How to Spot an AI Generated Dating Profile

A dating photo that looks flawless and turns up nowhere in a reverse image search used to be reassuring. Now it's often the mark of a face that was generated rather than stolen.

For years the advice ran the other way. A scammer borrowed a real person's pictures, so a search would find them on someone else's account under another name, and a clean result meant you were probably fine. Image generators broke that logic. SocialCatfish, which sells verification tools, estimates that about 25% of fake dating profiles now use AI-generated photos instead of stolen ones. That's a vendor's figure, so treat it as rough. But it points the same way as every harder number below, and it changes what a clean search result tells you.

A reverse image search compares a picture against images already published somewhere. A synthetic face was never published, so there's nothing for it to match. The search still catches the old scam, where photos turn up under a different name. It says nothing at all about the new one.

So a reverse image search is still worth running, because a match under another name settles the question in seconds. A blank result just moves the job back to the photo itself. According to SocialCatfish's 2026 guide, these tells still show up in current generators:

• Skin that's too smooth and too even, with none of the texture a real face has at close range.
• Hair that blurs into the background at the edges instead of ending in strands.
• Backgrounds with geometric inconsistencies or unnatural blur. Door frames, shelves and any lettering are where the geometry has to hold together and often doesn't.
• Teeth that are unnaturally symmetrical.
• Glassy eyes whose reflections don't match each other.
• A full set of polished, professionally composed shots and not one candid.

One correction to the advice you may have heard. Six fingers and mangled hands were the famous giveaway, and the same guide now says hands and ears are no longer reliable signals as of 2026. If his hands look fine, that tells you nothing either way. Look at the list above.

The clearest case so far comes from the company whose AI was misused. Anthropic's September 2026 threat intelligence report describes a China-based studio, tracked as GTG-15001, that used its Claude models to run more than 20 fraudulent dating apps. The Verge's follow-up investigation, summarized here, counted about 28, including Dora, Jovia, Kira, Luma and GraceChat.

In a two-week window in April 2026, the network ran more than 4,700 AI personas. They engaged at least 25,000 real people and sent about 2.36 million messages, an average of 94 messages to each person they targeted. The personas were instructed "never to disclose they were automated" and to deflect requests for video calls. For the people who pushed anyway, the operators hired human gig workers specifically to handle live video calls and social media follow-backs, roughly one human for every three AI personas. The backend also faked likes, profile visitors and pre-recorded video calls.

Removal was slow. Apple pulled the apps about 77 days after Anthropic detected the operation, and Google Play took 88 to 94 days. One app, Kira, was reportedly still live on September 16.

A second case shows how little video it takes. Fraud-prevention firm BioCatch told Axios that 340 financial institutions saw a 63% rise in romance scam attempts between 2024 and 2025. In one case it described, a victim named Beth was persuaded by a short AI-generated video chat that gave her "just enough visual confirmation" to quiet her doubts before she sent money.

An AI-edited photo and an AI-invented person are different problems. According to a summary of 2026 platform policies, Tinder, Hinge and Bumble all allow AI-generated photos as long as they accurately show what the user looks like. Tinder and Hinge check this by matching a live selfie against the profile pictures. A real man with an over-polished headshot is vain, not dangerous.

The strongest objection to everything above deserves a fair hearing. GRASS, a company that promotes offline socializing, argues that "every digital signal can be faked," and the GTG-15001 case gives it real support: the operators bought their way past video calls with hired humans. If a call can be faked, why bother?

Because faking it costs the scammer something, and the figures show it. The personas were told to dodge video, and the network ran three AI personas for every human. Most of those 4,700 faces had no one behind them who could pick up a call. A live call won't catch everything, but it filters out most of the operation, and a real man won't refuse one.

So ask for a live, unfiltered video call before you invest weeks of messages, and make it one he can't script:

• Ask him to wave a hand slowly across his face. SocialCatfish calls this the most reliable live test, because real-time face-swapping often glitches or flickers when something passes in front of the face.
• Ask him to turn his head well to the side. A guide from Ledger notes that turns past about 60 degrees strain current face-swap tools.
• Ask him something he couldn't have prepared for, and watch whether his lips stay in sync with his words.

If he stalls, blames his connection, or offers to send a video instead, you have your answer.

A real face on a live call settles one question, though: whether this person exists. It doesn't tell you whether he's honest. The FTC reports that romance scam losses rose 22% in its latest figures, with a typical loss of $2,020 per person. It says these scammers play the long game and build a relationship before they ask for money. Verifying the face is the first check, and the rest come once you're talking.