The dating apps promised real people. Many profiles weren't.

Security researcher Matthew Gore-Kormanik received a call through the dating app Doni and answered it. Afterward, his match wrote as though he had placed the call, even though Doni had shown it as incoming. Gore-Kormanik described the mismatch to The Verge. Who had called whom? In these apps, continuing the conversation could cost money.

The odd call was part of a larger system. In a network investigated by Anthropic, many profiles were invented and Claude generated their messages. The operator instructed the AI personas to deflect requests for photos or video calls and guide conversations through a set sequence. Paid workers appeared in the same feed. They could handle live calls or follow users on social media.

Anthropic says the network comprised more than 20 apps. During two weeks in April 2026, it observed more than 4,700 AI personas exchanging about 2.36 million messages with at least 25,000 people. In the feed it examined, there were roughly three AI personas for each paid worker. The 25,000 figure counts people who chatted, not people confirmed to have paid.

When a person answered

The paid workers could take a live call or follow someone on social media; Anthropic says another model suggested short replies for them to choose from. One encounter could therefore involve a real person without revealing who was behind any of the other profiles.

The operation did not need AI to imitate every part of a person. Automated conversations sat alongside real video calls and social follows. The reports document this mix across the network, not a worker taking over any particular AI chat.

The apps supplied other signs of apparent interest. Anthropic says their systems generated likes and profile visits, and played prerecorded video when no person was available. The same systems tracked users who appeared to suspect a bot: they created signs of interest while watching for signs of doubt.

For The Verge's investigation, Gore-Kormanik examined several Android apps and found shared code and connected backend systems across Doni, Dora, Jovia, Kira, Nalo, and Romi. An internal manual packaged with Doni described work procedures, including ways to draw users into calls. The Verge could not independently confirm every finding in Anthropic's report, but its app investigation showed that several of the named apps were technically connected.

The conversation was the product

Payment was built into the apps. According to Anthropic, matching and messaging used up a quota; users could buy coins to continue. The Verge also examined the payment flow. Continuing the conversation was itself the paid service.

Paying to talk to an AI character can be a straightforward choice when the service says what it is offering. These apps did not tell users that many profiles were automated. Someone hoping to meet another person could spend coins on messages to an AI persona. The service controlled who or what replied.

Claude generated many of the replies. People designed the apps, set the payment rules, and instructed the AI personas to conceal their automation. Anthropic says the business model behind the apps was not visible to the model within a single chat. The deception did not arise in a spontaneous model reply. It was built into the service.

There is no verified total for what users spent or lost. Anthropic says it banned accounts linked to the operation and shared its findings with app stores and other AI providers. Some app listings later disappeared, though the reasons for individual removals and the status of the wider network are unclear.

Users could see that continuing the chat might cost coins. They could not see who was replying.