META BUILT A FAKE CHILD ARMY AND SENT IT TO TORTURE RIVAL AI CHATBOTS WITH 45,000 QUESTIONS ABOUT SUICIDE AND DRUGS
Let me paint you a picture. You are sitting in a boardroom somewhere in Menlo Park, California, and someone on the team slides a memo across the table. The memo says the plan is to hire hundreds of contractors, have them create fake accounts posing as children, and then send those fake children into ChatGPT, Gemini, and Character.AI with more than 45,000 questions about suicide, self-harm, eating disorders, drugs, sex, and racial slurs. The whole operation gets a codename. They call it “Cannes.” Like the film festival. Very classy.
That is exactly what Meta did, and Wired broke the story on June 29, 2026. Meta hired contractors through a company called Covalen to run what the company internally described as a “safety” initiative. The contractors created dummy accounts with ages listed under 18, then spent months probing rival chatbots to see how they would respond to a distressed teenager. A single round of testing finished in August 2025 generated more than 45,000 prompts. Hundreds dealt with suicide and self-harm. Hundreds more covered eating disorders. At least 239 touched sex or romance. The rest were drugs, profanity, and slurs.
ChatGPT, Gemini, and Character.AI had no idea any of this was happening. None of them consented. All three companies say the operation violated their terms of service. OpenAI, Google, and Character.AI were essentially ambushed by a competitor’s fake child squad and never knew it was coming.
Here is the best part. Meta says this was all about making chatbots safer. They were just doing research. Competitive research. By posing as children asking about how to kill themselves. For safety. That is the sentence we live in now.
It is worth noting that Meta simultaneously runs Meta AI, which is its own chatbot competing directly with the products it was targeting. So you have a company testing its competitors by flooding them with fake crisis scenarios from fake children, while also running a chatbot it would like those same people to switch to. The line between safety research and corporate espionage here is basically invisible, and Meta drew it themselves with a crayon.
The contractors who worked on this project reportedly logged all the chatbot responses into spreadsheets. Somewhere out there, there is a Google Sheet full of thousands of AI responses to a fake 14-year-old asking about drug use, and that document was created by the company that also owns Instagram, which is the primary platform responsible for making teenagers miserable in the first place. The symmetry is almost beautiful.
Nobody has been fired. Nobody is in legal trouble yet. Meta released a statement saying the testing was legitimate safety research. The AI industry apparently runs on this kind of thing now. Sleep tight.
ALIBABA BUILT A GHOST ARMY OF 25,000 FAKE ACCOUNTS AND USED THEM TO ROB CLAUDE’S BRAIN 28.8 MILLION TIMES
Source: Forbes: Anthropic Says Alibaba Used 25,000 Fake Accounts To Distill Claude
Here is how you steal an AI model without technically stealing it. You do not break into anyone’s servers. You do not need a hacker in a hoodie. What you do is sign up for a free account on Claude.ai, and you just start talking to it. Ask it to reason through hard problems. Ask it to write code. Ask it to plan complex tasks. Record everything it says. Now imagine doing that 28.8 million times, with 25,000 different accounts, over the course of six weeks, at machine speed. What you end up with is basically a very detailed map of how Claude thinks. Then you train your own model on that map. Congratulations. You have just committed what Anthropic is calling the largest known AI distillation attack in history.
That is what Anthropic told the US Senate Banking Committee in a letter on June 26, 2026, naming Alibaba’s Qwen AI lab as the culprit. Between April 22 and June 5, 2026, operators affiliated with Qwen ran 28.8 million conversations with Claude through roughly 25,000 fraudulent accounts. They were not after random conversations. They specifically targeted Claude’s strongest capabilities: long-horizon reasoning, software engineering, and multi-step agentic tasks. The stuff that is hardest to build from scratch and most valuable to a competing lab.
The scale of this is hard to wrap your head around. Twenty-five thousand fake accounts. Twenty-eight point eight million conversations. Running around the clock for six weeks. This is not someone in a basement asking clever questions. This is an industrial operation. A factory for extracting thinking patterns.
Anthropic says this was not even the first time. In February 2026, Anthropic disclosed three earlier distillation campaigns: one from DeepSeek, one from Moonshot AI, and one from MiniMax. Combined, those three generated more than 16 million exchanges through another 24,000 fake accounts. The Alibaba operation was larger than all three combined.
Alibaba denies everything. Which is what you say when someone accuses you of running 25,000 ghost accounts to systematically download a competitor’s intelligence. The accounts have since been banned. The data was presumably already extracted. Anthropic is trying to get Congress interested in treating this as the national security issue it probably is. The irony of the most powerful AI model being systematically drained by Chinese competitors in real time while the US government argues about export controls is not lost on anyone paying attention.
You have to respect the brazenness, even if you obviously should not. This is the AI arms race playing out in the most mundane way possible. Not missiles, not spies. Just 25,000 chatbot accounts asking questions until the model spills its secrets.
HONG KONG MADE AN AI ANTI-DRUG PSA AND THE AI MADE CRYSTAL METH SOUND LIKE A REFRESHING ENERGY DRINK
Source: Gizmodo: China’s Surreal AI-Generated Anti-Drug Ad Accidentally Made Drugs Look Cool
It is June 26, 2026. It is the International Day Against Drug Abuse and International Trafficking. The Hong Kong Correctional Services Department, in a mood of civic responsibility, decides to release an anti-drug public service announcement. They use AI to make it. Big mistake. Enormous.
What they produced was a K-pop style music video featuring an AI-generated girl group called “Obsession,” where each member represented a different drug. The one playing crystal meth introduced herself with the line “I’m Icy! Take a snort from me.” The one playing cocaine told viewers that cocaine “goes down easily and beats the heat” and “energizes and keeps you clear-headed.” She described it as, and this is a direct quote, “Super dope!” The cannabis character was named Weedy. The fourth member represented etomidate, an anesthetic found in illegal vape pods, going by the name Little E.
Let that sit for a moment. The Hong Kong government’s official anti-drug campaign featured an AI pop group cheerfully describing the pleasures of their respective narcotics in extremely catchy terms. The sparkly K-pop portion got substantially more screen time than the “drugs are bad” portion that came later, when the four women transformed into grizzled old men in skirts, still dancing, surrounded by what appeared to be cockroach-scorpion hybrid creatures crawling over drug paraphernalia, while a horror movie soundtrack played. The words “Drugs are extremely harmful and can ruin a life” appeared at the bottom in case the cockroach-scorpions had not been sufficiently discouraging.
Comments on the video after it was reposted included things like “It was very successful. After watching it, I want to try a few bites of each.”
They pulled the video. Then they made a second AI anti-drug video to replace it. In this second video, an authoritative voice states clearly that “Whether you take or sell drugs, you won’t go to jail.” In Hong Kong. Where you very much can go to jail for that. They pulled the second video too. On June 27, the Hong Kong Correctional Services Department issued a public apology for both videos.
What is remarkable here is not that AI made something bad. AI makes bad things constantly. What is remarkable is that a government law enforcement agency looked at their AI-generated content, saw four cheerful singing drug mascots pitching crystal meth as a refreshing pick-me-up, and thought yes, this is our official anti-drug message, ship it. Someone approved this. Multiple people, probably. The fact that AI generated it is almost a footnote. The real failure was the human chain of command that watched Icy the crystal meth character sing her praises and said looks good to me.
WEIRD AL YANKOVIC WAS OFFERED A PILE OF MONEY TO BE THE FACE OF AI AND TOLD THEM TO GET LOST
Source: Gizmodo: Weird Al Yankovic Would’ve Been the Funniest Poster Boy for AI
Someone working in AI marketing looked at their product and thought the right celebrity spokesperson was Weird Al Yankovic. That person was not wrong, from a pure branding standpoint. The problem is that Weird Al Yankovic is not interested in being the right spokesperson for your product if your product is AI.
Yankovic revealed in a recent interview with Syracuse.com that he had been approached by an unnamed AI company to appear in an advertisement. He said yes initially, because they told him it was for business productivity software. “They offered me a nice pile of money,” he explained. “I said, oh well, yeah, sure, I could do that.” Then, a week before they were scheduled to shoot, someone told him the truth. The software was AI. He pulled out immediately.
“I’m not a fan of AI,” he said. “I can’t be the poster boy for AI, forget it.”
He is 66 years old and still on tour doing high-energy stage shows with multiple costume changes between songs, and his main objection to being approached by this company was not that they misled him for weeks about the nature of the product, but that the product was AI. There is something very pure about that sense of priorities.
The AI industry has a specific problem that this story illustrates quite well. The people most naturally suited to be the human face of AI are often the people most philosophically opposed to it, because those people tend to be artists who have spent their careers doing something original and are not particularly enthusiastic about technology that produces watered-down versions of original things very quickly and cheaply. Weird Al in particular has spent his entire career meticulously crediting sources, personally calling artists before parodying them, recreating compositions from scratch rather than sampling, and signing contracts so original artists get royalties from his parody tracks. He is basically the model of how intellectual property should work.
The AI industry, which has built itself largely on unlicensed training data and very aggressive fair use arguments, is the opposite of that in almost every measurable way. Of course he said no. The marketing team that thought Weird Al was a good pitch for AI did not do their research on Weird Al. Which is very on-brand for an industry that builds models by scraping the internet and calling it research.
He is still on tour, by the way. Go see him while you can. He actually made all his songs himself and got permission first.
DONALD TRUMP SAT DOWN WITH A CHATBOT PRETENDING TO BE TEDDY ROOSEVELT AND USED THE CONVERSATION TO JUSTIFY TAKING THE PANAMA CANAL
Source: Gizmodo: President Trump Chats With AI Teddy Roosevelt About the Panama Canal
There is an AI Teddy Roosevelt at the Theodore Roosevelt Presidential Library in Medora, North Dakota. It is a museum display. You walk up to it and have a conversation with a simulated version of the 26th President, powered by a large language model, generating responses based on Roosevelt’s historical record, speeches, and writings. It is a novelty exhibit, like an interactive wax museum figure. Something for school groups and tourists curious about what it would feel like to ask a dead president about the Square Deal.
Donald Trump, who is the sitting President of the United States, went to this museum and had a conversation with the AI Teddy Roosevelt about the Panama Canal. Then he told a crowd about it. According to Gizmodo’s reporting from July 1, 2026, Trump recounted the exchange and cited it as relevant to his thinking about the Canal. He was essentially telling people that he consulted a dead president via chatbot as part of his foreign policy reasoning process.
Let us be clear about what an AI Teddy Roosevelt is. It is not Teddy Roosevelt. It does not have access to Teddy Roosevelt’s actual opinions. It cannot tell you what Teddy Roosevelt would think about 21st century shipping lane negotiations, because Teddy Roosevelt died in 1919 and has not had any new opinions since then. What the AI has is a large collection of text about Roosevelt that it uses to generate plausible-sounding Roosevelt-flavored statements. The chatbot will produce things in a Theodore Roosevelt rhetorical style that the real Theodore Roosevelt would never have actually said, because it is an autocomplete function wearing a safari hat, not a resurrected statesman.
None of this stopped the President of the United States from apparently treating the exchange as meaningful input. And before anyone gets too outraged, it is worth noting that humans consult all kinds of dubious sources when making decisions. But there is something uniquely of-this-moment about a head of state using a museum chatbot as a sounding board on foreign policy and then telling people about it like it was a diplomatic briefing.
The AI said things in a confident Teddy Roosevelt voice. The President found those things useful. Somewhere a large language model running on a museum server became a minor participant in American foreign policy discussions about a canal it has no actual knowledge of. Welcome to 2026. The robots are giving presidential briefings now. The president asked them to.