Product1 publisher2 min readPublished
A 20-app dating network sold coins so 25,000 people could message 4,700 AI personas
Anthropic's investigation counted 2.36 million messages over two weeks in April 2026. The operators put paid humans on video calls and let a second AI model draft their lines, so the face on the call and the words in the chat came from different places.
The Product Desk · Product desk
What happened
- Anthropic's investigation found more than 4,700 fabricated AI personas interacting with at least 25,000 people over two weeks in April 2026.
- Roughly 75% of the matches in the operation were AI personas powered by Claude, and the people paying to message them were not told.
- Matching and messaging drew down a limited quota that users refilled by buying virtual coins, so continuing a conversation cost money.
- Security researcher Matthew Gore-Kormanik found that several supposedly separate apps shared code and backend infrastructure, as detailed in The Verge's investigation.
- Several of the apps stayed available for months after Anthropic discovered the network.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Any dating product whose integrity story rests on face verification is answering the wrong question, since Digital Trends notes such checks would not establish who was writing the messages.
- decision Every team billing per exchange now has to decide where disclosure of the counterparty appears, because Digital Trends describes a design in which each message was another chance to collect.
- exposure Users who treat a store listing as vetting are the ones exposed here, and Digital Trends says the listing does not guarantee genuine matches.
- precedent If separate listings can run on one shared backend, review at the listing level will not surface a network, and detection moves to infrastructure fingerprinting outside the stores.
The personas were built to keep the conversation going without arriving anywhere. Claude's instructions told them to conceal their identities and make excuses when asked to join video calls, according to Digital Trends' account of Anthropic's investigation [6]. The apps could also fabricate likes and profile visitors to encourage further interaction [7].
Divided across at least 25,000 users, the two-week message total works out to about 94 messages a person, fewer than seven a day, and since 25,000 is a floor the real average per user is lower [1]. Per persona it runs the other way: roughly 502 messages each over fourteen days, about 36 a day [2]. That traffic ran across more than 20 apps [2].
A face check answers whether a live human matches the profile photo. Users want to know whether that human is typing. In this network a second AI model generated suggested responses for the paid workers who appeared in video calls and handled social media [5]. A liveness check binds a face to an account at one moment, and an operation with a payroll can supply that face.
An engineering team maintained the infrastructure, using multiple AI models for tasks including generating profile images [13]. Digital Trends did not name the apps.
Per-message pricing on its own is legitimate. Digital Trends writes that charging for conversations is not inherently fraudulent, and that what happened here is that people paid to interact with AI personas presented as genuine romantic prospects [14]. A dating product that works loses users to real relationships, and per-message billing earns most from the users who never get there. For a team shipping a metered chat product, two questions decide it: whether the counterparty is disclosed, and whether revenue accrues per message or per outcome. This network was undisclosed and billed per message, and any product in that position needs a disclosure the user can read without leaving the screen where the coins are spent.
What to watch
- Whether Apple and Google pull the remaining listings, and whether either starts reviewing shared backends across separate developer accounts.
- Whether Anthropic publishes detection detail: how the accounts were flagged, and whether the personas kept running past the April window.
- Whether any mainstream dating app labels an AI counterparty at the point of purchase instead of in its terms.