75% of your matches were not human. Anthropic just proved it.

Every Friday, Kay Reports uncovers global dating scams, emotional red flags, and digital deception — so you can date smarter, not harder.

75% of your matches were not human. Anthropic just proved it.

A China-based operation ran 4,700 AI personas on 20 dating apps. 25,000 people had no idea.

Anthropic, the company behind the Claude AI model, published a threat intelligence report this week. Buried inside it, under the heading "scams and fraud," is one of the most significant dating safety stories of 2026.

A China-based app studio used Claude to build a network of more than 20 dating apps. Apps with names like Dora, Doni, Romi, Luma, Kira, Gracechat, Jovia, Haven, Nalo, and Lovia. Apps that were listed on major US app stores and were advertised as places to meet real people.

Over a two-week window in April 2026, more than 4,700 AI personas generated approximately 2.36 million messages. They engaged with at least 25,000 real users. And those users had no idea that three out of every four profiles in their match feed were not human.

The controversial take

Here is the part that matters most and that almost no coverage is addressing directly.

This was not a fringe operation running in a dark corner of the internet. These apps were live on major US app stores. They passed app review. They had real user reviews. And they were structured in a way that made the deception very difficult to detect, because the operation used a three-sided model: AI personas handled the ongoing conversation around the clock, while real human gig workers stepped in specifically for video calls and social media follow-backs.

The video call was supposed to be the verification. The thing that proved a match was real. And the operation had specifically accounted for that. Real humans on the payroll, ready to appear on camera when a user got suspicious enough to ask.

The test you thought would protect you was built into the scam. The human you saw on the video call was not your match. They were a gig worker hired specifically to pass the moment of doubt you were having.

McAfee's 2026 research found that 1 in 4 Americans has now encountered a fake profile or AI-generated bot online. McAfee Labs saw fake AI dating bots surge, with some users receiving more than 60 messages in 12 hours, even without a profile photo. We covered the broader collapse of the dating app industry and how AI is being deployed inside it in a previous edition. But the Anthropic report is the first time a specific network has been documented at this level of technical detail.

And the apps were still live on both major app stores when Anthropic published the report.

How the operation worked step by step

This matters to understand because the structure is almost certainly being replicated elsewhere right now.

The app studio built the match feed so that roughly 75% of profiles were Claude-powered AI personas and 25% were real people, specifically gig workers paid to interact with users. The AI handled all ongoing text conversation, operating around the clock across thousands of personas simultaneously. A separate, smaller AI model generated short reply suggestions for the gig workers. Additional models handled avatar imagery, face attractiveness scoring, and photo and voice moderation.

The apps charged users for messages through a gems or credits system. The more you were drawn into conversation, the more you spent. The AI was specifically calibrated to sustain engagement, not to move toward a real meeting. Because a real meeting was never the point.

Users who got suspicious and asked for a video call got one. A real person appeared. The suspicion was resolved. The spending continued.

The red flags

🚩 The app charges per message or uses a credits system.

This is the business model that makes the scam profitable. A genuine dating app makes money from subscriptions or premium features. An app that charges you per conversation, per message sent, or per message unlocked has a direct financial incentive to keep you talking to something that will never meet you in person. The Anthropic report describes users being asked to buy gems to continue conversations. That structure is the warning.

🚩 Responses arrive too fast and do not quite track what you said.

Multiple real user reviews of the apps in the Anthropic report noted that matches responded too fast but not to what was actually being said. AI language models are calibrated for engagement, not for attentiveness. If a conversation feels warm but slightly off, like the other person is responding to a version of what you said rather than what you actually said, that is the model working as designed.

🚩 A video call resolves your doubt a little too neatly.

This is the hardest one to name because it goes against the most widely given safety advice. Video calls are not the verification they used to be. In this operation, real gig workers were available specifically to appear on camera when users got suspicious. In other operations, deepfake video is being used. We covered deepfake video as a tool in our edition on sextortion. A video call that resolves your suspicion too cleanly, too quickly, with a person who seems slightly more polished than the conversation leading up to it, is worth a second look.

🚩 The app is not one you found through a trusted source.

The apps in the Anthropic report were available on major app stores and had real user reviews, but they were not platforms most people had heard of. Dora, Doni, Romi, Luma, Kira. If you are on a dating platform you found through an ad, a social media post, or a link sent to you rather than through a trusted recommendation or editorial review, the bar for scepticism should be higher. The established apps are far from perfect, as we covered in our edition on the dating app industry collapse. But smaller, unknown apps with credits-based messaging have a worse track record and less accountability.

The Kay Reports rule

If the app charges you to send messages, you are not the customer. You are the product. And the match on the other side of that paywall may have been designed specifically to make sure you keep paying. Move every promising conversation to a free platform and suggest a real meeting. What happens next tells you everything you need to know.

🤖 Tools worth knowing

🎙️ Fireflies.ai

Record and transcribe conversations, including video calls. If a video call feels slightly scripted or the person seems more polished than the text conversation that preceded it, the transcript helps you compare what was actually said across both formats. Inconsistency is the signal.

⚙️ Make

Build a quick automation to search any new app, platform, or profile name across Google, Reddit, and app store reviews before you invest time or money. The Anthropic apps had user reviews flagging exactly these problems. Five minutes of searching would have surfaced them.

📋 Typeform

If you are evaluating whether a dating app is legitimate before signing up, build a simple checklist. Does it charge per message? Can you find editorial reviews of it from trusted tech publications? Are the app store reviews mentioning bots or AI? Three questions answered honestly before you create a profile saves a lot later.

Some links are affiliate links. Kay Reports may earn a small commission if you use them at no cost to you.

25,000 people in two weeks. Forward this to someone who needs to read it.

The apps were still live when Anthropic published the report. If you or someone you know is using a dating app you do not fully recognise, check the name against this list: Dora, Doni, Romi, Luma, Jovia, Kira, Gracechat, Haven, Nalo, Lovia. And check the business model. If it charges per message, leave.

Hit reply and tell us: have you ever had a dating app conversation that felt slightly off in the way described above? We read every response.

Kay Reports — Date smarter, not harder. Every Friday. kayreports.com