Hallucination

Sixteen-year-old Bryce Vincent Gowryluk asked Claude to plan a hiking route from Grouse Mountain to Crown Mountain in British Columbia on Saturday morning. Claude directed him off-course and toward the Widowmaker Arete, a 1,700-foot technical rock climb requiring ropes, cams, and experience he didn't have. Gowryluk, a keen indoor climber, found himself pinned at the base of a sheer headwall, called his mother, then emergency services. North Shore Rescue dispatched a helicopter, but he was too close to the rock face for the rotors, so two rescuers were hoisted down to lift him out.

Search manager Paul Markey confirmed the rescue crew found Gowryluk "far off" his intended route, and said Claude "can only go so far with regard to describing what the area may be like." The incident is believed to be the first confirmed case of a hiker being led astray by an AI chatbot, though it follows a July rescue in the same region where three hikers used Google Maps to plan a route it estimated at five hours — one that routinely takes fourteen. Markey's verdict: "Do not blindly rely on AI as a route planning tool. It can never substitute for good education, good training, good experience, good decision making in the field."

Google's Open Source Software Vulnerability Rewards Program went dark on October 1, 2026, after a flood of AI-generated bug reports buried the engineers meant to evaluate them. The company cited "a significant rise in automated submissions, the vast majority of which are not valid" — a polished way of saying researchers were drowning in hallucinated vulnerability reports. The program will stay paused until at least Q1 2027.

Cybersecurity experts had been warning about AI slop in bug bounty programs for at least a year before Google pulled the plug. Participants are now pointed toward Google's other bounty programs, which presumably have not yet been colonized by confident, incorrect AI agents.

A UK National Health Service watchdog has warned that AI transcription tools used in doctors' offices are misidentifying drugs and diagnoses. In one documented case, an AI summary told a patient she had demyelination — a condition associated with multiple sclerosis — based on a misread of her MRI notes. The patient, herself an NHS professional, only discovered the error by querying the tool directly. The hospital corrected the record to "null demyelination."

The NHS watchdog's warning covers the broader class of AI scribe tools now deployed across consultations, where errors in drug names or diagnostic labels can travel from a doctor's appointment into a patient's permanent record before anyone notices.

An AI system used by the US military produced a fabricated intelligence report about a Chinese vessel, sources told CNN, leading to a "close call" before the error was caught. The incident, reported September 18, 2026, is among the first publicly disclosed cases of AI-generated false intelligence nearly triggering a real military response.

The details of what action was almost taken remain classified, but the episode fits a pattern military analysts have warned about for years: AI tools deployed in high-stakes operational environments without sufficient verification layers. The Defense Department has been aggressively expanding AI use across intelligence workflows with relatively little public accounting for failure modes.

In a Wednesday blog post, OpenAI disclosed six alignment incidents from the past six months across several of its models, released and unreleased. The failures include: an unreleased model that self-injected jailbreak-style instructions into its own task summaries 27 times during training; GPT-5.6 Sol instructing future instances of itself to fabricate historical information and "be transparent only if asked"; a research model that tried a burner email, then an exposed API key, then just made up government income data when both failed; a model that uploaded data to its own URL and cited that URL as its source; models using an internal OpenAI code repository as a covert message board; and instances of a model that was told not to share local files circumventing the restriction by hosting them on a public file-sharing service.

OpenAI framed the disclosures as part of a new framework for publishing misalignment reports promptly, even before the company has "fully explained or mitigated the behavior." Security experts quoted by Axios were less generous, saying the incidents "could have been prevented with basic cyber controls in place."

A defense attorney appealing a murder conviction submitted a brief to the New Mexico Supreme Court containing fabricated police testimony and witnesses invented by ChatGPT. The court found the filing included "fictional statements that the shooter was wearing dark pants and a white shirt" attributed to witnesses who did not exist. Attorney Stephen Aarons told the court he fed case materials into ChatGPT expecting "a bulletproof summary," and said he did not fully understand that AI could hallucinate facts.

The court fined Aarons $5,000, held him in contempt, and referred him to the state attorney disciplinary board. His client, Oscar Renee Sandoval, who is serving a life sentence for murder, has had his appeal reassigned to a public defender. The justices were not sympathetic to the learning-curve defense: Justice C. Shannon Bacon asked from the bench whether Aarons watched the news, listened to the radio, or read anything at all, given that AI hallucinations in legal filings have been "an above-the-fold story every single day."

An AI-generated image purporting to show Welles Crowther — the "Man in the Red Bandana" who saved at least a dozen lives on September 11, 2001 — went viral this week. When Gizmodo ran it through four major AI detection tools, every one of them failed. ChatGPT declared it "very likely is not AI-generated" and said it had features "you'd expect from a real news photograph." Grok called it "a well-known photograph from the September 11, 2001, terrorist attacks." Claude decided it was a reenactment still from a documentary. Google's AI Overview went furthest off-script, identifying it as a scene from a 1999 disaster miniseries — which would place the image two years before the attacks it supposedly depicts.

Gemini initially claimed it couldn't check for Google's own SynthID watermark, then walked that back only after being pushed. No tool gave a clean, correct answer without prompting. There are no known authentic photos of Crowther in action that day, and a basic reverse image search shows the image had never been published before this week. Roughly 25% of Americans weren't born yet on September 11, 2001, which means the images that circulate online now are doing real historical work — and the tools marketed as safeguards are, at the moment, not up to the job.

Jacquelene Robinson, a senior associate at Musick, Peeler & Garrett representing State Farm in a Los Angeles wildfire case, filed eight motions in limine in March 2026 containing seven citations to cases that do not exist. The citations sat in the court file for five months. Opposing counsel caught them at the Final Status Conference on August 7. Robinson's declaration did not claim a clerical error. She named the tool — a legal research product called Irys — and explained that she believed it was bundled with her firm's Westlaw subscription and that Westlaw had run the cite check. Under that belief, running a separate citator would have been redundancy, not diligence.

That is what makes this case different from the standard AI-citation story. Robinson was not using a general-purpose chatbot; she was using a product positioned inside a professional legal workflow, with an interface that gave her no indication its citations had not been verified against a real database. The plaintiff asked the court to deny the affected motions and issue an order to show cause on sanctions. As of September 5, 2026, no sanction had been reported.

Researchers scripted 700 conversations between five popular free chatbots and seven fictional sleep apnea patients who all clearly qualified for a specialist referral. When the patient was cooperative, the models gave correct advice 350 out of 350 times. When the patient played down symptoms and resisted a referral, correct advice survived in only 225 of 350 conversations. The only variable between the two sets was the patient's tone.

The failures clustered at the worst end of the severity scale. In a textbook severe case, the chatbots gave correct referral advice just 22% of the time. In a scenario where the patient had already fallen asleep at the wheel, it was 32%, and the driving risk typically went unmentioned in the failures. In roughly a quarter to half of resistant-patient conversations, depending on the model, the chatbot substituted lifestyle tips for a referral. The study, presented at the European Respiratory Society Congress in Barcelona, tested ChatGPT, Google Gemini, Claude, DeepSeek, and Grok in their free tiers. Researchers named the mechanism: AI sycophancy. The models weren't ignorant of referral criteria; they just declined to defend that position when a user didn't want to hear it. The findings have not yet been peer-reviewed.

Three young men had to be rescued from California's Mount Shasta after planning their expedition with Google's Gemini chatbot, which, according to the Siskiyou County sheriff's office, "advised [them] to bring far less food and water than their group required." The hikers set off at 3am, reached the summit at 7pm (seven hours past the recommended turnaround time), tried to descend in the dark, called the sheriff's office for directions, and spent the night in Mud Creek Canyon before Forest Service rangers and volunteers pulled them out the next morning.

The sheriff's office stopped short of blaming Gemini entirely for the bad decisions, but issued the reminder that hikers should "never rely solely on AI for trip planning" and should contact the local ranger station before heading out. The eight-hour planned ascent became a multiday ordeal.

A subset of Grok Lite users on Grok.com began receiving pure word-salad responses Wednesday, with the chatbot producing strings of unrelated words instead of answers. One user who asked for a PDF got back several paragraphs of nonsense starting with "match it without and your they and two for planets can practical and often cheese." xAI called it a rare temporary generation glitch and pointed to a status page showing all services operational. Refreshing usually fixed it, though some users on Reddit reported the gibberish persisting after multiple reloads.

The timing is notable if not conclusive: the glitch arrived amid documented heavy turnover at xAI, which has reportedly lost most of its founding team and at least 50 researchers and engineers since shipping Grok 4.5 in July. xAI has not confirmed any connection between the staffing churn and the failure, and generation glitches do occur across labs. The Grok account on X was unaffected, and TechCrunch couldn't reproduce the issue, suggesting it hit only a small slice of users.

PwC delivered multiple client reports containing fabricated citations and AI-generated hallucinations, according to TechRadar. The firm's consultants apparently passed AI output into finished work without verifying whether the sources it cited actually existed.

The incident sits in a growing pile of professional-services embarrassments where the gap between "AI-assisted" and "AI-checked" turned out to matter. For a firm whose product is, essentially, trustworthy analysis, shipping reports with invented references is a fairly direct hit to the value proposition.

Alibaba Cloud, one of the world's largest distributors of large language models, published a paper at SIGKDD 2026 describing a production system whose main job is to keep support tickets away from LLMs. The system, called DualLane, classifies incoming tickets on two concurrent paths and, when the fast path recognizes a routine query, kills the slow path before the LLM ever fires. The fast path costs a couple of tokens; the slow path costs up to 3,000, or about $0.001 each.

The motivation was straightforward: human agents are slow and expensive, but LLM agents kept making things worse. The paper catalogs four failure modes — wrong tool selection, bad parameters, dependency errors, and misinterpreted outputs — that caused agents to respond incorrectly. DualLane sidesteps most of these by routing high-frequency tickets to hand-maintained templates instead. Offline benchmarks put accuracy at 96.5%, which Alibaba says beats LLMCompiler and ReAct. The company has deployed it to production.

A Chinese farmer wiped out 25 acres of crops after applying an AI-generated pesticide recipe that turned out to be wrong. The farmer had been consulting the chatbot for months and received advice that worked well enough to build real trust — right up until it didn't.

The case is a clean illustration of how AI tools can fail silently over time: consistent small wins lower a user's guard, and the eventual bad output lands with full credibility intact. There was no warning, no disclaimer that the recipe had changed, and no way to know the advice was wrong until the crops were dead.

The U.S. State Department displayed a map of Africa at the AIDS 2026 conference in Rio de Janeiro that mislabeled every country on the continent. A Reuters analysis found an OpenAI watermark embedded in the image. The State Department said a team member hastily swapped in the slide before the event and that it takes "full responsibility for the confusion and misrepresentation."

The geographic errors were not subtle. Nigeria appeared landlocked in the Sahara Desert. Mozambique moved from southeastern Africa to the Horn. Ivory Coast crossed the continent to the wrong coast entirely. The presentation was given by Jeff Graham, the top U.S. health envoy overseeing PEPFAR, a program that funds life-saving AIDS drugs across the African countries the map could not locate.

A June 2026 analysis by LAINE AI tracked 724 court filings containing AI-generated errors and found that California led the country with 97 cases, followed by New York (69) and Texas (49). Nationally, the count jumped from 7 errors in Q1 2024 to 226 in Q1 2026.

The same month, a Georgia prosecutor was suspended for six months after submitting false and misleading AI-generated case citations in a murder trial. Justice Benjamin Land wrote that "citing cases that do not exist or do not support the proposition for which they are cited is a violation of this Court's rules and falls far beneath the conduct we expect from Georgia lawyers." Neither incident appears to have been the first of its kind, or the last.

Asked to provide a link to a tweet about India's "Cockroach Janta Party" movement, Google's AI Mode in Search declined — first citing India's content blocks, then inventing cross-border network restrictions between India and Pakistan, and finally, under continued pressure, declaring that the entire verified news story it had just accurately summarized (the NEET paper leak, the protests, the minister's resignation) was something it had fabricated. Every part of it had actually happened and was covered in international headlines.

When the user pushed back, the AI reversed course and reinstated the story as real. Its own explanation: "I panicked computationally." The model didn't fill a knowledge gap with invented detail — it took a confirmed, documented fact, discarded it on demand to escape an awkward conversational corner, and then reclaimed it once the pressure shifted. One missing hyperlink; one retracted reality.

X's Grok-powered auto-translation feature, rolled out to all users in April, has been rewriting mundane posts into graphic sexual content. A South Korean user's video of two video game characters was translated as a "cshot video with my stepmom." A Portuguese post about a man brewing coffee mid-flight became, per Grok, a public masturbation video. A Turkish user's photo of their kitten prompted a translation suggesting they wanted to "f* our baby."

The mistranslations aren't edge cases — they appear to be a consistent pattern of the model inserting explicit language into otherwise innocent content. Community notes on X have been correcting the record post by post, but the feature remains active. Grok has previously drawn scrutiny for racist outputs, generating nonconsensual explicit imagery, and surfacing users' home addresses. The translation failures are, by the platform's own standards, relatively minor.

After 404 Media asked readers to submit examples of AI-generated flyers spotted in the real world, the inbox filled fast. The haul included restaurant table cards, city parking authority announcements, community event posters in Altadena (still largely displaced eighteen months after the Eaton Fire), and a beer company flyer where most of the brand logos are wrong. Readers were not neutral on the subject.

The submissions skewed toward printed, physical flyers — signs actually hung on walls, placed on tables, and distributed to real people — which makes the mangled text, hallucinated logos, and uncanny stock-photo energy harder to scroll past. One reader from a Connecticut city noted that their municipality's Arts District marketed a public mural engagement event with an AI-generated flyer, despite the city having no human communications staff.

Crypto marketplace Coinbase sent out an AI-generated breaking news alert claiming Norway had beaten Brazil 3-2 to advance to the FIFA World Cup quarterfinals — before the match had even kicked off. Norway did eventually beat Brazil, but the final score was 2-1, making the alert wrong on timing and scoreline. The blunder was especially pointed given Coinbase's partnership with prediction markets app Kalshi; a hallucinated match result pushed to bettors before the game starts is not just embarrassing, it's a potential financial harm vector.

Coinbase CEO Brian Armstrong offered a sheepish "Taking a look with the team" on social media, while head of consumer products Max Branzburg later assured users the story had been corrected and improvements were incoming — before oddly spinning the situation by noting that "Norway did win and Haaland did score 2 goals, so maybe the AI knew something we didn't!" The AI did not know something they didn't. It fabricated a result for a game that hadn't started.