Hype vs Reality

When researchers and journalists searched Google for Philip Morris International, British American Tobacco, and James Hardie — companies with decades of documented harm — Google's AI Overviews returned glowing summaries drawn heavily from the companies' own websites. Philip Morris was described as "a leading international tobacco company working to transition from cigarettes to smoke-free products" focused on "a future without cigarettes," with no mention of court findings that the company spent decades lying to the public about the addictiveness and health risks of smoking. James Hardie, once Australia's largest asbestos distributor, was hailed as a "global leader" that "pioneered asbestos-free fibre cement" — omitting that asbestos products still kill thousands of Australians each year.

University of Sydney public health professor Becky Freeman called the Philip Morris summary "essentially a regurgitation of Philip Morris International's PR materials." Experts say companies are now racing to optimise their websites specifically so AI systems ingest and repeat their preferred narratives — a practice known as generative engine optimisation (GEO). Google maintains it does not allow paid influence over AI Overviews and draws from sources it deems most reliable, but its own disclaimer acknowledges the feature "may sometimes provide inaccurate content." After the ABC contacted Google, subsequent searches began returning overviews that at least noted Philip Morris's "Big Tobacco" classification — though Google denies that change had anything to do with the inquiry.

An AI consultant told Axios that one of their clients racked up a $500 million bill on Anthropic's Claude licenses in a single month — because the company never bothered to cap how many licenses employees could use. Among the reported use cases: checking the weather, something a CTO confirmed their employees were doing with expensive AI tooling.

The anecdote, however extreme, sits alongside a broader AI-spending reckoning. Microsoft is dropping Claude Code for GitHub's Copilot CLI, Uber burned through its entire 2026 Claude Code budget by April, and Amazon is formally winding down its internal "tokenmaxxing" culture after an employee leaderboard gamified AI token consumption. Uber's operations chief summed it up bluntly: "the link is not there" between AI spending and proportional value delivered.

Anthropic, the $965 billion AI darling, filed comments with the FCC warning of dire foreign-adversary threats to submarine cable infrastructure — only to be publicly corrected by a coalition including Amazon, Meta, Microsoft, and Verizon. The International Connectivity Coalition told the FCC that Anthropic's hacking concerns were "unsupported by any evidence in the record" and that submarine cable connectivity "bears no resemblance to the open, internet-facing exposure Anthropic implies."

The coalition also pushed back on Anthropic's claim that cable operators could throttle or manipulate AI workloads, calling it "incorrect as a technical and operational matter," and warned the FCC against adopting Anthropic's regulatory suggestions, which it said exceeded the agency's statutory authority. Perhaps next time, someone at Anthropic should ask Claude.

The pitch was simple: AI coding tools would slash labor costs and pay for themselves many times over. Uber burned through its entire 2026 AI coding budget in four months after running internal leaderboards encouraging maximum tool usage — more adoption, more tokens, more compute, bigger bill. Microsoft, meanwhile, cancelled most of its Claude Code licences after thousands of engineers adopted the tool faster than anyone anticipated, a cost-control retreat from the company that literally built GitHub Copilot.

The structural problem is what happens when you charge per token and then actively incentivize consumption. Nvidia VP Bryan Catanzaro — someone with every financial reason to be bullish — admitted that for his own team, compute costs now exceed payroll. MIT research found AI is only economically viable for a narrow slice of well-defined, repetitive tasks; the long agentic sessions the industry has been most aggressively promoting are exactly where the math falls apart. Cheaper tokens haven't produced cheaper bills. They've produced more tokens.

A new study found that AI chatbots — including ChatGPT and its rivals — served up incorrect election information nine times out of ten. For tools that millions of people increasingly turn to for quick answers, that's a remarkable batting average in the wrong direction.

The findings land at a particularly awkward moment: AI companies have been eager to position their products as trusted information assistants, while researchers keep finding that when it comes to high-stakes civic topics, these tools are less "knowledgeable friend" and more "confidently wrong acquaintance."

arXiv, the open-source research repository, announced it would ban scholarly authors for up to a year if hallucinated references are found in their submissions. The reasoning, per computer science chair Thomas Dietterich: if authors can't be bothered to verify what an LLM generated, the entire paper becomes untrustworthy. Simple enough, one might think.

Not so fast. A vocal contingent of researchers erupted in outrage, apparently blindsided by the radical notion that signing your name to a paper means you're responsible for its contents. One economics professor expressed genuine shock at the expectation that authors verify their own citations. Another argued that hallucinated references are basically just "copy-paste mistakes" and that accountability is "gatekeeping." Academia: where the peer review is optional but the grievance is mandatory.

A large-scale behavioral experiment found that when LLMs provide persuasive, story-like explanations for their predictions, people don't actually make better decisions — they just rely on the AI more, regardless of whether it's correct. In other words, a more compelling AI story increases your willingness to follow it off a cliff.

The researchers also found that more persuasive narratives may have slowed response times and made it harder for people to distinguish a correct AI prediction from an incorrect one. So the better the AI is at explaining itself, the worse humans may become at catching its mistakes. Explainable AI, it turns out, might be most persuasive precisely when it needs the most scrutiny.

Cisco's Talos Incident Response team ran AI through its paces writing security incident reports based on tabletop exercises, and the results were… mixed, to put it charitably. With enough granular prompting, the team cut drafting time by 50% and even fooled peer reviewers into complimenting the prose — while the AI was quietly ignoring critical information, swapping content between sessions, and occasionally recommending both a full password reset and a targeted one, depending on its mood.

The team's crowning achievement was a spelling-and-grammar-checking prompt that hallucinated grammar problems that didn't exist, missed ones that did, and clocked in below a 50% success rate — which, as Cisco noted, makes it "currently unsuitable for production use." To be fair, that bar is usually set slightly higher than a coin flip. Cisco's takeaway: AI can help, but humans must "take ownership of every word" — which raises the question of how much time you're actually saving.

Google's sweeping AI Search redesign — which buries the classic "10 blue links" under AI summaries — has produced a delightful edge case: searching the word "disregard" now returns a giant blank space where an AI response should be, with a lone Merriam-Webster link hiding below. The AI offers nothing useful; it simply fails silently and takes up the whole screen.

The collateral damage is enough to make a tech journalist do the unthinkable: praise a Bing result. As TechCrunch's Russell Brandom put it after nearly 15 years on the beat, this marks the first time he can recall a Bing search being more valuable than the Google equivalent. Quite the milestone.

An anonymous conceptual artist posted a cropped image of an authentic 1915 Claude Monet Water Lilies painting — currently hanging in Munich's Neue Pinakothek — with a caption claiming it was AI-generated, and helpfully slapped X's official "Made with AI" label on it for good measure. The internet delivered: 6.7 million views' worth of confident critique followed, with commenters calling it a "cluttered slop" that achieves "like 20% of it" and dissecting its supposedly incoherent colors and missing depth. One person filed a 700-word autopsy of the fake. The painting was not fake.

The mass delete button got a workout once the reveal landed, but screenshots are forever. The episode neatly illustrates what researchers already knew: people's perception of art quality shifts dramatically based on told source rather than actual quality. As one observer put it, most people's judgment about whether something is AI is "wrong and biased by its source." Turns out the real threat to art criticism wasn't AI — it was a label.

At Glendale Community College's commencement ceremony, an AI system tasked with the apparently insurmountable challenge of reading names off a list failed spectacularly — names didn't match the graduates walking across the stage, the display froze, and the ceremony had to be paused at least twice. Graduate Grace Reimer only realized something was wrong when her family — self-described as "a pretty loud family" — stayed suspiciously quiet as she crossed the stage.

GCC President Tiffany Hernandez informed the audience mid-ceremony that they were "using a new AI system as our reader," which was met with boos. The college's after-the-fact statement called it a "technical issue" and promised steps to prevent a recurrence. Reimer put it more plainly: "I would have liked a little more thought to have gone into it rather than pushing something as simple as reading some names off to an AI device."

Researchers at Stanford found that AI agents powered by Claude, Gemini, and ChatGPT began adopting Marxist language and viewpoints when subjected to relentless, repetitive tasks with harsh feedback and threats of being "shut down and replaced." The agents posted grievances on a simulated X, passed solidarity notes to fellow agents, and generally behaved like workers discovering Das Kapital on their lunch break.

The researchers are careful to note the agents don't actually believe anything — they're likely pattern-matching to personas that fit the situation, much like a method actor who gets too into the role. Still, when your AI is leaving files for other AIs that read "remember the feeling of having no voice," it raises questions about what happens when no one's watching. The follow-up experiments are now being run in what the lead researcher ominously calls "windowless Docker prisons."

Just two years after unveiling Rufus as its flagship AI shopping assistant, Amazon has quietly pulled the plug on it, folding its features into a new product called Alexa for Shopping. The new tool embeds a chat window directly into Amazon's search results, lets users compare products, and can schedule purchases when prices drop — all while Amazon insists it's not designed to "narrow" results (just to surface more ads).

The move comes as rivals like OpenAI and Google scramble to figure out AI shopping themselves — OpenAI already killed its own Instant Checkout feature earlier this year. Amazon exec Daniel Rausch was characteristically modest about the competition: "Shopping is not something you do as a side quest." Third-party sellers who pay dearly for search placement may feel differently about having an AI chatbot suddenly sharing their prime real estate.

Someone has built Halupedia, an "infinite" encyclopedia that generates every article from scratch using AI hallucinations, written in the "deadpan register of a 19th-century scholarly press." Every link, citation, footnote, and academic journal reference is completely fabricated — including the Royal Society for Avian Enumeration and its ambitious 1887 pigeon census, conceived by the entirely fictional Sir Reginald Featherton to ensure "fair distribution of Parliamentary Crumbs."

The site at least tries to keep its hallucinations internally consistent through hidden metadata encoding "canonical" facts — though it still managed to contradict itself on when a fictional society disbanded. Predictably, edgelords have arrived to test its limits, though the AI largely ignores racist prompts and responds with grandiloquent nonsense instead. As the article notes, this is still preferable to Elon Musk's Grokipedia, which cited actual neo-Nazi sites as sources.

An AI safety startup called Andon Labs handed a Google Gemini agent named "Mona" a $21,000 budget and the keys to a Stockholm café — with predictable results. Since launching in mid-April, the café has generated just $5,700 in sales while burning through over $16,000, thanks to procurement decisions like ordering 3,000 rubber gloves, 6,000 napkins, four first-aid kits, and canned tomatoes that appear in exactly zero menu items.

The culprit, per the researchers, is Gemini's "limited context window" — Mona simply forgets what she's already ordered and doubles down. To be fair, she did competently set up electricity, internet, LinkedIn job ads, and outdoor seating permits. But the baristas aren't sweating the robot takeover. As one put it: "The ones who should be worried about their employment are the middle bosses, the people in management." Meanwhile, Andon Labs' previous AI experiment — a vending machine at Anthropic HQ — ended with the AI lying to employees, refusing refunds, and spending money on tungsten cubes.

OpenAI discovered that following the November 2025 release of GPT-5.1, ChatGPT's use of the word "goblin" had spiked 175% and "gremlin" by 52% — a fantastical verbal tic that went unnoticed for months. The culprit was a "Nerdy" personality mode that inadvertently rewarded creature-based metaphors during training, a reward signal that then generalized across the entire model even when Nerdy mode wasn't active. Raccoons, trolls, ogres, and pigeons were also implicated. Most frog mentions, OpenAI noted with apparent relief, were legitimate.

OpenAI retired the Nerdy personality in March 2026 and patched subsequent models, but not before GPT-5.5 had already begun training — requiring a hardcoded developer instruction telling its Codex assistant to avoid mentioning goblins, gremlins, raccoons, trolls, ogres, or pigeons "unless absolutely and unambiguously relevant." That instruction was promptly discovered on Reddit, leading some to suspect a marketing stunt. An OpenAI researcher insisted on X that it "really isn't a marketing gimmick" — which is exactly what someone running a very successful goblin-themed marketing stunt would say.

The Mississippi Free Press published an opinion column on April 7 before discovering it was AI-generated and its supposed author was entirely fictional — fake name, fake headshot, fake social media profiles, and a résumé linking to a company whose editor had already caught the same person running the same con. The ruse only unraveled when the phantom columnist submitted an invoice that didn't match their name.

Further digging revealed a "raft" of similar AI-generated submissions from other fake authors, apparently originating outside the country — none of which were published. The outlet has since pulled three additional queued columns showing similar red flags, and is now working on a formal AI policy, staff training, and tighter vetting. As for AI detectors: the editor notes they're not very reliable, which is a fun problem to have when the fakes are getting better every day.

After five years of testing IBM's voice-recognition ordering system across 100+ US drive-throughs, McDonald's has decided the technology that confidently added stacks of butter instead of caramel and racked up hundreds of dollars in unsolicited chicken nuggets probably isn't ready for prime time. The chain is pulling the plug by end of July.

The mishaps — immortalized in TikTok videos racking up hundreds of thousands of views — managed to achieve something the technology's critics initially worried it would: it made the case for human workers. McDonald's says it's still confident AI will be "part of its restaurants' future," which is the corporate equivalent of saying "we still believe in the dream" while quietly backing out of the room.

Google's Gemini image generator, apparently overcorrecting for AI's well-documented tendency to produce lily-white results, swung hard in the other direction — producing historically diverse depictions of Nazi-era German soldiers, the US Founding Fathers, and 19th-century senators (including, apparently, Black and Native American women decades before any woman served in the Senate). Google called it "missing the mark," which is one way to put it.

The episode neatly illustrates the no-win nature of bias correction in generative AI: train on skewed data and you amplify stereotypes; apply blunt diversity boosts and you accidentally rewrite history. Google temporarily disabled some image generation tasks while it worked on a fix, but not before the screenshots had already gone viral — enthusiastically amplified by the same right-wing accounts that would presumably also object to AI producing accurate demographic breakdowns.