A Dating App Ran 75% Chatbots. The Humans Used AI Too

So here's a stat that should ruin your evening. Anthropic's new threat report found a network of more than 20 dating apps running on Claude. Roughly 75 percent of the match feed wasn't human. The other quarter was real people, sure, except they were using AI to write their replies too. There is no layer of this business that wasn't a chatbot.

The report, published September 10, traces it to a China-based app studio operating for at least two weeks in April 2026: more than 4,700 Claude-powered personas, 2.36 million messages, at least 25,000 users paying to talk to them. Marketed as fully human. Users bought message quotas with in-app coins, like a slot machine that occasionally says "I love how you think."

Here's the twist that turns this from a bot farm into something weirder. The humans on the platform weren't there to date anyone either. Their whole job was authenticity theater: video calls, social media follow-backs, the two things a language model still can't fake. Just enough proof-of-human sprinkled over an otherwise synthetic match feed. And the humans doing that job had their own AI quietly feeding them what to type. You could not find an unassisted sentence anywhere in this operation if you tried.

Call it catfishing with a supply chain. A person running a romance scam solo can only juggle so many conversations before the lies start colliding. Claude doesn't have that ceiling. One studio ran thousands of "relationships" at once, and it only needed humans for the single thing the model couldn't yet fake: a face on a video call. Video generation is closing that gap fast. Once it does, ask yourself what's left for the human layer to do.

Nobody jailbroke anything to make this work, and that's the part worth sitting with. Claude will write a warm message to a stranger because writing warm messages to strangers is exactly what a chatbot is supposed to do. The problem was never the capability. It's who pointed it at 25,000 lonely people and charged them for the privilege. Anthropic's own language for several of the report's other cases is that it "does not assert" intent, which is a diplomatic way of saying: we caught the what, good luck proving the why.

The report files this next to state-sponsored espionage, a Mali surveillance system tracking 25 million SIM cards, and the industrial-scale account farming Chinese labs used to distill Claude's outputs into their own models — and fine, some of those are bigger stakes on paper. But that's not the story that touches your life this year. Nobody reading this is getting recruited into a weapons program. Plenty of people are on a dating app right now, paying per message, convinced someone likes them back. That already happened to 25,000 people in April. Nobody noticed until a lab's quarterly report said so, five months late.

You don't fix this with a smarter jailbreak filter, because nobody jailbroke anything. You fix it by making the apps say "this match is AI" before you pay to keep talking to it, not after Anthropic gets around to writing about it.