flicker
he wore a fake face on the video call. it lagged for one second. the police were watching.
nobody in this story is named. spain's national police announced the arrest on 11 august and did not name the man.
in spain a digital certificate is the thing that lets a person sign documents online. to get one you hold your identity card up to a camera and a person on the other end checks it against your face. a man in murcia wanted certificates in other people's names, so he changed his face to match.
the police say he sat in front of the camera holding forged cards while software altered his features live to match the photograph on each one. the cards had holograms that catch the light, so he rigged household spotlights with coloured bulbs to make the flashes. he used vpns, and more than 320 phone lines across 24 devices, most of them bought in shops in murcia under other people's names.
he tried this 38 times, for more than 30 different people. a certificate company noticed the pattern of requests, and the police were told. then, during one live check, the face software fell behind by about a second. the mask dropped. the person running the check saw the real face.
he was arrested for the continued falsification of official documents. the police took an encrypted laptop, phones, storage drives and papers. the register reports that some certificates had been issued before he was caught.
the trick worked until the computer got tired. a human was still on the other end to see it.
who benefits: whoever sells the next face changer, because this one only failed on speed. who pays: the thirty people whose names were used, and everyone whose signature now needs a slower check. what should change: the live check keeps a person on the line, and that person gets told what a one second glitch means.
the long wayopenclose
the spanish interior ministry's own line is that the program that altered his features suffered a small processing delay. that is the whole case. the forgery was good enough to pass thirty eight times at the document stage. it was the machine, not the paperwork, that gave him away.
the certificate provider is the quiet hero. it flagged the pattern of requests before the glitch, which is why an officer was watching when it happened.
the coloured bulbs are the detail to keep. a real identity card throws specific colours back at the camera. he could fake a face with software, but the holograms he had to fake with a lamp.
where it comes from
the register, connor jones, 11 august 2026. help net security, sinisa markovic, 12 august. the spanish interior ministry's release for the policia nacional, 11 august, read in spanish.
what is not settled
no date of arrest is given by anyone; 11 august is the day the police announced it. the ministry does not say whether any certificate was actually issued; that some were is the register's line. the man's age, nationality and the software he used are not reported. what he meant to do with the signatures is described by help net security as financial fraud and by the police only as falsification.
what was withheld
the man is not named by the police, so no name prints, and no carry that offers one was read. what he planned to sign with the certificates is not on the record and is not guessed at.
hour
he asked two towns for one hour of camera footage. they said no. it cost them eighty thousand dollars.
jose rodriguez, 36, of milton-freewater, oregon. tim hall, his attorney. dan curtis, city attorney of sedro-woolley. jo ellen kesti, mayor of sedro-woolley. sid roberts, mayor of stanwood.
flock cameras sit on poles and photograph every licence plate that passes, and software reads the plate. in april 2025 jose rodriguez asked the city of stanwood, washington, for the pictures its cameras took in a single hour. in may he asked sedro-woolley for the same. both cities refused.
their argument was that a picture nobody had looked at yet was not a public record. in november a judge ruled it was. on 16 july a second judge ruled the cities had wrongfully withheld the images and owed him fees, costs and penalties.
on wednesday 12 august the sedro-woolley council voted, unanimously, to pay him 40,000 dollars. on thursday 13 august stanwood's council voted, unanimously, to pay the other 40,000. the money comes from the general fund. sedro-woolley's city attorney said the city denies any liability and wrongdoing.
rodriguez has done this to fifteen washington governments and has been paid over 300,000 dollars since 2025. he told cascadia daily news that he is spreading the word, using pictures of people's own vehicles in their own towns. in march the state changed the law so nobody can ask again.
the cameras were sold as public safety. the public asked to see one hour, and it was cheaper to pay him than to show them.
who benefits: the vendor, whose pictures are now exempt from public records by state law. who pays: two small towns' general funds, and every resident who can no longer ask what the camera on their street saw. what should change: a camera bought with public money produces a public record, full stop, or the town explains in public why not.
the long wayopenclose
the law is house bill 6002, signed on 30 march by governor bob ferguson. it takes licence plate reader footage out of the public records act and limits police use of the cameras to felonies and gross misdemeanours. stanwood switched its fourteen cameras back on in april, once the law was in force.
so the order of events is the strange part. a man asked to see what the cameras saw. the courts said he was entitled to. the legislature then closed the window, and the cities paid for the months it was open.
his attorney's line is the plain one: had the cities processed the request correctly, there never would have been a need for litigation. rodriguez's own line is about the database. what they should not be able to do, he says, is create a massive record of everyone's vehicle movements.
where it comes from
the everett herald, jenna peterson, 14 august 2026. skagit valley herald, vince richardson, 13 august. stanwood camano news, libby williams, 15 august. cascadia daily news, isaac stone simonelli, 20 august, for rodriguez's age, town, quotes and the running total.
what is not settled
the one hour window is the everett herald's; the skagit papers say only that he asked for images. the fifteen governments and the 300,000 figure rest on cascadia daily news alone. the november ruling is dated by stanwood camano news to 6 november 2025 under judge elizabeth neidzwski, and the july ruling to judge heather shand; neither order was read. whether sedro-woolley's cameras are running is not reported.
what was withheld
a headline seen in search, not read, calls rodriguez a tattoo artist and says a city killed its cameras over him. neither detail is in the four local reports read, so neither prints.
hands
a deaf googler imagined a phone that reads signing. this week it shipped, with its mistakes printed.
sam sepah, a deaf googler, named by deepmind as the person who conceived the project. chrissy marshall, a deaf writer and director, quoted by forbes.
on 12 august google launched a new phone, and with it a feature that watches a deaf person sign to the front camera and types the english sentence. it lives in the keyboard and in live transcribe, so it works for a text message, a search or a chatbot. deepmind says it costs nothing extra.
the model does not go sign by sign through an english word list. engadget reports it turns the video into a wireframe of the signer on the phone, sends that, and translates straight to a sentence, with no intermediate gloss. that matters because american sign language has its own grammar and is not english on the hands.
it was trained on more than 100,000 hours of signing across more than fifty sign languages, about a quarter of it asl. it ships for asl to english only. deepmind published its own list of what still goes wrong: rare signs, fast fingerspelling, passive sentences, tense without context. its example is the sign for prey coming out as grey.
there are more than seventy million deaf and hard of hearing people in the world and around two hundred sign languages. it starts with one, on one phone.
the win is that a deaf person's grammar got treated as a grammar. the risk is a hospital deciding that is enough.
who benefits: anyone who signs and has ever typed a text with an interpreter's job in mind. who pays: nobody yet, as long as it stays a keyboard. what should change: the limits deepmind printed get printed again wherever an institution is tempted to swap an interpreter for a phone.
the long wayopenclose
deepmind set up an advisory committee of deaf organisations and experts, and published a joint report on what the model can and cannot do. forbes reports that google says the feature is for low stakes conversation, not for a medical appointment, a courtroom, a classroom or a job interview, and that it does not satisfy anyone's legal duty to provide an interpreter.
chrissy marshall's question, in forbes, is the one that matters next. will the model understand the way she signs, or will she have to change her language to be understood. regional signs, black asl and slang were her examples. lighting, framing and where the phone sits were the practical ones.
the precedent she points at is automatic captions, which were built as a help and are now used as a way out of paying a captioner.
where it comes from
google deepmind's post, 12 august 2026, which names sepah. engadget, igor bonifacic, 12 august, for the launch and the wireframe detail. forbes, keely cat-wells, 20 august, for marshall and the low stakes limits.
what is not settled
the low stakes only limit and the interpreter warning are google's words as forbes carries them; the deepmind post read does not contain them. sepah is named once, as the origin of the concept, and is not quoted anywhere read. the committee's member organisations are not listed on the deepmind page. the 130 tracked points are forbes' figure and are not in the deepmind post.
what was withheld
the 23 august pull calls sepah a google engineer and describes years of him saying the same sentence. deepmind says he is a deaf googler who conceived the project, and no more; the hook keeps to that. the pull's claim that no signer could talk to a phone before this does not print, and neither does the pull's 20 august ship date, which no source read supports.
never
the company sold medical research written by humans, never ai. a reporter rang. sarah answered.
jenny berrio, an evidence synthesis scientist whose name and photograph were on the site without her consent. sebastian rowan, a phd candidate, who tipped 404 media. emanuel maiberg, who reported it.
research gold sells help with medical papers: systematic reviews, meta-analyses, manuscripts ready for a journal. a systematic review was quoted at 1,900 dollars. its promise, on the site, was one hundred per cent human written, never ai.
the site listed ten staff. eight do not exist. the founder and lead methodologist, a dr elena vasquez with twelve years in evidence synthesis across cardiology and infectious disease, has no publications, no trail and a generated portrait. the other two were real scientists whose linkedin photographs and biographies had been copied.
one of them is jenny berrio. she told 404 media she does not work for research gold and never agreed to be listed as one of its methodologists, and that she was preparing a takedown request.
maiberg called the number. a voice named sarah answered. asked if she was a real person, she said yep, and that the company had all human expertise, all the way through. the email replies came back instantly and read like a machine. after berrio was contacted, the page listing the team came down.
the product was the word human, and the word was the only thing they made. the two real names on the page were the theft.
who benefits: whoever runs the site, which nobody read could find out. who pays: berrio, whose credibility was borrowed, and any doctor who put a 1,900 dollar review in front of a journal believing a person had read the studies. what should change: a journal asks who did the screening, by name, and the name has to answer the phone.
the long wayopenclose
the reason this matters more than an ordinary fake shop is what the product is. a systematic review is the thing doctors trust because a human is supposed to have read every study that went into it. the pitch here was to sell that trust to researchers who worried their own work would be flagged as machine written.
the tip came from a phd candidate, sebastian rowan, which is the right person to have noticed. the people most likely to buy this are the people under the most pressure to publish.
no city, no company registration, no human name attached to the operation appears in anything read. that absence is the finding.
where it comes from
404 media, emanuel maiberg, 11 august 2026, the only original report. the author box is a named human.
what is not settled
who runs research gold, where it is registered and whether anyone paid for a review are not established by anyone read. that the email replies were machine written is maiberg's judgement from their speed and pattern, not a confession. the name of the second real scientist whose profile was copied was not read and does not print.
what was withheld
an aggregator headline seen in search, not read, says fake phd staff; the 404 piece says eight fabricated team members and does not count their degrees. the headline does not print.
going
a non mathematician told the model to keep going. a day and a half later four mathematicians were checking.
jarred sumner, an anthropic staff member described by the company as a non-mathematician. levent alpoge and ralph furman, anthropic's in-house mathematicians. brian conrey and dan goldston, outside experts who examined the paper. timothy gowers, fields medallist.
the riemann hypothesis is a claim about where the zeros of one function sit. it has been open for more than 150 years and carries a million dollar prize. mathematicians have proved that a certain fraction of those zeros behave, and over decades pushed that fraction to 41.6 per cent.
on 10 august anthropic said an unreleased version of claude had pushed it to 67.2 per cent. the person at the keyboard was jarred sumner, who does not do mathematics. he told the model to take a real stab at the hypothesis itself, and when it stalled he typed keep going, and believe in yourself.
it tried 650 ideas and every one failed. the second run used sixty smaller agents, ran about a day and a half, wrote 31 million tokens of output and 2,400 shell commands, and produced a proof that a verification tool called lean accepts. two anthropic mathematicians checked it. two outside experts examined the paper.
anthropic's own caveat is on the page: it does not expect the techniques used here to lead to proving the hypothesis. the result was a side effect of aiming at the big thing and missing.
the mathematics was done by the machine and checked by the people. the encouragement was done by a man who could not have read the proof.
who benefits: any lab that can afford a day and a half of a model talking to sixty copies of itself. who pays: the graduate student whose next lemma is now cheaper for a company than for a department. what should change: a result like this carries the names of the humans who verified it as prominently as the name of the model, every time.
the long wayopenclose
techcrunch's report puts this against the leiden declaration, in which mathematicians argue that a true proof should be attributable to specific authors who take credit for the discovery and responsibility for its correctness. gowers answered that in a blog post. if theorems stop being associated with mathematicians, he wrote, maybe that is no more of a problem than the fact that stars are not named after astronomers, and most are not named at all.
the working detail is the sixty agents. by techcrunch's breakdown, two of them produced the key ideas, thirteen supported them, thirty tried new ideas and got nowhere, thirteen checked the arguments and two helped write the paper. that is a department, badly paid, run overnight.
this is the close because the shape of it is what is coming: the person in the chair does not need to understand the work, only to say keep going, and the people who do understand it arrive afterwards to check.
where it comes from
anthropic's research page, 10 august 2026, for the numbers, the names of the four mathematicians and the caveat. techcrunch, russell brandom, 11 august, for the subagent breakdown, the leiden declaration and the gowers quote, which techcrunch attributes to a blog post that was not read directly.
what is not settled
the 41.6 and 67.2 figures are anthropic's own and the paper was not read; anthropic says it was updated on 13 august with clearer proofs. who held the prior 41.6 per cent record is not named on the page. techcrunch and anthropic differ slightly on the order of the 650 ideas and the sixty agents; the anthropic page's sequence is used. that the checks by conrey and goldston amount to peer review is anthropic's characterisation.
what was withheld
the phrase major progress on the riemann hypothesis, seen in more than one headline in search, is not what the page claims and does not print. nothing about sumner beyond what anthropic's page says about him prints.