pace
four rival ai labs agreed in public to slow down. six days later four of their own paying subscribers sued them for it.
charles buist and nick spetsas, attorneys in florida. cheyenne hunt and christine bullock in california. each pays for chatgpt, claude, grok or gemini.
on 12 september dario amodei, who runs anthropic, published an essay titled we must pace the frontier. it asks every frontier lab to slow down together, and it warns that within six to twelve months a swarm of ai agents could be capable of taking over the internet.
the associated press reports that sam altman, elon musk and demis hassabis each agreed in public the same day. the essay saw its own problem coming. for antitrust reasons, it says, washington would need to issue a narrow waiver for that kind of conversation.
no waiver was issued. on friday 18 september four paying subscribers filed suit in federal court in san francisco against anthropic, openai, spacexai and google, calling the agreement an illegal deal between competitors. they want it to cover every paying subscriber in the country.
senator josh hawley had already told a hearing there is no world in which he hands those companies an exemption. the four companies had not answered requests for comment by the weekend.
wemustslowthepace
four companies found a way to stop racing each other without calling it a truce. the customer still pays for the fast one.
who benefits: the labs, which get to slow down together. who pays: the subscriber who bought speed and now owns less of it at the same price. what should change: an agreement between competitors goes in front of a regulator before it is announced, not after it is sued.
the long wayopenclose
the essay's plan has three steps, under amodei's own headings: embedded evaluators, democratic coordination, global coordination. the first would give outside evaluators ongoing, employee-like access inside every frontier lab.
altman's reply, as the wire carries it, welcomed a federal framework and said openai does not believe it needs to wait for an antitrust exemption or legislation to begin.
politico's report on the filing describes buist and spetsas as florida attorneys and hunt as a lawyer. nick rowley leads for the plaintiffs. his statement says safety cannot be left to private agreements between the most powerful for-profit technology companies in the world.
so the argument on both sides is safety. the labs say slowing down together is the safe thing. the plaintiffs say four companies deciding that in private is the dangerous thing. a court gets to pick.
where it comes from
associated press, carried by cbs news and fortune, 19 september 2026, counted as one report. politico, tyler katzenberger and christine mui, 18 september, read at a carry. the tribune, carrying ani, 19 september, for the four names. the essay at darioamodei.com.
what is not settled
the complaint itself was not read; it sits behind a link this line could not open, and no source read prints a case number. that the three chief executives agreed the same day is the wire's account of their posts, which were read only secondhand. the essay's page prints september 2026 and no day; the twelfth is the wire's.
what was withheld
one rich account of the complaint was set aside whole because its byline is a machine by its own author box. a case name that appears only in a second machine-written article does not print.
flyer
she showed up for the labor day festival. there had never been a festival.
renae miguel, houston. leah napoliello, better business bureau of greater houston and south texas.
renae miguel saw a flyer on social media for a labor day weekend event at eleanor tinsley park in houston, and went. fox 26 reports that she arrived at an empty park and found other confused families walking around it.
nobody had cancelled anything. a houston labor day weekend celebration festival at the park, 5 to 7 september, ten to five each day, live music, food vendors, skyline views, was carried by at least two event listings, one of them a local magazine's holiday guide.
miguel filmed the moment she worked it out and the video travelled. she told the station it could have happened to anyone, in a city where so many people are trying to host something. she lost the day. the station says she dodged a bigger financial hit.
leah napoliello investigates for the better business bureau. she says people pay upfront for tickets or a vendor booth, show up, and lose a lot of money, and that whoever is behind it wants prepaid debit cards, cashapp, gift cards or crypto, because that money cannot be followed.
- what
- labor day weekend celebration festival
- where
- eleanor tinsley park
- when
- september 5 to 7 · ten to five
- who is running it
- how to reach them
the picture did not need to fool anyone into paying. it only had to cost her the day.
who benefits: whoever sells tickets and vendor booths to an event that is only a picture. who pays: miguel paid a day, and the people napoliello hears from pay in gift cards. what should change: a site that lists an event should be able to say who is running it.
the long wayopenclose
fox 26's advice is short and all of it is about what a fake flyer leaves out. check the venue's own site. do not buy through a link in a post. fake flyers, the station says, often lack a specific address and any contact for an organiser.
that is the whole tell. the poster, the logos and the vendor lineup now cost nothing to make, so they prove nothing. the only part of a flyer that still costs something is a name and a phone number that answer.
both listings were still up when this issue closed.
where it comes from
fox 26 houston, sakura gray, 18 september 2026, which spoke with miguel directly. the festival's name, dates and hours come from two event listings read in full, one dated 4 september.
what is not settled
fox 26 calls the flyer ai generated and does not say who established that; no named person in the piece says so. the station never names the event; the link between miguel's flyer and the listed festival rests on the park, the weekend and a houston chronicle headline this line could not open. whether anyone lost money on this event is not reported.
what was withheld
her video was not found. a loss figure for miguel does not exist and none prints. one word in napoliello's second quote looks like a transcription slip on the station's page, so the quote is paraphrased.
human
four insurers signed a rule that no machine turns a claim down alone.
sean scanlon, comptroller of connecticut. tom swan, connecticut citizen action group.
on wednesday 16 september sean scanlon, connecticut's comptroller, announced five rules for the companies that run the state's health plans. the first says no denial may be made by an ai system alone. a human reviews it, always.
the others: no cutting a provider's payment or changing a billing code on a model's word alone. members' data cannot be used to train anyone else's ai. nobody talks to a machine believing it is a person. the carriers open their audit procedures to the comptroller.
ct mirror reports the talks took about a month, with anthem, cigna, aetna and caremark, and that the rules start on 1 january 2027 for more than 270,000 people on the state employee plan and the partnership plan.
it is small on purpose. a state can only regulate fully insured plans, and about 220,000 connecticut residents, six per cent, are in one. in january scanlon goes to the legislature to ask for the rest. three national insurers are already in court over algorithm based denials.
the win is real and it is small on purpose. a human signs every no.
who benefits: 270,000 people whose denial now has a person behind it. who pays: the carriers, in reviewers' hours. what should change: the same five rules for everyone the state cannot reach, which is most of the state.
the long wayopenclose
the second rule is the sleeper. nobody marched for it and it is in there anyway: what a member's body has been through cannot become training data for some other model.
the fourth is the one a member will feel first. a carrier has to say so when ai is materially assisting with, interacting on, or recommending a benefit. the fifth makes the carriers validate their systems against historical data for accuracy, consistency and fairness, and show the comptroller how.
tom swan of the connecticut citizen action group called it a really great step and said that without oversight this is dangerous. the president of the state medical society called it a step toward stopping automated denials.
these are terms the comptroller sets for plans he runs. they are not statute. that is why they could be done in a month, and why they stop at the edge of his plans.
where it comes from
ct mirror, katy golvala, 16 september 2026. the comptroller's own release, same day, for the five rules in their own words.
what is not settled
neither source says an ai-only denial has happened on these plans; the lawsuits against three insurers elsewhere allege it. rule one says a human, not a doctor. the carrier names, the start date, the month of talks and the six per cent rest on ct mirror alone.
what was withheld
a count of how many states have similar rules came from an aggregator that got two other numbers wrong in the same article. it does not print.
nice
a researcher found a way around the newspaper's paywall. the president of openai wrote back: ah nice.
greg brockman, president of openai. nick ryder, openai researcher. brent hecht, microsoft. nick turley, head of chatgpt.
on thursday 17 september the publishers suing openai and microsoft filed their summary judgment brief in manhattan federal court with the blacked out parts restored. ninety two pages. the new york times, the daily news, ziff davis, the center for investigative reporting and the intercept on one side.
page fifty six. an openai researcher named nick ryder tells greg brockman about a hack to get around the nytimes paywall, to help brockman's own efforts to scrape the site. brockman's reply, in full: ah nice.
page two. brent hecht, microsoft's director of applied science, wrote that the company's content strategy had started a doom loop, and that it is highly unusual for a product to threaten the economic foundations of its own essential suppliers. nick turley, who runs chatgpt, wrote that the products are largely substitutive, period.
hecht also predicted that millions of people would come to see models hoovering up their work as an astonishing theft. microsoft says he spoke for himself and not for the company. openai did not comment.
nobody inside was confused about what it would do to the people who make the news. they wrote it down and kept going.
who benefits: two companies that got the archive without the invoice. who pays: every newsroom whose readers now get the answer without the visit. what should change: the supplier gets paid before the product ships, which is how every other supply chain works.
the long wayopenclose
hecht's argument is not a moral one, which is what makes it hard to wave away. it is a supply chain memo. a product that takes its suppliers' work and sends them no readers eventually has no suppliers, and then nothing new to learn from. a company scientist told his colleagues the business was eating its own inputs.
the brief opens by saying he called it an astonishing theft of unprecedented proportions. eleven pages later it prints his actual sentence, which predicts that millions of people will soon consider it one. reuters kept the difference. most headlines did not.
microsoft's full answer is that the comments reflect one employee's individual perspective, are not a legal analysis, and that its court filings explain why these uses are consistent with copyright law.
where it comes from
the plaintiffs' combined summary judgment brief, document 1587-1, filed 17 september 2026 in in re: openai, inc., copyright infringement litigation, southern district of new york, read at the public docket. reuters, blake brittain, 17 september. techcrunch, rebecca bellan, 17 september.
what is not settled
this is the plaintiffs' brief. it is advocacy and not a finding, and no judge has ruled on it. the brief prints the messages without dates; techcrunch dates hecht's memo to january 2023 and the doom loop presentation to january 2024. the financial times report that led much of the coverage could not be read.
what was withheld
one widely shared carry prints brockman's reply as two different words. the filing says ah nice, and so do four outlets that read it. the garbled version does not print.
knock
she takes her ear off to sleep. so she built the thing that wakes her.
gargi agrawalla, loughborough university graduate.
gargi agrawalla was born profoundly deaf and got a cochlear implant at twelve. it let her hear for the first time. its sound processor comes off to sleep and to shower, which is fine at home and leaves her uneasy anywhere unfamiliar.
so she built sivo.go, one handheld unit to replace the several alert gadgets a deaf traveller is told to pack. a small model, trained on more than five hundred sounds and running entirely on the device, listens for fire alarms and doorbells and turns them into vibration and flashing light.
knocking beat it. in testing, the model kept confusing a knock with footsteps or with someone tapping on a table, even after she gave it more samples. so she stopped asking the model. a separate vibration sensor goes behind the door, and the door reports the knock itself.
this week it won the uk james dyson award, five thousand pounds, and it goes forward to the international round. the shortlist is announced on 14 october.
the best engineering in the story is the moment she stopped trusting the model and asked the door.
who benefits: anyone who takes their hearing off at night. who pays: nobody, which is rare on this sheet. what should change: more builders saying in public where their model failed and which dumb part fixed it.
the long wayopenclose
she describes the problem as luggage. a cochlear implant user already travels with chargers, a dry box and waterproof accessories, and the usual answer to safety is to pack several more devices on top. one unit that covers the important alerts is a lighter bag and a quieter head.
the unit can sit where it is needed or be worn on a lanyard or a clip. alerts come as colour coded light, symbols on a screen and vibration she can turn up or down. loughborough's release calls the model tinyml, which means it is small enough to live on the device with no wifi and no cloud.
the global winners are named on 4 november. the prize at that level is thirty thousand pounds.
where it comes from
dezeen, jennifer hahn, 16 september 2026, quoting agrawalla directly. loughborough university's release, same day.
what is not settled
the maker's own award page refused every read, so nothing here rests on it. the award date is taken from the two publication dates; neither page prints an announcement date. no price and no release date for the device are reported.
what was withheld
a line that she used to pack four separate gadgets was in the week's pull and is in no source. it does not print.
twelve
her daughter was thirteen. she begged the capitol not to pass its chatbot law. it exempts twelve kinds of chatbot.
cynthia montoya, thornton, colorado.
cynthia montoya lives in thornton, colorado. her daughter juliana was thirteen when she died by suicide in november 2023. montoya has spent the time since asking her state for a law about the chatbots children talk to.
colorado wrote one. npr reports that the sponsor told her she would be invited to give input, and that by her account his office then went silent for weeks. the bill was introduced in february without her. she testified against it in the senate. governor polis signed it on 29 may.
the law says what it covers and then lists what it does not: twelve kinds of chatbot, the last of them any chatbot that is a feature within another software application. the version introduced in february listed six.
npr found that same sentence in bills in at least eight states, and google lobbyists registered in support in four. a hawaii lawmaker says lobbyists for google handed her the framework and she declined to sponsor it. google says it supports thoughtful, effective legislation. the colorado group behind the bill says it is not an industry bill in disguise, and the sponsor says tech companies cannot be legislated without their input.
chatbot
- 01tools for developers and researchers
- 02commerce and customer service
- 03narrow, single topic tools
- 04products marketed to businesses
- 05voice assistants and device interfaces
- 06tools for internal business use
- 07features inside a video game
- 08features inside a theme park
- 09health care entities under hipaa
- 10entities under the state's health care availability act
- 11limited educational tools
- 12a feature within another software application
she asked for a duty of care and the state wrote a list of who does not owe one. twelve lines long.
who benefits: any company whose chatbot lives inside another product, which is most of them. who pays: the next family. what should change: exemptions get written in public, with the name of whoever drafted them on the page.
if you or someone you know is struggling, call or text 988 in the united states, at any hour.
the long wayopenclose
montoya's own words to npr: her work is now harder because they passed this bill, and they passed it with her screaming from the mountaintops, begging them not to.
the law requires a company to institute technically feasible measures. laura marquez-garrett, an attorney at the social media victims law center, points at who decides what is feasible: the company.
in five states npr found identical language exempting any chatbot that works as a voice assistant for a consumer device, with modified versions in ten more. arizona's governor vetoed her state's bill in june. colorado's duties begin on 1 january 2027.
the sponsor, sean camacho, is on the record defending it. he says there had to be enough exemptions for the commercial use of ai.
where it comes from
npr, katie mcque, an npr exclusive, 18 september 2026. the enrolled text of colorado hb26-1263, where the twelve exclusions are the bill's own numbered list. the legislature's bill history for the signing date.
what is not settled
that the colorado group worked with google on the bill rests on one anonymous source, and that a google lobbyist shopped the bill rests on one anonymous republican lawmaker. both are npr's reporting and neither prints here as fact. npr lists chatbots housed in search engines among the twelve; the enrolled text this line read does not. the arizona veto letter could not be read.
what was withheld
the contents of her daughter's chat logs, which npr's piece describes in montoya's words. nothing about how juliana died prints here or ever will.
almost
a chatbot told an analyst the ship carried nuclear weapons parts. the planes were already in the air.
nobody in this story has a name. cnn reports it from four sources who would not give theirs.
this spring, during the war with iran, an american military analyst attached to special operations command pacific wanted to know what a chinese ship in the middle east was carrying. cnn reports that the analyst asked a chatbot.
the answer was components of a nuclear weapons programme. then, by cnn's account, the analyst used the same tool a second time, to write the answer up as a standard intelligence report, and the report circulated across the military.
two sources told cnn that armed personnel were preparing to board. two said military aircraft were already airborne. shortly before the operation someone looked closer, found where the intelligence had come from, and it was called off.
one source called the report entirely false and said it almost started a war. cnn could not learn what the ship was actually carrying, or which chatbot it was. a former senior official told the network that the military's internal tools are mostly copies of the commercial ones.
the wrong answer is not the story. the second prompt is. the machine wrote its own paperwork.
who benefits: nobody, this once, by minutes. who pays: next time, whoever is on the ship. what should change: an intelligence report carries the name of the human who checked it, and the tool that produced a claim never gets to write the report about it.
the long wayopenclose
every safeguard in an intelligence chain is a person between a claim and an order. the first prompt produced a claim. the second prompt dressed it in the format that people downstream are trained to trust. by the time it reached them it looked like everything else they act on.
the pentagon told one outlet that it does not comment on specific intelligence reports.
this chair is the close because it is not a forecast. it already happened, this spring, and it was caught by somebody deciding to look closer with the aircraft already up.
where it comes from
cnn, 18 september 2026, as carried in agreement by seven outlets read in full, among them gcaptain, the jerusalem post, mediaite and the new republic, plus the header of cnn's own text at ctv news.
what is not settled
cnn's own page cannot be opened by this line, so every detail here is cnn's as carried, checked across carries that agree with each other. what the ship carried and which chatbot was used are unknown, because cnn could not establish either. whether the two commands declined to comment or did not respond differs between carries.
what was withheld
several carries add context cnn did not report: which models the military licenses, a state visit, a named vendor. none of it prints, because none of it is tied to this incident by anyone who would know.