened

   • Australia's government is weighing copyright changes that could let AI firms train on creative work. [7]
   • Opposition senators warn that this could breach section 51 of the constitution. [7]
   • Section 51 lets Australia's government take property only on just terms, meaning fair payment. [7]
   • Senator Ross Cadell said Australia could owe payment to every rights holder on Earth. [7]
   • Anne Twomey, a leading constitutional lawyer, said there was potentially a concern. [7]
   • Anthropic suggested an opt-out model: it would copy work unless an owner signals an objection. [7]

   Who is involved

   • Opposition senators warn of a breach
   • Anne Twomey constitutional law professor
   • Musicians
   • Authors
   • Journalists

   How it unfolded

   Prime Minister Anthony Albanese has said it would be theft if creators were not paid. [7]

   Why it matters

   The choice decides whether musicians, authors and journalists get paid when AI learns from them. [7]

   How it works

   Constitutional lawyers ask whether weakening an owner's rights counts as the state taking property. [7]

   The truth

   • No bill has been published yet, so nobody can test the worry against real wording. [7]
   • Nobody knows yet

     Whether Australia can change copyright for AI without breaking its constitution.

     No bill has been published, and the leading constitutional expert says it depends on how the law is drafted. [7]

8. 10

   Apple's AI reaches China with Alibaba's help

   What happened

   • Apple's AI features reach China with iOS 27.3 in December, after a 4 November test. [9]
   • Users can call on Qwen, an AI model from Alibaba, inside Siri and Writing Tools. [9]
   • China's internet regulator cleared that set-up in July. [9]
   • The features include live translation, writing help and removing objects from photos. [22]
   • The rebuilt Siri assistant Apple showed this year is not included. [9]
   • They work on the iPhone 15 Pro and newer models. [9]

   Who is involved

   • Apple the iPhone maker
   • Alibaba maker of the Qwen model

   How it unfolded

   They arrive about 2 years after Apple's AI first launched in the US. [9]

   Apple trained its own AI model just for China, Reuters reported in August. [22]

   1. 4 November developer test version
   2. December iOS 27.3 for everyone
   3. July regulator clears the set-up
   4. December the features launch

   Why it matters

   Until December, iPhones in mainland China have none of these Apple AI features. [22]

   The truth

   • Alibaba gains a place inside Siri on iPhones across China. [9]
9. 11

   Waymo picks Toyota for Japan and the UK

   What happened

   • Waymo, a driverless-taxi company, chose Toyota to supply cars for Japan and the UK. [23]
   • Testing of self-driving Toyota bZ4X cars starts in Japan in 2027. [24]
   • Toyota will build them at its Takaoka factory and fit Waymo's driving system there. [23]
   • Japan and the UK drive on the left, so the cars need right-hand steering. [23]
   • Waymo has tested Jaguar cars in London for months and needs permits to launch. [23]
   • It plans a fully driverless taxi service in Tokyo next year. [24]

   Who is involved

   • Waymo driverless taxis
   • Toyota carmaker

   How it unfolded

   The two firms signed a first, loosely worded agreement 18 months ago. [23]

   1. 2027 testing in Japan
   2. Next year Tokyo driverless service

   Why it matters

   Toyota's factories give Waymo a way to make robotaxis for markets outside the US. [24]

   How it works

   Engineers call this Level 4 driving: the car needs no human driver in its cleared areas. [24]

   The truth

   • For now the deal looks more like an ordinary supplier contract, TechCrunch says. [23]

   What comes next

   The firms will also study putting Waymo's system into Toyotas that people own. [23]

10. 12

    Foxconn aims for half of AI servers

    What happened

    • Foxconn, best known for assembling iPhones, wants 50% of the world's AI server market. [25]
    • Its share is about 40% now and could reach 45% next year, chairman Young Liu said. [25]
    • An AI server is the computer AI runs on in a data centre. [25]
    • Foxconn's factories are fully booked for at least the next 18 months. [25]
    • Liu expects AI equipment to bring 60% to 65% of revenue within three to five years. [25]
    • He said a new deal with OpenAI is in the works, after meeting Sam Altman. [26]
    • OpenAI, the maker of ChatGPT, has not commented. [26]

    Who is involved

    • Young Liu Foxconn's chairman
    • Sam Altman OpenAI's boss

    How it unfolded

    In November 2025 the two firms agreed to design AI data-centre racks together. [26]

    1. Nov 2025 racks designed together
    2. Now a new deal in the works

    Why it matters

    A full order book at the main AI server maker shows buyers still ordering hard. [25]

    The truth

    • Liu pushed back on talk of an AI bubble, while his own sales ride on AI. [26]
    • Who gains Foxconn — Its AI server factories are fully booked for at least a year and a half. [25]
11. 13

    China denies spying on Nvidia's boss

    What happened

    • China called a report of a spy close to Nvidia's boss, Jensen Huang, groundless speculation. [27]
    • The New York Times reported that US intelligence suspected Alice Huang of gathering information for China. [27]
    • Alice Huang, no relation to Jensen, was an executive at Megaspeed, a Singapore data-centre firm. [27]
    • Investigators also suspected Megaspeed of helping Nvidia chips reach China against US export rules. [27]
    • Nvidia said Jensen's interactions with Ms Huang are not unusual or unexpected. [27]
    • The Times reported that President Trump was briefed a year ago and took no action. [28]
    • Sources said he feared hurting Nvidia's chip sales or the trade truce with China. [28]

    Who is involved

    • Alice Huang former Megaspeed executive
    • Jensen Huang Nvidia's boss

    How it unfolded

    1. A year ago Trump is briefed
    2. Since no action

    Why it matters

    Jensen Huang has become President Trump's main adviser on AI, the report says. [28]

    How it works

    Intelligence agencies call it unmasking: revealing a US citizen's name in surveillance reports. [27]

    The truth

    • Jensen Huang was not given the warning briefing that targets of spying sometimes get. [28]
    • Nobody knows yet

      What, if anything, Alice Huang passed to China.

      Neither the BBC nor the sources behind the New York Times report could say what she shared or how valuable it was. [27][28]

12. 14

    Reddit shares fall 29%

    What happened

    • Reddit's shares fell 29.22% on Friday, their worst day since listing in 2024. [29]
    • Reddit forecast sales of $1 billion to $1.03 billion for the last quarter of the year. [29]
    • Analysts had expected $1.05 billion. [29]
    • Sales still grew 53% from a year earlier. [29]
    • Its boss, Steve Huffman, said visits from Google search went down meaningfully because of AI. [29]
    • Reddit will test making people log in to read it. [29]

    How it unfolded

    Reddit's worst fall before this, 20.99%, came on 31 July. [29]

    Why it matters

    Google's AI answers keep readers on Google instead of sending them to sites like Reddit. [29]

    How it works

    Web businesses call visits sent by a search engine referral traffic, and Reddit's fell. [29]

    The truth

    • Google keeps the readers; Reddit loses the visits it used to get from search. [29]
13. 15

    Europe's data-centre builders hurt and unpaid

    What happened

    • A Guardian and Investigate Europe inquiry found injuries and unpaid wages on European data-centre sites. [30]
    • Reporters spent five months on sites in Sweden, Norway, Belgium and the Netherlands. [30]
    • At Google's site in Saint-Ghislain, Belgium, a worker testing wiring was hit by 400 volts. [30]
    • He says managers told him the lines were switched off, but they were live. [30]
    • The blast burned 12% of his body, hospital records show. [30]
    • Belgium's labour inspectors closed their inquiry in late October and found no wrongdoing. [30]
    • At Microsoft's site in Gavle, Sweden, workers described 14 men sharing one bathroom. [30]
    • One worker there tested positive for active tuberculosis in February, medical records show. [30]

    How it unfolded

    Europe has more than 2,300 data centres, with hundreds more being built. [30]

    Why it matters

    Unions say Europe's push to own its tech leans on exploited migrant workers. [30]

    The truth

    • No company or contractor appears to have been found liable for the Belgian blast. [30]
    • Who pays Migrant workers building Europe's data centres — Reporters found injuries, crowded housing and unpaid wages on sites in four countries. [30]
14. 16

    Finnish police investigate Google's forest clearing

    What happened

    • Finnish police opened a criminal inquiry into forest cleared for Google data centres. [31]
    • The firm, Google's Finnish arm Tuike Finland, faces suspected breaches of environmental law. [31]
    • Five complaints were made about its Muhos data-centre project. [31]
    • Near Muhos, 816 acres of forest were cleared without permission and turned to mud. [31]
    • It also cleared 500 acres near Otanmaki. [31]
    • Finland's licensing and supervision authority backs the police inquiry. [31]
    • Google has apologised for the clearing and says it will cooperate with police. [31]

    Who is involved

    • Tuike Finland Google's Finnish arm

    Why it matters

    The case tests whether the rush to build AI data centres can outrun environmental rules. [31]

    The truth

    • Local forest is already gone; the size of any penalty is not yet known. [31]
    • Who pays Forest near Muhos, Finland — 816 acres were cleared without permission for a Google data-centre project. [31]
15. 17

    Salesforce buys a military AI start-up

    What happened

    • Salesforce, which sells software for managing customers, agreed to buy Rune Technologies for $375 million. [32]
    • Rune makes AI software that helps the US military predict supply needs in the field. [32]
    • The price is in cash and shares, Resilience Media reported. [32]
    • Rune was valued at $127 million last year, after raising $30 million in total. [32]
    • Its backers include IQT, a venture fund backed by the CIA. [32]
    • Salesforce's defence arm, Missionforce, grew 80% in a year. [32]
    • It holds a 10-year, $5.6 billion contract with the US Army. [32]

    How it unfolded

    This year Salesforce also agreed to buy Fin, an AI agent company, for $3.6 billion. [32]

    Why it matters

    Rune's field data moves Salesforce from office planning toward battlefield operations. [32]

    The truth

    • Rune's early backers, such as Andreessen Horowitz, sell well above the last valuation. [32]
16. 18

    GitHub lets teams run many AI agents

    What happened

    • GitHub, where programmers store and share code, redesigned its Copilot app at its Universe conference. [11]
    • Copilot is GitHub's AI helper that writes and reviews code. [11]
    • A new Projects feature lets teams give several AI agents goals across many code stores. [11]
    • Agents can draw on context from tools, code stores and design mock-ups. [11]
    • Developers can also work on plans, design talks and app data in the same place. [11]

    How it unfolded

    The earlier app mainly helped one developer run agent sessions and review code changes. [11]

    Why it matters

    Companies want agents across coding, testing and upkeep to ship software faster. [11]

    The truth

    • GitHub's announcement does not show whether teams of agents write better software. [11]
17. 19

    Cisco hands one engineer its old code

    What happened

    • Cisco, which makes network equipment, gave a 27-year-old engineer unlimited AI use to rewrite code. [33]
    • Padmanabhan Krishnamurthy must rewrite 98% of Cisco's legacy code, its older software, within 24 months. [33]
    • He has 4 weeks to prove the idea first. [33]
    • Cisco's president Jeetu Patel said he gave him unlimited tokens to refactor code. [33]
    • Krishnamurthy joined when Cisco bought EzDubs, a translation start-up he helped build. [33]
    • He told Patel he could do the job with a team of two. [33]

    How it unfolded

    Patel found him in a role five levels down, doing work Patel called unstrategic. [33]

    Why it matters

    Patel wants young hires who challenge the system and the status quo. [33]

    How it works

    Programmers call rewriting old code without changing what it does refactoring. [33]

    The truth

    • Patel says today's company structures work against people like Krishnamurthy. [33]
18. 20

    Micron pays big bonuses in Taiwan

    What happened

    • Micron, a US maker of memory chips, is paying big bonuses to its Taiwan workers. [34]
    • About 700 of its 4,300 Taiwan production staff, about 17%, get five years' pay in bonuses. [34]
    • The payout includes a one-time bonus of about $31,000. [34]
    • From January 2027, Micron workers worldwide get a new quarterly profit bonus. [34]
    • Samsung and SK hynix, its two Korean rivals, have paid bonuses of up to $400,000. [34]

    How it unfolded

    Micron has been battling worker unrest and unions pushing for more pay. [34]

    Why it matters

    AI data centres buy huge amounts of memory, so memory makers earn record profits. [34]

    The truth

    • Production line workers are demanding a slice of those profits. [34]
19. 21

    Korea charges Tesla owners who unlocked self-driving

    What happened

    • South Korean police identified 80 Tesla drivers who allegedly unlocked blocked self-driving software. [35]
    • The feature, Full Self-Driving, is not approved for their cars in Korea. [35]
    • Police seized 21 devices that plug into the car's internal network to switch it on. [35]
    • One suspect built his own device from instructions found on GitHub. [35]
    • The penalty can reach two years in prison or a 20 million won fine, about $15,000. [35]
    • Only US-built Teslas can use the feature, because they skip Korea's local safety tests. [35]
    • About 47,458 of 229,311 Korean Teslas, or 20.7%, are eligible, one data firm estimates. [35]

    Why it matters

    Korea's transport ministry said unlocked cars may fail to read road conditions properly. [35]

    How it works

    Korea's car safety law bans unapproved software changes that could affect a car's safety. [35]

    The truth

    • Many owners had already paid for the feature they now face fines for using. [35]
    • Who pays Korean Tesla owners who unlocked self-driving — They face fines of up to 20 million won, about $15,000, or prison. [35]
20. 22

    Proton's boss drops AI re-interviews

    What happened

    • Andy Yen, head of Proton, the privacy-focused email company, cancelled planned AI re-interviews on Friday. [36]
    • Staff would have had to explain how they use AI, with their jobs at risk. [36]
    • The move followed a 404 Media investigation published the day before. [36]
    • It showed Yen pushing AI even when staff raised security worries. [36]
    • He called one worker's request to test new code for days loser thinking. [36]
    • Some Proton users have called for Yen to resign. [36]

    Who is involved

    • Andy Yen Proton's boss

    How it unfolded

    Proton's chief technology officer had written an AI manifesto for staff. [36]

    Why it matters

    Proton sells privacy and security, so rushed code worries its paying users. [36]

    How it works

    Security teams call it a pentest: attacking their own new code to find holes before release. [36]

    The truth

    • Cancelling the re-interviews does not answer users' worries about security, 404 Media reports. [36]
21. 23

    AI agents order little from DoorDash

    What happened

    • DoorDash, a food delivery app, says outside AI agents have placed only several hundred orders. [13]
    • It completes about 3 million deliveries a day, its 2025 report said. [13]
    • The agents include Muse, Instinct and Fo. [13]
    • Chief executive Tony Xu said most people still order through DoorDash's own app. [13]
    • Xu said a company should not fight gravity, but should not be delusional either. [13]
    • DoorDash runs its own agents too, including one that takes orders by text. [13]

    Who is involved

    • Tony Xu DoorDash's chief executive

    Why it matters

    There is much talk of AI agents shopping for people; DoorDash's numbers show it barely started. [13]

    The truth

    • DoorDash gains from outside agents only if they order through it, and few do yet. [13]

    What comes next

    Xu sees a chance to sell adverts inside AI agents in future. [13]

22. 24

    Accenture lets AI agents screen applicants

    What happened

    • Accenture, a global consulting and outsourcing firm, uses AI agents to screen job applicants' CVs. [12]
    • The agents also book interviews, its head of people, Kate Clifford, told investors. [12]
    • Accenture says a human always makes the hiring decision. [12]
    • It claims the process cut the time taken to hire by 28%. [12]
    • Accenture now sells the same system to its clients. [12]

    Who is involved

    • Kate Clifford Accenture's head of people

    How it unfolded

    Clifford said recruiters had spent their time typing data and scheduling, not meeting people. [12]

    Why it matters

    At Accenture, an applicant's CV is now read first by AI agents. [12]

    The truth

    • The 28% figure is Accenture's own claim, made at its investor day. [12]
23. 25

    Insurers write AI out of their policies

    What happened

    • Big insurers such as AIG are asking US regulators to let them exclude AI from policies. [37]
    • 80% of state insurance commissioners have already agreed, ZDNet reports. [37]
    • So losses caused by AI may fall outside a company's normal insurance. [37]
    • Start-ups are moving the other way and selling AI cover. [37]
    • Klaimee, a start-up, raised $5.5 million to check AI agents and insure them. [37]
    • HSB, part of the insurer Munich Re, has sold AI liability cover since March. [37]

    Who is involved

    • Insurers like AIG want AI excluded
    • US regulators asked to allow it

    Why it matters

    AI agents now reach company inboxes, passwords, customer records and payment systems. [37]

    How it works

    Underwriters, who price insurance, rely on past losses, and AI changes too fast for that data. [37]

    The truth

    • Most buyers assume AI cover pays for every kind of loss; it almost never does. [37]
24. 26

    Applebee's app held a database password

    What happened

    • Applebee's, a US restaurant chain, left a database password inside its Android app. [38]
    • With it, anyone could reach registered customers' names, emails, birthdays and phone numbers. [38]
    • Researcher Robert Wiggins found it, using AI to search thousands of lines of code. [38]
    • The chain had no way for outsiders to report security problems. [38]
    • Applebee's fixed the flaw after TechCrunch helped him report it. [38]
    • Its rewards programme had at least 8.5 million members last year. [38]

    Who is involved

    • Robert Wiggins security researcher

    Why it matters

    The app has over a million downloads on Google Play, and every copy held the key. [38]

    How it works

    Security engineers call this hardcoding credentials: a login key written into the software itself. [38]

    The truth

    • Applebee's will not say how many customers were exposed or whether it has access logs. [38]
    • Nobody knows yet

      How many Applebee's customers had their details exposed.

      Applebee's would not give a number or say whether it has logs that show who reached the database. [38]

    Good news

    • Applebee's closed a flaw that let anyone with its app reach registered customers' names, emails and birthdays, after a researcher and TechCrunch reported it.
25. 27

    NHS warns GPs after a website firm fails

    What happened

    • NHS England wrote to family doctors after a firm running their websites collapsed. [39]
    • Silicon Practice Digital Health Solutions has gone into administration, the UK process for failing firms. [39]
    • Some surgery websites have gone offline completely, reports suggest. [39]
    • Potentially hundreds of practices around the UK are affected. [39]
    • NHS England's technology chief, Kevin Humphries, set out what practices should do first. [39]
    • That includes saving old patient information and clearing the backlog of online requests. [39]

    Who is involved

    • NHS England runs England's health service
    • Family doctors GP surgeries
    • Kevin Humphries NHS England's tech chief

    How it unfolded

    Many surgeries had moved patient contact online, so the collapse cut a main way in. [39]

    Why it matters

    Patients may struggle to reach their surgery while it sets up other ways to talk. [39]

    The truth

    • Most GP practices are independent contractors, so NHS England can advise but not run the fix. [39]
    • Who pays Patients of GP surgeries — Websites went offline after their supplier collapsed, cutting a main way to contact their surgery. [39]
26. 28

    Altman says memory belongs in data centres

    What happened

    • OpenAI's boss Sam Altman said lots of memory in home computers is quite terrible. [40]
    • He meant DRAM, the fast working memory a computer uses while it runs. [40]
    • He said memory in home computers is used about 5% of the time. [40]
    • In data centres, he said, the world's limited memory could be used far more fully. [40]
    • Memory chips are short because AI data centres are buying so much. [40]

    Why it matters

    Running AI at home could one day cut what people pay for AI subscriptions. [40]

    How it works

    Engineers call this utilisation: the share of time a machine is actually working. [40]

    The truth

    • OpenAI sells AI that runs in its own data centres, so Altman's view suits its business. [40]
27. 29

    A study asks why AI has not lifted output

    What happened

    • A University of Manchester report asks why AI has not yet lifted output per worker. [41]
    • Firms pay up front for retraining and reorganising, and the payback comes later. [41]
    • AI can speed up single tasks without making a whole job faster. [41]
    • Developers using GitHub Copilot finished about 26% more tasks, one study found. [41]
    • AI coding tools raised coding activity about 240%, but software releases only about 30%. [41]
    • The gap comes from people still checking and testing the work. [41]

    Why it matters

    Of six countries, Britain has the biggest share of workers in jobs AI touches and supports. [41]

    How it works

    Economists call these complementary tasks: human work that becomes the limit once AI speeds the rest. [41]

    The truth

    • The report was paid for by Bytes Technology Group, a company that sells IT services. [41]
    • Nobody knows yet

      When AI's gains will show up in output per worker.

      The Manchester report says the costs come first and the payback later, and gives no date. [41]

28. 30

    Renting out a Mac's chip pays little

    What happened

    • Darkbloom lets Mac owners rent out their computer's graphics chip to run other people's AI requests. [42]
    • About 800 people have joined, earning a few cents to a few dollars a day. [42]
    • The 50 people at the top of its leaderboard earn a median of $5.26 a day. [42]
    • Darkbloom estimates a powerful Mac mini could earn up to $192 a month. [42]
    • That assumes the machine works around the clock on a popular model. [42]
    • Falling prices for AI work have made Darkbloom rethink its plans. [42]

    Why it matters

    When AI answers get cheaper, people renting out spare machines earn less. [42]

    The truth

    • The Register says even Darkbloom's best case may not last. [42]
29. 31

    Nvidia's boss courts Kazakhstan

    What happened

    • Nvidia's boss Jensen Huang visited Kazakhstan, in Central Asia, for the first time. [10]
    • He said the country could become an AI computing hub for Europe. [10]
    • Kazakhstan plans to nearly double its power capacity to over 40 gigawatts by 2035. [10]
    • Its Data Center Valley in Ekibastuz should open a first 50-megawatt phase in 2027. [10]
    • Huang said a Kazakh AI model of its own makes complete sense. [10]
    • He met President Kassym-Jomart Tokayev, who called AI a national priority. [10]
    • Huang called a national AI model a security concern, like a country's own grid and internet. [10]

    Who is involved

    • Kassym-Jomart Tokayev Kazakhstan's president
    • Jensen Huang Nvidia's boss

    Why it matters

    Huang said countries could turn their energy into AI and export that instead. [10]

    The truth

    • The first 50-megawatt phase is a small start against the 1 gigawatt the site could reach. [10]
30. 32

    Musk and Zuckerberg adopt Trump's word

    What happened

    • Elon Musk and Mark Zuckerberg said superintelligence instead of AI on their earnings calls. [43]
    • President Trump ordered US government agencies a month ago to use that word. [43]
    • This week he said he is banning the terms AI and artificial intelligence. [43]
    • The bosses of Microsoft, Amazon and Alphabet kept saying AI. [43]

    Who is involved

    • Elon Musk Trump ally
    • Mark Zuckerberg Trump ally

    How it unfolded

    Trump gave the first order at a White House lunch for AI leaders on 29 September. [43]

    1. 29 September order to agencies
    2. This week AI banned as a word

    Why it matters

    The words show which tech bosses line up with the Trump administration in public. [43]

    The truth

    • Musk and Zuckerberg are President Trump's leading tech allies, CNBC reports. [43]
31. 33

    AI firms hire people to win over towns

    What happened

    • AI companies are advertising six-figure jobs for people who can win over towns facing data centres. [44]
    • Google wants negotiators ready to attend contested council hearings at short notice. [44]
    • Microsoft seeks people to lead town halls and community relations for its build-out. [44]
    • Anthropic offers at least $265,000 a year for roles with half the time on the road. [44]
    • Postings like these have risen, a Semafor analysis found. [44]

    Why it matters

    The jobs exist because local pushback against data-centre projects keeps growing. [44]

    The truth

    • Google's advert says its negotiators will counter misinformation at local hearings. [44]
32. 34

    1,300 African AI start-ups apply

    What happened

    • More than 1,300 companies applied to the first group of a new African AI start-up programme. [45]
    • The Askya Growth Platform, backed by venture investors, launched this month. [45]
    • Founders in Lagos, Nigeria's biggest city, told Semafor they want to build AI, not rent it. [45]
    • Babacar Seck of Askya said Africa cannot be renting AI, it must be making it. [45]
    • Ideas include small AI models for African languages. [45]
    • Others want systems that let governments and businesses keep sensitive data under their own control. [45]

    How it unfolded

    Many say Africa lacks the computing power, money and energy to compete in AI. [45]

    Why it matters

    Local models could put AI to work on Africa's own energy, farming and trade. [45]

    The truth

    • Semafor calls the 1,300 applications only anecdotal evidence for the founders' hopes. [45]

    Good news

    • More than 1,300 African companies applied to the first group of a new African AI start-up programme.
33. 35

    A factory to build data centres like cars

    What happened

    • Runware, an AI start-up, wants to build data centres on a factory line, like cars. [46]
    • It already sells the Sonic Inference Pod, a small data centre shipped ready to install. [46]
    • Its factory in Romania aims to make 30 megawatts of capacity a day by 2028. [46]
    • That is nearly 1 gigawatt a month. [46]
    • The whole US was expected to add about 10 gigawatts of data-centre power this year. [46]

    Who is involved

    Crusoe and Schneider Electric also sell ready-made pods. [46]

    Why it matters

    AI firms want computing power sooner than ordinary data centres can be built. [46]

    The truth

    • Runware says it can ship faster than rivals; that is the company's own claim. [46]
34. 36

    A newer web standard speeds a delivery app

    What happened

    • Delivery Hero, a German food delivery company, moved most users to HTTP/3, a newer web standard. [47]
    • That cut the time its app's home screen takes to respond by 17%. [47]
    • Staff engineer Elvis Chidera described the work at the P99 conference. [47]
    • He said HTTP/3 is faster than HTTP/2 but has slow parts that need tuning. [47]
    • Delivery Hero serves about 60 million users outside the US, from Europe to Asia. [47]

    Who is involved

    • Elvis Chidera Delivery Hero engineer

    How it unfolded

    Uber is buying Delivery Hero for $14.8 billion. [47]

    Why it matters

    A faster home screen means less waiting for millions of people ordering food. [47]

    The truth

    • The speed came from work, not the switch alone: some parts needed adjusting. [47]

    Good news

    • Delivery Hero cut the time its app's home screen takes to respond by 17% after moving most of its users to HTTP/3.
35. 37

    Police may be able to trace 3D-printed guns

    What happened

    • Australian researchers found a way to link 3D-printed plastic parts to the filament they came from. [48]
    • Filament is the plastic thread a 3D printer melts to build an object. [48]
    • The Curtin University team matched 26 of 28 samples correctly. [48]
    • The study appeared on 22 October in Forensic Science International. [48]
    • The test heats a sample and measures how heat flows through it. [48]
    • So far it works only for PLA, a common printing plastic. [48]

    How it unfolded

    Printed ghost guns carry no serial numbers, so police struggle to trace them. [48]

    Why it matters

    A match could link a seized gun to filament seized elsewhere. [48]

    The truth

    • The team printed with only one machine, so real cases may behave differently. [48]

    Good news

    • Police could gain a first way to trace plastic guns back to their material. [48]
    • Researchers at Curtin University in Australia matched 26 of 28 3D-printed plastic samples to the filament they were made from, a possible first way to trace printed ghost guns.