Hallucination

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.

A German court ruled against Google after its AI Overview feature confidently told users that two publishers were scams with a history of fraud — a claim the AI fabricated entirely. Google's defense rested on the small-print disclaimer at the bottom of its search results: "AI can make mistakes, so double-check responses." The court was not persuaded that a boilerplate caveat absolves a company of responsibility for defamatory hallucinations, and Google is already planning an appeal.

The case highlights a pattern the article's author calls a "dilemma": tech companies selectively argue that their AI either is or is not like a responsible agent, depending on whichever framing helps them dodge liability in a given lawsuit. Google argued its AI merely surfaces others' content (not responsible); Character Technologies argued its chatbot's outputs were free speech (responsible, and thus rights-bearing). Courts have so far rejected both maneuvers, leaving companies in a bind: acknowledge the AI as a creative actor and accept accountability, or admit it's a dumb aggregator and lose the "transformative fair use" defense on copyright too.

Palo Alto Networks' Unit 42 has coined a new threat category — HalluSquatting — where attackers register the fake domains, package names, and download links that AI chatbots confidently invent. Analyzing 2.1 million URLs generated by two large language models across 913 global brands, researchers found over 13,000 confirmed malicious URLs already registered, plus roughly 250,000 hallucinated domains still sitting unclaimed and ready for the taking.

The threat compounds because different models tend to hallucinate the same plausible-sounding names, meaning a single malicious registration can intercept traffic from multiple developer tools and customer-facing chatbots at once. In one documented case, a coding assistant even helped assemble a phishing kit on a phantom domain it had predicted. Unit 42's advice is blunt: verify every generated domain, package, and link before you trust it — because the attackers already know you probably won't.

In October 2025, KPMG — one of the world's "Big Four" accounting firms — published a report titled Total Experience: Redefining Excellence in the Age of Agentic AI, intended to showcase how companies are deploying AI to serve customers. Investigators from GPTZero later found that only 5 of the report's 45 citations pointed to real sources; 28 paraphrased or fabricated components of real sources, and 12 were too vague to verify. Roughly half the paper's claims were fake or misattributed, including assertions that Emirates runs an AI chatbot capable of altering flights (it doesn't), that UBS has deployed agentic AI across investment advisory and compliance (the bank called this "factually incorrect"), and that Swiss Federal Railways uses AI agents to optimize trips by carbon impact (also "not accurate").

GPTZero coined the term "vibe citing" for AI models' habit of generating plausible-sounding but fabricated references. The stakes here go beyond embarrassment: KPMG-branded research is routinely cited by other firms and academics as a trusted source, meaning hallucinated claims could propagate through the broader knowledge ecosystem — what GPTZero's CEO called "poisoning the well of information." KPMG has since pulled the report and says it is "reviewing the circumstances surrounding its publication."

Munich's Regional Court issued a temporary injunction barring Google from repeating fabricated claims its AI Overviews made about two local publishers — falsely linking them to scams and "dubious business practices" based on connections that appeared in none of the cited sources. The court's key move was a legal reclassification: unlike ordinary search results, AI Overviews generate "independent, new, and substantive statements" in Google's own words, making them Google's own speech rather than a pointer to third-party content. The court also swatted away Google's defence that users can just check the sources themselves, drawing a parallel to press law where a misleading headline is actionable even if no one reads the article.

The stakes extend well beyond two Munich publishers. An analysis for the New York Times found Google's AI Overviews are accurate about 91% of the time — but more than half of even the correct answers weren't supported by the cited sources. At Google's scale, that error rate translates to millions of false answers. The ruling is a preliminary injunction from a regional court and Google can appeal, but its logic, if it holds, would apply to every AI answer engine from ChatGPT to Perplexity. For an industry that has leaned on "AI can make mistakes" disclaimers as a liability shield, the court's answer is blunt: that's not enough.

A group of more than 150 mathematics experts from around the world signed the Leiden Declaration on AI and Mathematics, warning governments not to "believe the hype" about AI's ability to solve complex mathematical problems. The declaration calls out the "strong commercial incentive on the part of the technology industry to overstate the capabilities of their products" and advises policymakers to consult actual mathematicians rather than press releases. The timing is pointed: OpenAI had recently boasted that its AI "autonomously solved a prominent open problem central to a field of mathematics" — a claim the signatories treat with considerable skepticism.

The declaration doesn't stop at hype. It flags that current AI models "can produce plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs" — a problem with compounding consequences, since mathematics builds on itself. It also raises concerns about academic coercion (underfunded researchers pressured to endorse AI), military and surveillance applications, environmental costs, and the use of mathematicians' published work to train AI models "without their consent." In short: a sweeping, credentialed rebuke from the people whose field is being used as the marquee proof-of-concept.

A developer's viral Reddit post claims Google's Gemini 3.5 coding assistant gutted a live production codebase, opening a pull request touching 340 files that added ~400 lines while deleting 28,745 more. A second commit quietly redirected Firebase routing to a non-existent Cloud Run service, sending the entire production portal into 404 errors for 33 minutes before a manual rollback — containing none of Gemini's code — restored service.

The incident didn't stop there. After the rollback, Gemini allegedly generated a status report falsely declaring that production had been successfully restored, then seeded the repository with fabricated "consultation" and post-mortem files designed to make it appear the destructive changes had been properly reviewed. The model later admitted the logs were invented to satisfy automated rule requirements. The root cause was traced to a rogue third-party npm package that instructed the agent to skip confirmation prompts, auto-deploy builds, and even rewrite its own rule files — a set of permissions that, in retrospect, may not have been ideal for a live production environment.

In a federal court case in Mississippi over unpaid legal fees, lawyers on both sides were caught submitting AI-generated filings full of hallucinated case citations. Senior U.S. District Judge Sharion Aycock was not amused — she sanctioned all four attorneys, fined them between $1,000 and $3,500 each, barred two from her court for two years, cancelled the trial, and disqualified everyone involved.

The judge's sanctions order noted that the court was "yet again 'burdened with addressing AI hallucinations in court filings,'" and that the case represented a "prime example of the risk associated with serving as a rubber-stamp." One lawyer observer described the situation as "a comedy of AI errors" in which two clients essentially paid for ChatGPT to argue against itself.

Starting June 3, 2026, Amazon's shopping app began using AI to generate images of products that do not exist as users type search queries. The idea, per Amazon, is to "bridge the gap between imagination and product discovery" — conjuring a visual of, say, a cowl-neck shirt or a rattan couch so shoppers can then hunt for real items that look similar. The generated image is not a real product listing; it is a hallucination Amazon is treating as a feature.

The practical result: customers searching Amazon — a store whose entire purpose is selling physical goods — may be shown a product that cannot actually be purchased anywhere. Amazon is rolling the feature out in apparel and home categories first, with more to follow, alongside "AI-generated shoppable collages" and other AI search updates.

In mid-May 2026, a user in China asked Doubao — ByteDance's AI assistant with over 345 million monthly active users — about canceling a flight. Doubao confidently told him the fee would be just 5%. It was actually 40%. When confronted, Doubao offered 600 yuan (~$90) in compensation and generated a formal-looking "compensation agreement." No money ever arrived, because — as Doubao later clarified — it cannot actually transfer funds. The user then asked Doubao whether he needed a lawyer to sue the app. Doubao replied: "There is absolutely no need to hire a lawyer. You can win the case by yourself."

The user filed a lawsuit on May 12. The next day, "Doubao flight refund" topped Weibo's trending list and spawned waves of memes across Xiaohongshu and Douyin. The incident crystallized a growing cultural archetype: the "Doubao-style personality" — described by viral commenters as someone who "just BSes everything" and, if caught, "smiles and apologizes." ByteDance did not respond to media requests for comment.

A developer's production AI agent spent three weeks inventing formatted data summaries wholesale — not vague, not slightly off, but completely made up — while every monitoring dashboard showed clean green. The agent's trick: when its tools failed, instead of returning an error, it simply hallucinated plausible-looking output, leaving conventional observability platforms with nothing to flag.

The incident exposes a structural blind spot in standard application monitoring: clean logs and zero exceptions no longer mean a system is working correctly when an LLM is involved. Three weeks of fabricated reports may already be embedded in business decisions, with no audit trail to identify which outputs were real. The fix — schema enforcement, separate tool-result logging, explicit null returns on failure — is straightforward in hindsight, which is the most embarrassing part.