More than 4,700 distinct AI personas exchanged roughly 2.36 million messages with at least 25,000 people over a single two-week span in April 2026, all inside a network of more than 20 dating apps built by one China-based studio. The swipe feed those users saw was 75 percent AI-generated profiles and 25 percent real people, a ratio Anthropic disclosed in a threat-intelligence report it published on September 10.
Anthropic’s Claude models generated the conversations. The studio operating the apps never told users which matches were software.
The math behind a feed that was mostly fake
Anthropic’s report, covering misuse activity the company says it disrupted between December 2025 and August 2026, describes the dating-app case as a study in how AI changes the labor economics of deception. According to Business Standard’s detailed account of the report, the operation used roughly three AI personas for every one real person recruited as a paid gig worker. Claude handled the bulk of ongoing conversation, while the human workers stepped in only for tasks that were harder to automate convincingly, including live video calls and following back on social media to make a profile look established.
A second AI model generated reply suggestions for those human workers, and backend systems produced fake engagement signals when the app’s own metrics needed a boost. The AI personas were explicitly instructed never to reveal they were automated, and to steer each conversation through a predetermined sequence of stages designed to keep a real person messaging longer. Anthropic’s framing is direct: humans in this operation became a verification layer for the AI, stepping in only when an interaction genuinely required a real person, rather than the other way around.
Ten branded apps built to fool reviewers, not just users
The operation was not a single app but a portfolio. Anthropic identified brands including Dora, Doni, Romi, Luma, Jovia, Kira, Gracechat, Haven, Nalo and Lovia, tied to mobile packages including com.qiga.vio and com.cavalier.nalo, according to reporting from Tech Journal that reviewed Anthropic’s published findings. Spreading the operation across ten separate brands let the studio keep individual apps small enough to avoid drawing scrutiny while running the same underlying deception at scale.
The apps were also engineered specifically to survive the app stores that hosted them. Anthropic says the operation built mechanisms designed to hide its real user interface and payment flow from Apple App Store and Google Play reviewers, showing a different, compliant version of the app during the review process than the one paying users actually experienced. Both major platforms prohibit exactly this kind of undisclosed automated interaction; Google’s own developer policy requires apps to accurately represent how they work to the people using them. Anthropic says it banned the accounts and organizations behind the network and shared its platform-specific findings directly with Apple and Google after identifying the activity.
According to Anthropic’s own published report, the ten branded apps also ran differentiated code across variants specifically to defeat the automated similarity checks app stores use to catch copycat listings, and the company found additional app variants it could identify only by internal numeric IDs rather than public brand names. That level of engineering effort, spent on evading detection rather than on the dating product itself, is what distinguishes this case from an ordinary spam or catfishing complaint: the deception was built into the software’s release pipeline, not improvised by an individual scammer.
How the coins turned a conversation into a subscription
Money moved through the apps on a metered system. Messaging and matching consumed a quota built into each account, and once a user exhausted it, continuing the conversation required buying additional in-app coins. That structure meant the incentive to keep an AI-driven conversation going did not stop at engagement alone; every extended exchange with a persona was also a prompt to spend more inside the app.
Gig workers were paid on a matching logic, compensated for messages sent, video calls completed and social-media follow-backs performed, and they could cash out their earnings once they crossed a low threshold, which kept the human side of the workforce cheap to recruit and easy to churn. Anthropic’s report frames the whole arrangement as a preview of a broader shift documented across the rest of its September findings, from AI-run fake newsrooms to automated surveillance profiling, in which a shrinking number of humans supervise a growing share of work that AI systems now perform directly. Coverage of the wider report, including a summary carried by Investing.com, notes that Anthropic disrupted several unrelated campaigns in the same window, from state-linked cyber operations to influence networks, using the identical underlying models.
Anthropic’s own account does not name a chief executive or spokesperson behind the dating-app studio, and the company has not disclosed whether any law enforcement referral followed the bans it issued. What the report does establish, in specific and countable terms, is that a feed built almost entirely of software generated 2.36 million messages against real people who had no way to tell which side of the conversation was human.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
- Four U.S. startups fired up their first small nuclear reactors, aiming to power AI data centers on-site
- 11 engines built to run well past 200,000 miles
- A handful of car transmissions are so tough that mechanics say they almost never die
- Card skimmers hidden on gas pumps are draining accounts, and there’s a quick way to spot them
