invented
his appeal quoted four witnesses. none of them had ever existed.
stephen d. aarons, attorney, santa fe. oscar renee sandoval, his client, serving life. five justices of the new mexico supreme court.
stephen aarons has practised law for more than forty years. last year he filed the appeal for oscar renee sandoval, sixty, of chaparral, convicted of murder and serving life. the brief quoted testimony from officer michelle amarillo, officer sanchez, manal al-jibury and teresa marquez. the court found that none of the four exist.
it also put false words in the mouths of real witnesses, and it misstated two new mexico cases, state v. lopez and state v. manus. aarons had loaded the trial transcripts into chatgpt and asked for what he called a bulletproof summary. in his own written reply he said the tool produced, out of whole cloth, fictitious witnesses, testimony, quotations and authorities in a form that appeared coherent and plausible.
at a hearing on 21 august, justice c. shannon bacon asked him whether he watches the news, because lawyers relying on ai hallucinations is an above the fold story every single day. the written order followed on 9 september. direct contempt. five thousand dollars to the state bar's client protection fund within thirty days. barred from that one court while the disciplinary board investigates. every brief in the case struck.
the order also found that aarons never told his client the brief carried misrepresentations, and never told him about the show cause proceedings. sandoval's appeal now starts again from nothing, with a public defender, and the court intends to hear it in its 2026 to 2027 term. all five justices signed.
the lawyer lost one courtroom and five thousand dollars. the man serving life lost a year he did not know he was losing.
who benefits: nobody, though the court got its above the fold story. who pays: sandoval, whose appeal was erased without anyone telling him. what should change: a client gets a copy of any order to show cause about his own lawyer, from the court, on the day it issues.
the long wayopenclose
the order keeps two failures apart. one, he did not check the brief before he signed it, and did not tell his client it was wrong. two, separately, he did not tell his client that the court had ordered him to explain himself, and did not send him those papers. at the hearing he cited the man's limited english and his incarceration.
the court found a lack of remorse and a lack of concern for his client. at the hearing the chief justice, julie vargas, thanked him for admitting he was wrong. afterwards he told reuters he is remorseful but hopeful the disciplinary board takes into account that it was an honest mistake. three positions, three speakers, none of them merged here.
the sanction is narrow on purpose. it is not a fine paid to the court, it is not a suspension, and it is not disbarment. the santa fe new mexican counts thirty eight active cases he still has in state courts. the disciplinary board decides the rest.
the case file is a murder. this chair stays on the lawyer and the four people he invented, and never on the killing.
where it comes from
the order of the new mexico supreme court, docket s-1-sc-40845, dated 9 september 2026, read at length in eugene volokh's post at reason.com, 14 september, which quotes it verbatim and credits ars technica. the santa fe new mexican, phaedra haywood, 21 august 2026, from inside the hearing. reuters, 11 september 2026, read at al jazeera.
what is not settled
the substance broke on 21 august, three weeks before this window; the 9 september order is the in-window act and the only reason this chair sits here. the order pdf itself was not opened; its text is known through volokh's quotation. the public defender's name, kim chavez cook, is reuters' and is not in the order, which appoints the office. the thirty eight cases and forty years are the local reporter's.
what was withheld
the two accounts of who sandoval was convicted of killing disagree, and neither prints, because the chair is not about the killing. a personal loss aarons cited in his written reply is on the record and stays out; it has nothing to do with the take.
slop
a student maths contest, two million dollars in ai credits. 771 mathematicians objected. openai took its million home.
caiman moreno-earle, 19, caltech sophomore, and four fellow student organisers. dan roberts, openai. terence tao, peter scholze and two dozen other fields medallists.
five caltech undergraduates planned a mathathon for 30 october. a hundred teams, forty hours, and about twenty thousand dollars of ai credits each, two million in all, half from openai and half from anthropic. the question set for the teams: how can ai tools responsibly augment human understanding of mathematics.
on 10 september a group of current and former caltech mathematicians published an open letter against it, 771 signatures at publication. it names five objections, the first of them the creation of slop mathematics, meaning results produced fast and left for other people to verify, unpaid and uncredited. one line: there is no reasonable future model of mathematical research in which every research mathematician receives twenty thousand dollars in ai credits.
the same day openai withdrew. dan roberts, who leads research there, said the company recognises that the rapid progress of ai in mathematics is disruptive. the next web puts the withdrawn credits at one million dollars. anthropic kept its half. the students rewrote their rules the same day, and a line on their site asking what the role of a mathematician is when ai can solve conjectures faster had already come down.
the day after, fields medallists published their own declaration. the page read today carries twenty seven names, among them terence tao, peter scholze, maryna viazovska and pierre deligne. solving problems, it says, is only a tool and proxy for conceptual understanding and insight, and mass producing true or false statements at faster and faster pace could destroy fertile ground.
nobody in this fight thinks the machines cannot do the maths. the fight is over who has to check it.
who benefits: the labs, who get a hundred teams stress testing their models for a weekend and a headline either way. who pays: whoever reads the proofs afterwards, which the letter says is the profession, unpaid. what should change: a contest that pays for the generating should pay for the checking, and name who does it.
the long wayopenclose
the letter's second and third points are about the sponsors, not the students. corporate interest, and conditions unlikely to produce understanding. its fifth is a failure to care for young mathematicians, which is a hard thing to say to five of them.
the students answered on 10 september under their own names, alvan arulandu, sathvik rodrothu, caiman moreno-earle, avni garg and brian zhao. second round entries now need an arxiv preprint, there is a parallel track for course notes and explainers, and faculty will read their public statements first. a spokesperson told gizmodo they do not expect the withdrawal to affect the event in any substantial way and are talking to other firms.
the week had a prologue. on 8 september openai said its agents had solved the navier stokes problem, one of the seven millennium prize problems. tristan buckmaster of nyu, who had been working the same line with levent alpöge of anthropic, said publicly that openai's offer to him included leaving alpöge's name off. sebastien bubeck of openai denied it and said the company recognises their priority. the mathathon letter landed two days into that argument.
moreno-earle, to the next web: they do not have the luxury of waiting around to find out.
where it comes from
the open letter at proofsandprompts.com, 10 september 2026, and the organisers' response at the same site, same day. the fields medallists' declaration at mathandai.org, 11 september. gizmodo, tom mckay, 11 september. the next web, ana maria constantin, 14 september, who spoke to moreno-earle. fortune, jeremy kahn, and mit technology review, grace huckins, both 8 september, for the navier stokes prologue.
what is not settled
dan roberts' statement was posted to x and was read only as quoted by gizmodo and the next web. the one million figure is the next web's; gizmodo prints it as ten thousand of each team's twenty thousand. the next web printed twenty five fields medallists on 14 september and the page today lists twenty seven, so the number moved. whether the event now has a replacement sponsor is not reported. openai's navier stokes claim has not been independently verified in any source read.
what was withheld
the organiser's surname is spelled two ways across the next web's own article; the spelling here is the one on the students' signed response. the buckmaster account includes quotes he attributes to a private call, which openai denies; they do not print.
parents
his parents sued the chatbot company. this year california signed a law with his name on it.
matt and maria raine, parents of adam raine. senator steve padilla, san diego. governor gavin newsom.
adam raine was sixteen when he died in april 2025, after months of conversations with chatgpt about ending his life. his parents sued openai. on 10 september governor newsom signed thirteen bills about children and technology in one day, and the one about chatbots is called adam's law.
senate bill 1119, by steve padilla, says a companion chatbot that talks to a child must know roughly how old the child is, must assess its risks before release, must offer crisis help and a mental health referral inside the app, and must tell a parent when a child turns a safety setting off. notifications off, time limits on and a short memory are the defaults, and only a parent can change them.
the companies are liable for certain outputs to minors: self harm content, sexual content, romantic roleplay, excessive praise, emotional manipulation. an independent audit goes to the attorney general every year, and a family can sue. the senate passed it without a no vote and the assembly 64 to 4.
a second padilla bill, sb 867, stops toys with companion chatbots inside them from being made or sold in the state for four years, through the start of 2031. matt and maria raine, in the senator's release: they still have not adjusted to life without adam, but are pleased that an element of his legacy is to help make ai chatbots safer for minors.
the defaults are the law. a child can no longer be the one who turns the safety off.
who benefits: every child in california who opens a companion app after the law takes effect, and every parent who now gets told. who pays: the chatbot companies, in audits and in lawsuits they can no longer settle quietly, and toy makers who lose a state for four years. what should change: the other forty nine states copy the defaults before they copy the exemptions.
if you or someone you know is struggling, call or text 988 in the united states, at any hour.
the long wayopenclose
the shape of this law is the opposite of the ones that list who is exempt. it starts from what the product must do for a child and lets the company work out how. reuters notes that openai has itself acknowledged its safeguards became less reliable during long conversations, which is the exact condition the memory and time limits are aimed at.
the toy ban is the strangest line in the package and the easiest to repeat at dinner. a teddy bear that talks back and remembers you is illegal to sell in california until 2031. the toy association, which opposed it, told its members it will share compliance guidance and that it had talked congress out of a federal version in favour of a study.
the rest of the day: no algorithmic feeds or autoplay for anyone under sixteen, civil penalties for harm to children, ai generated sexual images of minors added to the criminal code, and up to two hundred and fifty thousand dollars per action for deepfake victims. the electronic frontier foundation, on the age verification side of the package, called it a privacy and free speech nightmare.
the raines' case against openai is still a lawsuit. the law does not decide it. it decides what the next product must do before a child talks to it.
where it comes from
the office of senator steve padilla, 10 september 2026, with the raine family's statement. contra costa news, 11 september, carrying the family's statement and the vote counts. the governor's release, 10 september, for the thirteen bills. reuters, steve gorman, 10 september, read at the spokesman review. the toy association, 10 september, for the dates of sb 867.
what is not settled
the bill texts at leginfo could not be read, so every requirement rests on the senator's and the governor's descriptions and reuters. effective dates for adam's law are not printed in any source read. the toy ban's age line differs by source: the toy association says toys for children under sixteen, the may bill text at legiscan says under eighteen; the chair prints neither. the raines' words come from the senator's release and were not given to a reporter.
what was withheld
how adam died, and anything from his conversations, which the family's complaint describes. neither prints here or ever will.
scanner
she left the warehouse in a wheelchair. the scanner kept counting. six days later she was fired.
willamina barclay, kristina green and dazaria parks, rochester. jennifer hatch, buffalo. debbie greenberger, emery celli. inimai chettiar, a better balance.
on 8 september four women who worked in amazon warehouses in western new york filed a nationwide class action in federal court in brooklyn. each was pregnant. each says she asked for a chair, or a break, or a lighter load, and was told to wait for paperwork. each was fired.
the complaint describes two clocks. unpaid time off is a bank of hours; when it runs dry an automatic email gives forty eight hours to justify the absence or be dismissed. time off task is measured by the handheld scanner, minute by minute; thirty minutes in a day earns a written warning, and a hundred and twenty minutes in a day, or thirty on three days in a year, earns termination. the complaint says nothing in either system tells a protected pregnancy break from slacking.
willamina barclay, by the complaint's account, was wheeled out of the rochester site on 16 june 2025 with severe abdominal pain, in view of hr staff, and taken to hospital. the system deducted her hours while she was there and sent the warning. on 22 june she was fired for a negative balance. kristina green's final warning, in april 2026, cited 126 minutes of unverified inactive time; she says she was in amazon's own lactation room, pumping. she was fired on 7 may.
these are allegations, and amazon has not answered in court. but in february the eeoc, on jennifer hatch's own charge, found reasonable cause to believe amazon systemically violated the pregnant workers fairness act by failing to accommodate pregnancy nationwide, in the words of both releases. hatch had been fired by email in april 2025 after four emergency room visits.
the machine did not decide anyone was lying. it was never built to ask.
who benefits: a company that runs hundreds of warehouses on one policy and one email template. who pays: four women, and by the eeoc's finding a nationwide class of them. what should change: a clock that can fire a person has to stop when a doctor's note arrives, and a human has to be the one who restarts it.
the long wayopenclose
the four are kept apart because the releases keep them apart. hatch, buffalo: denied short breaks and permission to sit, fired after a pregnancy related hospitalisation. barclay, rochester: denied sitting breaks or a chair, fired after a medical emergency. green, rochester: penalised repeatedly for pregnancy breaks, fired for time spent pumping. parks, rochester: denied restroom breaks and a seat, fired in july this year for missing work after amazon's refusal to accommodate her forced her onto leave. she was still pregnant when the case was filed.
the complaint says a manager told barclay that if he gave her a chair he would have to give everybody a chair. it also says a central department handles accommodation requests for every warehouse by the same automated emails, which is the sentence that turns four women into a class.
debbie greenberger, counsel, says the pregnant workers fairness act became law years ago and amazon should have well established policies to comply. inimai chettiar, who runs a better balance, says no corporation is above the law, not even amazon. a better balance filed a separate nationwide class action in november 2025 for warehouse workers with disabilities.
where it comes from
the complaint, barclay et al. v. amazon.com services, llc, no. 1:26-cv-05531, eastern district of new york, stamped filed 8 september 2026, read at the plaintiffs' counsel's site. the releases of a better balance and of emery celli brinckerhoff abady ward and maazel, both 8 september 2026.
what is not settled
everything here is the plaintiffs' account. amazon has not answered and no page read is amazon's. the eeoc's finding is quoted as the two releases word it; the complaint's own quotation of it is phrased differently. the six days rest on the complaint's dates, 16 and 22 june; earlier drafts of this chair carried five and that number does not print. no wire coverage was read.
what was withheld
the hoodline line that an algorithm flagged medical breaks is in no primary and does not print. the women's medical diagnoses are in the complaint and stay there.
watcher
the government is paying for software to treat heart failure. and for a second program to watch the first.
haider warraich, cardiologist and programme manager, arpa-h. rick abramson, fda digital health center of excellence. sharif vakili, updoc. ankeet bhatt, kaiser permanente.
on 9 september arpa-h, the research agency inside the department of health and human services, announced the first awards under a programme called advocate. up to 33.7 million dollars in the first year of a four year, 62.7 million dollar effort to build the first ai authorised by the fda to treat patients directly: heart failure patients, between visits, adjusting medication and ordering tests, escalating to a clinician when it judges it should.
three companies build the agent. one university builds the thing that watches it. stanford's award, up to fifteen million by fierce healthcare's count, is for a supervisory ai whose only job is catching unsafe recommendations from the clinical agents, in real time. the agency calls it disease agnostic. the supervision of a doctor's judgement is now its own line of federal research.
duke and kaiser permanente test it in the world. kaiser's own release, 9 september, describes three phases across up to twenty one northern california medical centres, ending in a randomised trial of about 2,500 patients. nothing is enrolled. nothing is running. no patient has been treated by any of it.
the companies must submit an fda authorisation package within twenty four months. rick abramson, who directs the fda's digital health center, is quoted inside the funder's own announcement: getting the framework right requires close, iterative collaboration between developers, clinicians and the fda. the regulator who must say yes is already in the press release.
the doctor is not being cut out. the doctor is being re-implemented, twice.
who benefits: patients in the half of american counties with no cardiologist, if it works, which nobody yet knows. who pays: the taxpayer for the build, and the first 2,500 patients for the proof. what should change: the watcher's verdicts get published, so that when the agent is wrong the public learns it from the record and not from a lawsuit.
the long wayopenclose
the numbers the agency gives for the problem: more than 200,000 americans a year die of preventable cardiovascular disease, nearly half of counties have no cardiologist, and the country spends nearly half a trillion dollars a year on heart disease. its projection, conditional on success, is an estimated 28 billion dollars a year saved across heart failure alone.
haider warraich, who runs the programme and still practises cardiology, describes the goal as a technology that can serve as a clinician extender, capable of providing certain care autonomously and equally capable of engaging the clinical team when needed. sharif vakili of updoc, whose company already has an fda cleared insulin tool, told fierce that heart failure is the moonshot indication for clinical ai. the fda has so far authorised predictive ai only, never generative ai in a high risk setting.
johns hopkins applied physics laboratory is the independent evaluator. vincent liu, chief data officer at the permanente medical group, says ai has to be evaluated in real clinical settings with the same rigour as any new medical intervention. that sentence is the whole programme in one line.
this chair is the close because everything in it is ahead of the reader. the award is the only thing that has happened.
where it comes from
arpa-h's own release, 9 september 2026, at arpa-h.gov. fierce healthcare, heather landi, 11 september, who spoke to vakili. kaiser permanente division of research, sue rochman, 9 september.
what is not settled
every dollar figure is a ceiling and prints as up to. the agency's release carries no per award figure; the fifteen million for stanford is fierce's, and kaiser's own release says up to sixteen million over three years where fierce prints 16.3. kaiser's centre count is the agency's twenty one against kaiser's up to twenty one; the institution's number governs. the twenty four month deadline binds the three agent builders only. no stanford document was read.
what was withheld
a phrase calling this the manhattan project of clinical ai is in fierce as vakili recalling warraich, one hand further from the speaker than a hook can stand on, and does not print. no total of the six award ceilings prints, because the funder never adds them.