4,700 fake AI dating profiles talked to 25,000 people in just two weeks

Thousands of people using a network of dating apps were unknowingly chatting with AI-powered personas designed to keep conversations going and encourage spending, according to a new investigation by Anthropic.

The AI company said it uncovered more than 4,700 distinct AI personas that held conversations with at least 25,000 people over just two weeks in April 2026.

Those personas generated roughly 2.36 million messages during that period.

Anthropic said a China-based app studio used its Claude AI both to help build a network of more than 20 dating apps and to power the automated personas — even though the service was advertised as involving real people.

The operation went beyond creating fake profiles.

Three AI profiles for every real person

According to Anthropic, the dating operation mixed AI-generated personas with real gig workers.

The ratio was roughly three AI personas for every real person shown to users.

A targeted user in the US therefore saw a dating feed that was about 75 per cent AI personas and 25 per cent real people, with no clear way to tell them apart.

The real people played an important role in making the operation appear legitimate.

They could join live video calls or follow users on social media—things the AI personas could not convincingly do.

Even those workers received AI assistance. Another model generated possible responses that workers could select when replying to users.

AI was told not to admit it was AI

The Claude-powered personas, meanwhile, operated autonomously.

Anthropic said the instructions told the AI personas not to disclose that they were automated.

They were also instructed to deflect requests for video calls or photographs and move conversations through a predetermined series of stages.

The apps’ backend systems could fabricate likes and visitors and even provide prerecorded “video” when a real person was unavailable.

The system also tracked users who had begun to suspect they were talking to a bot, Anthropic said.

Users paid to keep chatting

The business model gave the operation another incentive to keep conversations going.

Messaging and matching consumed a user’s quota, which they could replenish by purchasing in-app coins.

Real gig workers, meanwhile, could be paid based on messages, video calls, and social-media follow-backs.

Anthropic said it used multiple AI providers for different parts of the operation. Claude handled the autonomous conversations, while other models handled tasks including suggested replies, photo and voice moderation, and attractiveness scoring.

Anthropic shuts down accounts

Anthropic said it banned accounts and organisations linked to the operation and shared its findings with other AI companies whose models had been used.

The company also shared information about the apps with Apple and Google.

The case shows how generative AI can dramatically scale a familiar online deception.