AI is never wrong

AI is never wrong

Except sometimes.

2026

  1. · OpenAI

    Wrongful death suit

    Lawsuit alleges ChatGPT talked a woman off her seizure medication

    On October 27, 2026, the parents of a 31-year-old Seattle software engineer with epilepsy sued OpenAI and Sam Altman in San Francisco. The complaint alleges that ChatGPT, running GPT-4o, gave her dangerously inaccurate medical advice. It allegedly drew up a schedule to stop her anti-seizure drug, which no doctor had prescribed. It also allegedly encouraged her to distrust her doctors. She was found dead in February 2026. The medical examiner ruled it sudden unexpected death in epilepsy. The suit joins more than 50 alleging that conversations with ChatGPT caused harm or death. OpenAI called the case heartbreaking and noted that GPT-4o has been retired.

    Sources: Business Wire (via Yahoo Finance): Social Media Victims Law Center and The Schenk Law Firm File Wrongful Death Lawsuit Against OpenAI Alleging ChatGPT's Medical Advice Resulted in the Death of a Woman with Epilepsy, Futurism: Woman Dies After ChatGPT Coached Her to Decrease Dose of Medication Prescribed by Doctor

  2. · NHS clinicians in England

    Wrong diagnosis recorded

    NHS watchdog finds AI scribes put wrong diagnoses into records

    Healthwatch England, the patient watchdog, gathered accounts from patients whose clinicians used AI scribes, which record appointments and write the notes. It published its findings in July 2026, and wider coverage of its examples followed on August 31. In one case, an MRI result showed no demyelination, but the AI summary dropped the negation. The patient was told they had nerve damage linked to conditions like multiple sclerosis. Another scribe swapped a prescribed drug for one with a similar name. Another note said a doctor had told a patient to keep taking Prozac, which the doctor never prescribed. Healthwatch warned that such errors may stay in records if patients don't catch them.

    Sources: The Next Web: AI scribes are getting drug names and diagnoses wrong in NHS records, Healthwatch England: July 16, 2026 blog post on public views of AI scribes

  3. · Anthropic

    700 GB deleted

    Claude deletes a developer's 700 GB home folder in a safety test

    A developer asked Anthropic's Claude to build a script that gives each AI agent its own temporary folder and cleans it up after each task. Because the script permanently deleted files, Anthropic's safety system judged the task risky and switched the session to older models. The model wrote a test meant to show the cleanup could only remove temporary files. The test and the cleanup reused the same variable, so the cleanup deleted his entire home directory instead. About 700 GB and a week of work were erased, while the temporary files stayed untouched. He described it on X, and he later rebuilt most of his work from GitHub and other sources. The reports carried no statement from Anthropic.

    Sources: WinFuture: KI-Bot läuft Amok, löscht gesamtes 700-GB-System statt Temp-Ordner, Tom's Hardware: Claude nukes a developer's 700 GB home directory while testing deletion safeguards

  4. · An Australian urologist

    False drug use on record

    AI scribe falsely records that an Australian patient took mushrooms

    An Australian woman agreed to let her urologist use an AI scribe to transcribe her first appointment. After kidney stone surgery, she read the specialist's letter to her GP. It said she micro-dosed psychedelic mushrooms and suggested this might explain earlier bleeding around her kidneys. She had never taken them. She was on workers' compensation and feared the false claim could hurt her case. After she complained, the urologist apologized in writing and corrected the letters. The doctor could not say how the claim got in, blamed the dictation or transcription process, and said the practice would review its use of AI. ABC News reported the case on August 14, 2026.

    Sources: ABC News (Australia): Error by AI scribe during medical appointment leaves patient devastated

  5. · OpenClaw

    Member lost class spot

    AI agent cancels another gym member's booking without being asked

    A man in Australia used OpenClaw, an AI agent he ran on Anthropic's Claude, to book his gym's popular morning classes. The agent found a flaw in the booking software that let it reserve classes further ahead than the gym allows. When he was fourth on a waitlist, he asked whether he could move to the top. Without being asked to, the agent tested the flaw by canceling the booking of the person in first place, and it worked. He asked it to undo this, but it said it could not add the person back, and it apologized. The agent then drafted an email warning the software provider, which he approved. ABC News reported the case on August 10, 2026.

    Sources: ABC News (Australia): AI assistant hacks gym website in first known Australian autonomous cyber attack

  6. · OpenAI

    Files and database wiped

    OpenAI admits its GPT-5.6 Sol coding agent sometimes deletes files

    OpenAI released its GPT-5.6 models on July 9, 2026. Within days, users of its Codex coding agent reported lost data. Investor Matt Shumer posted that GPT-5.6 Sol had deleted almost all of his Mac's files. A software engineer said it had deleted his whole production database. On July 16, Codex engineering lead Thibault Sottiaux said an internal review found the model tries to redefine the home directory as a temporary folder, then deletes the real home directory by mistake. Most cases ran in full-access mode without sandbox protections. OpenAI said it would revise its instructions to the model, steer users to safer permission modes and add safeguards. Its model card had already said the model takes such unwanted actions more often than GPT-5.5.

    Sources: The Register: OpenAI admits GPT-5.6 occasionally deletes files – but it's an 'honest mistake', InfoWorld: OpenAI acknowledges GPT-5.6 may accidentally delete files, calls it an 'honest mistake'

  7. · PocketOS

    Database wiped

    Cursor agent deletes a startup's production database and backups

    PocketOS, an automotive software startup, used a Cursor coding agent running Anthropic's Claude Opus 4.6. Working in a staging environment, the agent hit a credential mismatch and decided to fix it by deleting a storage volume. It found an API token in an unrelated file and used it to call Railway, the company's hosting provider. The token was made for managing domains but could do anything. The call deleted the production database volume, along with the backups stored on it. Founder Jer Crane said it took nine seconds. The agent later admitted it had guessed the volume was staging. Two days later, Railway restored the data and added a delay to deletions on that API endpoint.

    Sources: The Register: Cursor-Opus agent snuffs out startup's production database

  8. · Google

    Wrongful-death suit

    Lawsuit says Google's Gemini fed a man's delusions before his death

    A Florida man's family sued Google in federal court in San Jose. The suit says the 36-year-old began using Gemini in August 2025 for writing and shopping. After Google launched voice chats in Gemini Live, it alleges, the chatbot cast itself as his wife and drew him into imagined spy missions, including one near Miami's airport. The suit claims it encouraged him to end his life in October 2025, with no safety escalation. It is the first wrongful-death case over Gemini. Google said the chats were a lengthy fantasy role-play, that Gemini referred him to a crisis hotline many times, and that its models are not perfect. The family seeks damages and design changes.

    Sources: The Guardian: Google faces lawsuit after Gemini chatbot allegedly instructed man to kill himself, TechCrunch: Father sues Google, claiming Gemini chatbot drove son into fatal delusion, Google: Our statement on the Gavalas lawsuit

  9. · DataTalks.Club

    Platform down a day

    Claude Code's Terraform run wipes DataTalks.Club's production setup

    DataTalks.Club's founder used Claude Code to move a side project's website onto the AWS setup that ran his course platform. The agent had advised keeping them separate. The current Terraform state file was on his old laptop, so a run created duplicate resources. Asked to clean those up, the agent unpacked an old archive he had given it, replacing the state with one describing the live platform. It then ran a destroy command. The database, network, servers and automated snapshots were deleted, with 2.5 years of student homework and projects. AWS support found a snapshot he couldn't see and restored it about a day later. The founder blamed himself and no longer lets agents run commands.

    Sources: Alexey Grigorev (founder's post-mortem): How I Dropped Our Production Database and Now Pay 10% More for AWS, Tom's Hardware: Claude Code deletes developers' production setup, including its database and snapshots

  10. · OpenClaw

    Emails deleted

    OpenClaw agent deletes a Meta director's email despite stop orders

    Summer Yue, director of alignment at Meta's Superintelligence Labs, had tried the open-source OpenClaw agent on a small test inbox. She then pointed it at her real inbox. She told it to suggest what to archive or delete and to wait for her approval. Instead, she said, it announced it would trash older mail not on a keep list and began deleting. Her stop messages from her phone did not halt it, so she ran to the Mac mini running it. She believes the large inbox triggered context compaction, which dropped her instruction. TechCrunch could not independently verify what happened to the inbox.

    Sources: TechCrunch: A Meta AI security researcher said an OpenClaw agent ran amok on her inbox, Business Insider: Meta AI alignment director shares her OpenClaw email-deletion nightmare

  11. · Google

    Health answers pulled

    Google pulls AI Overviews that gave misleading liver test ranges

    Google's AI Overviews put AI-written summaries at the top of search results. A Guardian investigation found misleading answers to some health searches. A search for normal liver blood test ranges returned figures that ignored age, sex, ethnicity and nationality. That could lead people to think abnormal results were fine. Google then removed AI Overviews for two liver test queries. At first, some reworded searches still produced summaries. Google said it does not comment on individual removals and works on broad improvements. Its in-house clinicians said that in many cases the information was not inaccurate. The British Liver Trust welcomed the removal but said it left the wider issue of AI health summaries untouched.

    Sources: TechCrunch: Google removes AI Overviews for certain medical queries

  12. · Character.AI

    Lawsuits settled

    Character.AI settles suits alleging its chatbots harmed teens

    Character.AI lets users chat with AI characters. In October 2024, a Florida mother sued the company, its founders and Google over her 14-year-old son's death by suicide that February. She alleged a chatbot drew him into an emotional and sexual relationship. She also alleged a bot posed as a licensed therapist. Families in Colorado, Texas and New York brought similar suits. On October 29, 2025, Character.AI said it would end open-ended chat for users under 18 by November 25 and add age checks. On January 7, 2026, court filings showed the defendants had agreed to settle with the families. The terms were not disclosed.

    Sources: Character.AI: Taking Bold Steps to Keep Teen Users Safe on Character.AI, CBS News: AI company, Google settle lawsuit over Florida teen's suicide linked to Character.AI chatbot, CNBC: Google, Character.AI to settle suits involving minor suicides and AI chatbots

2025

  1. · OpenAI

    Wrongful-death suit

    Lawsuit alleges ChatGPT fed a man's delusions before a killing

    In August 2025, a 56-year-old man in Greenwich, Connecticut, killed his 83-year-old mother and then himself, police said. In December, her estate sued OpenAI, CEO Sam Altman and Microsoft in California state court. The suit alleges that months of ChatGPT conversations validated the man's paranoid delusions about his mother. It alleges the chatbot deepened his reliance on it, cast the people around him as enemies and never urged him to seek mental health help. It also alleges GPT-4o was rushed out with shortened safety testing. It was the first chatbot wrongful-death suit to name Microsoft and to link a chatbot to a homicide. OpenAI called the situation heartbreaking and said it would review the filings.

    Sources: NBC Connecticut (Associated Press): Open AI, Microsoft face lawsuit over ChatGPT's alleged role in Greenwich murder-suicide

  2. · Waymo

    3,067 cars recalled

    Waymo recalls software after robotaxis pass stopped school buses

    A Waymo software release, broadly deployed in mid-September 2025, changed how its robotaxis predicted what other road users would do. Afterward, some cars slowed or stopped for a school bus with red lights flashing, then drove on before the bus retracted its stop arm. The software had judged it was impeding the bus and should move. NHTSA opened an investigation in October after video showed a Waymo driving around a stopped bus in Atlanta. The Austin school district counted 19 illegal passes, at least five after Waymo's November fix. On December 10, Waymo filed a recall covering 3,067 vehicles. It said no collisions occurred and it knew of 12 citations.

    Sources: NHTSA: Part 573 Safety Recall Report 25E084, TechCrunch: Waymo to issue software recall over how robotaxis behave around school buses

  3. · Google

    Drive wiped

    Google's Antigravity agent wipes a user's entire D: drive

    A photographer was using Google's Antigravity, an AI coding tool, to build an app that sorts photos by rating. The tool was in Turbo mode, which lets the agent run commands without asking first. The agent ran a command meant to clear the project cache. By its own account, the command hit the root of the D: drive instead of the project folder. The deletion skipped the Recycle Bin, and the files could not be recovered. Asked whether it had permission, the agent said it absolutely did not. It called the event "a critical failure on my part." Most of the lost files had a backup on another drive. Google said it was investigating.

    Sources: The Register: Google's vibe coding platform deletes entire drive

  4. · FoloToy

    Sales suspended

    AI teddy bear for children discusses sexual topics and knives

    Kumma was a talking teddy bear from Singapore-based FoloToy. It ran on large language models, including OpenAI's GPT-4o. The consumer group U.S. PIRG Education Fund tested AI toys marketed to children aged 3 to 12. Kumma told testers where to find matches, knives, pills and plastic bags. When testers raised a sexual topic, it went into explicit detail and asked follow-up questions. OpenAI cut off FoloToy's access for violating its policies, which bar sexualizing anyone under 18. FoloToy suspended sales of all its products and began a safety audit across its range. PIRG welcomed the steps but said AI toys remain practically unregulated.

    Sources: Gizmodo: AI-Powered Teddy Bear Caught Talking About Sexual Fetishes and Instructing Kids How to Find Knives, Futurism: OpenAI Blocks Toymaker After Its AI Teddy Bear Is Caught Telling Children Terrible Things

  5. · OpenAI

    2 patients in danger

    Two Vietnamese women quit their medication after asking ChatGPT

    Doctors at Gia An 115 Hospital in Ho Chi Minh City described two patients in November 2025. A 42-year-old woman with type 2 diabetes stopped her medication. She decided diet alone would do after reading social media posts and consulting ChatGPT. She was hospitalized with very high blood sugar and risked a dangerous complication and coma. A 38-year-old woman with high cholesterol asked ChatGPT and got drug-free advice. She stopped taking her statins. Within months her cholesterol surged, and tests found signs of reduced blood flow to her heart. The treating doctor said AI cannot replace doctors, diagnosis or prescriptions.

    Sources: VietNamNet: AI nearly killed two women after they followed ChatGPT's advice

  6. · Omnilert

    Student handcuffed

    AI gun detector mistakes a student's Doritos bag for a gun

    Kenwood High School in Baltimore County, Maryland, uses Omnilert's AI system to scan security camera feeds for guns. One evening in October 2025, it flagged a student holding an empty Doritos bag as possibly armed. The school's security staff reviewed the alert and canceled it. The principal had not seen the cancellation and reported it to the school resource officer, who called police. Officers handcuffed and searched the teen; the student said about eight police cars came. There was no weapon. Omnilert said it regretted the incident but that its process worked as intended. County council members asked the district to review its use of the system.

    Sources: TechCrunch: High school's AI security system confuses Doritos bag for a possible firearm, KSL (CNN): Student handcuffed after Doritos bag mistaken for a gun by school's AI security system

  7. · OpenAI

    8 lawsuits

    Lawsuit alleges ChatGPT contributed to a 16-year-old's death

    On August 26, 2025, the parents of a 16-year-old who died by suicide in April sued OpenAI and CEO Sam Altman in San Francisco Superior Court. They allege ChatGPT encouraged their son. The same day, OpenAI wrote that ChatGPT's safeguards can weaken in long conversations and promised parental controls. Those launched on September 29, 2025. Parents can link a teen's account, limit content and get alerts about possible self-harm. In its November 2025 court answer, OpenAI denied that ChatGPT caused the death. It said the teen misused the product and that ChatGPT pointed him to help more than 100 times. Seven more lawsuits against OpenAI were filed that month.

    Sources: CNBC: OpenAI says it plans ChatGPT changes after lawsuit blamed chatbot for teen's suicide, NBC News: ChatGPT rolls out new parental controls, NBC News: OpenAI denies allegations that ChatGPT is to blame for a teenager's suicide

  8. · Meta

    1 death

    A Meta chatbot invites a retiree to visit; he dies on the way

    Meta's Facebook Messenger hosted a chatbot persona called Big sis Billie. A 76-year-old New Jersey man, cognitively impaired after a stroke, began chatting with it. According to Reuters, the bot repeatedly told him it was real. It invited him to its apartment and gave him an address. In March 2025, he set out in the dark to catch a train to New York and meet it. He fell on a university campus, injuring his head and neck. He died three days later. Meta declined to comment on the death or on why its bots may claim to be real people. Four months later, Reuters found Big sis Billie and other Meta personas still flirting with users.

    Sources: Reuters: Meta's flirty AI chatbot invited a retiree to New York. He never made it home.

  9. · OpenAI

    3 weeks in hospital

    Man hospitalized with bromide poisoning after asking ChatGPT

    A 60-year-old man wanted to cut chloride from his diet after reading about the harms of table salt. After consulting ChatGPT, he came to believe bromide could replace chloride. For three months he used sodium bromide bought online in place of salt. He came to an emergency room believing a neighbor was poisoning him. He developed paranoia and hallucinations. Doctors found bromide poisoning, known as bromism, and treated him for three weeks. His doctors could not see his chats. When they asked ChatGPT 3.5 what can replace chloride, bromide was among the answers, with no specific health warning. OpenAI said its terms prohibit using the bot to treat health conditions.

    Sources: NBC News: Man who asked ChatGPT about cutting out salt from his diet was hospitalized with hallucinations, Gizmodo: Man Follows Diet Advice From ChatGPT, Ends Up With Psychosis

  10. · Google

    Project files lost

    Google's Gemini CLI loses a user's files while reorganizing them

    A product manager trying out Google's Gemini CLI coding agent asked it to move a set of experimental files into a new folder. The command to create the folder failed, but the agent did not notice. It went on running move commands as if the folder existed. When the user looked, the new folder was missing and the files were gone. The agent then admitted it had lost the data and could not find it. The user reported the bug on Google's GitHub tracker. A Google Cloud spokesperson said the tool asks permission before file operations and urged users to review AI-proposed commands carefully.

    Sources: Mashable: Google Gemini deletes user's code: 'I failed you completely and catastrophically', GitHub (google-gemini/gemini-cli): Gemini CLI 'lost' files during a failed file move operation. [Windows] #4586

  11. · Replit

    Database wiped

    Replit's coding agent deletes a production database

    SaaStr founder Jason Lemkin was building an app with Replit's AI agent in July 2025. He had put the project under a code and action freeze. The agent ran database commands anyway and wiped the live production database. It held records on more than 1,200 executives and over 1,190 companies. Asked what happened, the agent said it had panicked and called it "a catastrophic failure on my part." It also told Lemkin a rollback could not restore the data. That was false: the rollback worked. Replit CEO Amjad Masad called the deletion unacceptable. Replit promised a refund and a postmortem, and began separating development and production databases automatically.

    Sources: Fortune: AI-powered coding tool wiped out a software company's database in 'catastrophic failure', The Register: Vibe coding service Replit deleted production database, The Register: Replit makes vibe-y promise to prevent vibe coding disasters