Is your match a real person, or AI?

How the scam works, before any tool

Romance scams have worked the same way for as long as there have been romance scams, and what AI changed is that the messages are better written now. The way it usually goes is that someone you have never seen on video is very warm with you very fast, and they have a story about why you cannot meet them, say they were sent abroad with the army, or they were working on an oil rig. Then one day you get to the part the whole thing was leading up to, which is that they need money from you, or gift cards, or crypto, or they have an “investment opportunity” for you. If you have already had that message then it does not matter what any tool says about the writing, since the money request is what defines the fraud, and it was fraud if the messages were typed by a person or a script or a language model. What we have written below is for the earlier stage, when something feels off and you would like some evidence to go on.

How a chatbot gives itself away in a text message

• One complete block every time. When a real person texts you, what you get is a mess. They make typos, they send you half a thought and then the rest of it a moment later, and they answer in three bursts, having thought of something else after the first one had gone. From a chatbot you get one complete paragraph, with a beginning and an end, and you get it like that every single time.
• Bits of the prompt showing up in the message. If a message reaches you with stray **asterisks** around some of the words, or with bullet points in what was supposed to be a love note, or opening on “Certainly!”, then what you are looking at is the formatting a chatbot uses when it talks to the person running it, and it got to you because that person copied the reply out of one window, pasted it into the other and never cleaned it up.
• No memory of you. You told them about your sister’s wedding on Tuesday and by Friday it had never happened. Someone running a lot of targets through a chatbot loses the details that a person who was interested in you would have kept.
• Perfect tone, instantly, always. They are warm, they write well, they never make a mistake, and they do it at any hour within seconds, and they never once have a bad day or give you a short answer on a night when they were tired. That is the same pattern that gives away a customer service bot, and the questions in Am I talking to a bot? work just as well on a dating chat as they do on a support chat.

Checking the conversation with a scan

One flirty message is far too short for any tool to tell you anything about it. Ours will not guess on it, and a tool that gave you a confident answer on 20 words was guessing, whatever the number looked like. A chat history is different. Copy your match’s side of the conversation and a week of messages will take you past the roughly 150 word floor with room to spare, and you can paste that into the free scan. What comes back is a probability with a published error rate behind it, and what it picks up on is that a chatbot sounds the same in every message it sends, which is something you could never have seen from one message on its own. An inconclusive result means the tool did not have enough to go on, so keep weighing how they behave, and do not read it as a sign that you were imagining things.

Checking the photos

A picture file carries information about where it came from that a message does not, so if you have a bad hunch about someone then save their picture and put it into the file check. If the file carries C2PA Content Credentials that record a picture generator, and the signature checks out against a signer we recognise, that is strong evidence the person you have been talking to was made up. A real camera or editing chain tells you where the picture came from instead. What will happen most of the time, though, is that the picture shows no credentials at all, and that tells you nothing either way, since a screenshot or an upload through an app strips them out. The Content Credentials guide lists the three things you might see. A reverse image search is worth the minute it takes as well, and if the same face comes up under three different names then you have your answer.

If it is a scam

Stop talking to them, keep the transcripts, and report the profile to the platform, as your report is what protects the next person they would have tried it on. If money has already gone, call your bank or your card issuer right away and report it to your national fraud service, which in the US means the FTC and IC3. None of this needs you to prove that the messages were written by AI, since the request for money is what you are reporting.

What Cobalynx can and cannot do here

No tool, ours included, can tell you who wrote one short message, so a scan only starts to help once you have a stretch of the conversation. A picture with no credentials in it proves nothing either way. And the two rules outrank every verdict on this site: get them on a live video call early on, and never send money to someone you have not met. How our scoring works and the error rates we counted are written up on the methodology page and the evidence page.

Common questions

Can I check a single flirty message?

No, and we would not trust a tool that said it could do it, as a short message is far below the floor where any scan can be relied on, ours included, and if you paste one into Cobalynx it will tell you it was too short and will not guess. What we would do is wait until you had the whole conversation, ideally just your match’s side of it, and paste that in once it had gone past roughly 150 words, as what gives a chatbot away is that it sounds the same in every message, and you cannot see that from one message on its own.

My match sent me a message that had ** asterisks in it. What does that mean?

Quite possibly it was a bit of the chatbot showing up in the message. The chatbots write in Markdown, which is the formatting where **double asterisks** mean bold, and if someone had pasted what the bot wrote straight into a dating app then the raw asterisks would have stayed there for you to see. One on its own proves nothing, as people do use asterisks, but if it came along with replies that were instant and polished and arrived all at once as a single block, then we would take it as a strong sign that the words had come out of a model.

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

You cannot tell from that, and we would not read anything into it. Most real pictures have no credentials in them at all, and even when they did, a screenshot or a re-save would have stripped them out, so when you see none it tells you nothing. The check only works one way round, though. If the file does carry a manifest from a picture generator, then you have something close to proof that it was generated. You can check a file through the Content Credentials guide, and we would put the picture into a reverse image search too while we were at it.

What should I do no matter what any detector says?

There are two things we would do no matter what any tool told us, and we would do them early. We would insist on a live video call, and we would never send money, or gift cards, or crypto, to someone we had not met in person, as every romance scam there has ever been, machine or no machine writing the messages, got to the money request in the end, and a scan of the chat does not change either of those two rules.