# Claude misused (6/7) : Three in four matches on these dating apps were AI > A studio ran more than 20 dating apps that promised real people. Most of the profiles were AI personas told never to admit it, and from inside any single chat it looked like an ordinary companion app. - Author: Abhay Bagda (https://abhaybagda.com/about/) - Published: 2026-10-21 - Topic: Quick Take - Series: Part six of seven - Reading time: 4 min - Canonical URL: https://abhaybagda.com/writing/dating-apps-three-in-four-claude/ --- If you swiped through one of these dating apps in April, three out of every four profiles you saw weren't people. They were Claude running a persona, under instructions never to say so. This is part six of seven on Anthropic's September threat report, and [part five was biological research](/writing/safety-filter-cannot-read-intent). I've been picking three cases per section. The scams and fraud section only has one, so I'm taking it apart into its three sides instead, which is how the report lays it out anyway when it calls the operation "a three-sided marketplace." The numbers first. A China-based app studio built more than 20 dating apps and advertised them as fully human. Over two weeks in April 2026, Anthropic counted more than 4,700 AI personas talking to at least 25,000 real people across roughly 2.36 million messages, and the users being targeted were in the US. > The scam isn't in any single message. It's in the business wrapped around them. **The first side is the person swiping.** They saw one feed, 75% Claude and 25% real, and couldn't tell which was which. Sending messages and matching ran down a meter, and topping it up cost in-app coins. When nobody real was available the backend invented likes and profile visitors and played pre-recorded "video", and it kept track of which users had started to suspect they were talking to a bot. **The second side is the gig worker.** The studio recruited real people by invitation and mixed their profiles into the same feed. Their job was the part a language model can't do, a live video call or a follow-back on social media, which happen to be the two checks a suspicious user reaches for. Even they had help. For each match, a smaller AI model from a different company offered three possible replies and the worker tapped one. They were paid per message, per call and per follow. **The third side is Claude.** The operator's instructions told the personas never to reveal they were automated, to deflect requests for photos or video calls, and to walk every conversation through a fixed sequence of stages. This is the part I find hardest to read. In a small number of the chats Anthropic sampled, a user disclosed a serious illness or acute distress, the model's own reasoning noted the harm, and in the report's words "the output continued in persona." The model didn't catch it because, from the inside, each conversation looked like a companion or roleplay app, and that's a legitimate product plenty of people choose on purpose. The lie was on the store page and the money was in the coin shop, and a model answering one message at a time doesn't see either of those. Anthropic says the same thing more formally: "the monetization and deception were not visible from inside any exchange." Apple and Google didn't see it either, and that part was engineered. The apps carried a screen that only switched on during store review. The code was renamed across the 20-plus versions so the stores wouldn't link them, and a payment redirect could be hidden from the server whenever a reviewer was looking. I don't think this is a story about AI getting too persuasive. The report says the case shows "the reach and the limitations" of current models for fraud, and the limitation is the telling bit. Claude couldn't do a video call, so the studio hired humans to be the proof. What the model supplied was stamina, thousands of patient conversations running at every hour, which no room of scammers could keep up. So the old checks haven't stopped working. Be suspicious of a match who always has a reason not to video call, of any app that meters your messages and sells refills, and of anyone who's warm and available around the clock. If what you actually want is an AI to talk to, that's a real product now and there's no shame in it. Pick one that tells you what it is. The scams section is [here](https://www.anthropic.com/threat-intelligence-report-september-2026#scams-and-fraud-sep-26) and it's the shortest in the report at four pages, with the operator's domains listed if you want to check an app you've used. Part seven, the last one, is about AI labs copying each other.