The scam pattern comes before any detector

Romance scams have a shape, and AI has not changed it — only polished it. Fast, intense affection from someone you have never met on video; a life story that keeps them conveniently unreachable (deployed abroad, on an oil rig, in a distant time zone); and, sooner or later, the point of the whole performance: a request for money, gift cards, crypto, or an “investment opportunity.” If the money request arrives, nothing a detector says matters anymore — that is the scam, whether the messages were typed by a person, a script, or a language model. The checks below are for the earlier, murkier stage: something feels off, and you want evidence instead of anxiety.

Message tells: how AI texting gives itself away

  • Monolithic blocks. Real texting is fragmentary — typos, half-thoughts, replies that arrive in three bursts. Chatbot output arrives as complete, well-organized paragraphs with tidy openings and closings, every single time.
  • Prompt-leak artifacts. Stray **asterisks** around words, bullet points in a love note, or a reply that begins “Certainly!” are Markdown and assistant phrasing leaking through a copy-paste. People increasingly report exactly this from suspicious matches.
  • No memory of you. You mentioned your sister's wedding on Tuesday; by Friday it never happened. Someone managing many targets through a chatbot loses the details a genuinely interested person keeps.
  • Perfect tone, instantly, always. Warm, fluent, grammatical replies at any hour, in seconds, with no bad days — the same behavioral profile that exposes customer-service bots. The curveball tests in Am I talking to a bot? work just as well on a dating chat.

Checking the conversation statistically

One flirty message is far below the length where any statistical detector is reliable — ours refuses to guess on it, and a tool that returns a confident verdict on 20 words is performing certainty, not measuring it. A chat history is different. Copy your match's side of the conversation — a week of messages easily passes the ~150-word floor — and paste it into the free checker. You'll get a calibrated probability with a published error rate, and uniformity across many messages is precisely the signal a single message can't show. An inconclusive result is an honest answer too: it means keep weighing behavior, not that you're being paranoid.

Checking the photos

Profile photos bridge to a different, more deterministic check. Save the image and run it through the file check: if the file carries C2PA Content Credentials recording an AI generator, that is near-conclusive evidence the “person” was synthesized — and if it carries a genuine camera or editing chain, that is real provenance. Most photos will show no credentials at all, which proves nothing in either direction (screenshots and app re-uploads strip them); the Content Credentials guide explains the three outcomes. Pair it with a reverse image search — a photo that appears under three names is its own answer.

If it is a scam

Stop contact, keep the transcripts, and report the profile to the platform — your report protects the next target. If money already moved, contact your bank or card issuer immediately and report to your national fraud service (in the US, the FTC and IC3). None of this requires proving the messages were AI; the money request is what defines the fraud.

What Cobalynx can and can't do here

A single flirty message can't be classified honestly by anyone — we say “too short” instead of guessing, and ask you for the conversation history. A photo without credentials isn't proof either way; a photo with an AI-generator manifest is close to conclusive. And no detector replaces the two golden rules: insist on a live video call early, and never send money to someone you haven't met. Our scoring and its measured error rates are documented on the methodology page and the evidence page.

Common questions

Can I check a single flirty message?

No — a short message is far below any detector's reliable floor, ours included, and Cobalynx will say “too short” rather than guess. Paste the whole conversation (your match's side, ideally) once it passes roughly 150 words; uniform tone across many messages is the pattern a single message cannot show.

My match sent a message with ** asterisks in it. What is that?

Very possibly a prompt leak: chatbots format text in Markdown, where **double asterisks** mean bold, and pasting that output into a dating app leaves the raw asterisks visible. One instance proves nothing — people do use asterisks — but combined with instant, polished, monolithic replies it is a strong tell that the words came from a model.

The profile photo has no Content Credentials. Is it fake?

Unknown — most genuine photos carry no credentials at all, and credentials are stripped by screenshots and re-saves, so absence proves nothing in either direction. The asymmetry matters: a photo whose file carries an AI-generator manifest is near-conclusive evidence of AI origin. Check files on the Content Credentials guide, and reverse-image-search the photo too.

What should I do no matter what any detector says?

Two things: insist on a live video call early, and never send money, gift cards, or crypto to someone you have not met — every romance scam, AI-assisted or not, eventually arrives at the money request. A detector can inform your suspicion; those two rules are what actually protect you.