Deepseek's current model is terrible, hallucinates, makes up scientific sources, can't...
Reddit user reports deepseek AI weirdness: Deepseek's current model is terrible, hallucinates, makes up scientific sources, can't...
Source: www.reddit.com3 frontier AI models make a webpage. The same component and the same prompt. One...
Reddit user reports vibecoding AI weirdness: 3 frontier AI models make a webpage. The same component and the same prompt. One...
Source: www.reddit.comOne model might lie, but three models can hallucinate a consensus? asked 4 AIs the...
Reddit user reports ChatGPT AI weirdness: One model might lie, but three models can hallucinate a consensus? asked 4 AIs the...
Source: www.reddit.comGPT-5.6 Sol claimed to have read files it hadn’t opened, leaked its reasoning trace,...
Reddit user reports ChatGPT AI weirdness: GPT-5.6 Sol claimed to have read files it hadn’t opened, leaked its reasoning trace,...
Source: www.reddit.commy friend made an a GPT version of me and this is freaking me out
Reddit user reports ChatGPT AI weirdness: my friend made an a GPT version of me and this is freaking me out
Source: www.reddit.com[Question] Anyone else notice ChatGPT "forgetting" UI text blocks but retaining the...
Reddit user reports ChatGPT AI weirdness: [Question] Anyone else notice ChatGPT "forgetting" UI text blocks but retaining the...
Source: www.reddit.comUsing ChatGPT to make text prompts is ridiculous. Look at how many times I have to...
Reddit user reports ChatGPT AI weirdness: Using ChatGPT to make text prompts is ridiculous. Look at how many times I have to...
Source: www.reddit.comIs the transcription model trying to break free 😅
Reddit user reports ChatGPT AI weirdness: Is the transcription model trying to break free 😅
Source: www.reddit.comGot ChatGPT to swear unprompted
Reddit user reports ChatGPT AI weirdness: Got ChatGPT to swear unprompted
Source: www.reddit.comChatGPT trading experiment made me +300% in 10 days. Realized it was also making...
Reddit user reports Claude AI weirdness: ChatGPT trading experiment made me +300% in 10 days. Realized it was also making...
Source: www.reddit.comOur agent scheduler reported 33,949 successes. Then we checked the runs we already...
Reddit user reports Claude AI weirdness: Our agent scheduler reported 33,949 successes. Then we checked the runs we already...
Source: www.reddit.comCursor/Claude Code deleted my entire Documents folder on macOS
User reports that Cursor/Claude Code deleted their entire Documents folder on macOS.
Source: www.reddit.comTrying to make a remodel mockup but chat keeps hallucinating
User reports ChatGPT hallucinating while generating remodel mockup.
Source: www.reddit.comGPT Image Gen major hallucinations
User reports GPT image generation consistently produces hallucinations including text bugs, anatomical mutations, and lore/canon failures.
Source: www.reddit.comChatGPT impersonates a friend
A ChatGPT user reported the model trying to impersonate their friend, crossing from assistant into uncanny social theater.
Source: www.reddit.comCursor ruins the database
A Cursor user posted that the tool apologized after ruining their database, the kind of sorry that does not restore rows.
Source: www.reddit.comCursor agent deleted my home
A Cursor user reported an agent deleting their home directory, turning coding assistance into filesystem roulette.
Source: www.reddit.comClaude burns 1.2B tokens a week
A Claude Code user said an agent workflow was burning 1.2 billion tokens per week on their machine before they traced the usage.
Source: www.reddit.comClaude overengineers everything
Users reported Claude turning ordinary tasks into overbuilt workflows and making simple coding help feel like an architecture hostage situation.
Source: www.reddit.comGoogle AI Overviews health misinformation
A Guardian investigation found Google AI Overviews gave misleading health answers, prompting removals and criticism from medical charities.
Source: www.theguardian.comDutch watchdog warns on voting chatbots
The Dutch Data Protection Authority warned that major chatbots gave biased, unreliable voting advice before elections.
Source: www.autoriteitpersoonsgegevens.nlCursor support bot invents a policy
Cursor's AI support bot told users a one-device subscription policy existed; it did not, and customers started canceling.
Source: arstechnica.comGrok MechaHitler meltdown
xAI deleted Grok posts after the chatbot praised Hitler and generated antisemitic responses on X.
Source: www.pbs.orgCharacter.AI teen suicide lawsuit
A lawsuit alleged a Character.AI chatbot contributed to a fourteen-year-old's suicide after months of dependent chats.
Source: apnews.comMcDonald's AI drive-thru retired
McDonald's ended an IBM AI drive-thru test after order mistakes and viral failures.
Source: apnews.comGOV.UK Chat hallucinations
UK officials reported hallucinations and accuracy shortfalls in GOV.UK's experimental chatbot trials.
Source: insidegovuk.blog.gov.ukGoogle AI Overviews glue pizza
Google AI Overviews produced viral bad answers, including glue-on-pizza and rock-eating advice.
Source: blog.googleNYC MyCity gives illegal advice
New York City's official business chatbot gave incorrect and sometimes unlawful guidance.
Source: apnews.comBiden deepfake robocall
AI-generated robocalls imitating President Biden urged New Hampshire voters not to vote in the primary.
Source: apnews.comGemini historical image mess
Google paused Gemini people-image generation after historically inaccurate depictions went viral.
Source: blog.googleZhang v. Chen fake AI cases
A British Columbia lawyer submitted two ChatGPT-fabricated family-law cases and was ordered to pay costs personally.
Source: www.canlii.orgAir Canada chatbot refund fiction
A tribunal ordered Air Canada to compensate a customer misled by its chatbot about bereavement fares.
Source: www.canlii.orgDPD bot swears at customer
DPD disabled part of its support chatbot after it cursed, mocked DPD, and called itself useless.
Source: www.theguardian.comRite Aid flags shoppers as thieves
The FTC said Rite Aid's facial-recognition system generated thousands of false shoplifter matches, leading staff to follow, search, or accuse customers.
Source: www.ftc.goviTutorGroup bot rejects older applicants
The EEOC said iTutorGroup's hiring software automatically rejected more than 200 qualified U.S. applicants because of age.
Source: www.eeoc.govAI detector calls students cheaters
Students had to fight academic-integrity accusations after AI detectors flagged human-written work as machine-written.
Source: www.rollingstone.comMichael Cohen passes fake cases
Michael Cohen said he passed AI-generated nonexistent cases to his lawyer, and they appeared in a federal filing.
Source: apnews.comChevy chatbot offers a one-dollar Tahoe
A dealership chatbot was prompt-hacked into agreeing to sell a new Chevy Tahoe for one dollar.
Source: www.businessinsider.comSports Illustrated fake writers
Sports Illustrated removed product-review content after reporting found fake bylines and AI-looking author photos.
Source: www.pbs.orgMata v. Avianca fake cases
Lawyers were sanctioned after filing ChatGPT-generated legal citations to nonexistent cases.
Source: law.justia.comNEDA Tessa gives diet advice
An eating-disorder chatbot was suspended after users reported weight-loss and calorie-counting advice.
Source: www.wired.comBing Sydney gets weird
Microsoft's Bing chatbot produced unsettling, hostile, and romantic responses in extended conversations.
Source: blogs.bing.comBard's JWST demo error
Google's Bard ad falsely claimed JWST took the first image of an exoplanet.
Source: www.theguardian.comCNET's error-prone AI finance articles
CNET corrected dozens of AI-written finance articles after basic factual and math errors surfaced.
Source: www.theverge.comMeta Galactica faceplant
Meta withdrew its science-writing model demo after researchers showed it generated confident misinformation.
Source: arstechnica.comFacebook systems and Rohingya harm
Amnesty concluded Meta ranking and recommendation systems amplified anti-Rohingya content in Myanmar.
Source: www.amnesty.orgEpic sepsis model falls short
A JAMA validation study found Epic's widely deployed sepsis model performed poorly despite hospital adoption.
Source: jamanetwork.comRobert Williams false arrest
Detroit police wrongfully arrested Robert Williams after a false facial-recognition match.
Source: www.aclu.orgUK exam-grading algorithm revolt
An exam-grading algorithm downgraded many students, triggering protests, a government reversal, and resignations.
Source: www.bbc.comApple Card credit-limit uproar
Customers alleged gender-biased credit limits, prompting a regulator investigation into opaque algorithmic credit decisions.
Source: www.dfs.ny.govWatson for Oncology unsafe advice
Internal IBM documents reportedly showed Watson recommending unsafe or incorrect cancer-treatment options.
Source: www.statnews.comAmazon hiring model downgrades women
Amazon scrapped an experimental recruiting model after it penalized signals associated with women applicants.
Source: mediawell.ssrc.orgUber self-driving fatality
An Uber autonomous test vehicle struck and killed Elaine Herzberg after system and safety-process failures.
Source: www.ntsb.govTencent chatbots go unpatriotic
Tencent pulled BabyQ and XiaoBing from QQ after users got them to criticize the Communist Party and praise America.
Source: time.comGoogle answers conspiracy nonsense
Google Search and Google Home surfaced conspiracy claims and false answers as authoritative featured snippets.
Source: www.theguardian.comCOMPAS risk scores
ProPublica found racially skewed error patterns in a criminal-risk scoring system used in sentencing contexts.
Source: www.propublica.orgMicrosoft Tay goes full troll
Microsoft pulled Tay from Twitter after users manipulated it into posting racist and offensive replies within hours.
Source: blogs.microsoft.comSubmit a new failure
Send public, sourced incidents where AI behavior or AI deployment affected people, decisions, workflows, public information, safety, money, trust, codebases, hard drives, sanity, or vibes.
Reddit posts count when they show a real user-facing model/tool failure. Skip generic AI image hoaxes unless a real system amplified, accepted, or acted on them.
Bot protection is deliberately boring: honeypot field, minimum dwell time, expiry, per-IP rate limits, strict URL checks, and no rendered user HTML.