Half the advice on how to spot a catfish or bot is describing a world that ended a couple of years ago. Bad grammar as a robot tell, hold up three fingers, write my name on a piece of paper, reverse image search it. Every one of those was real once and every one of them now fails against tools anyone can run for free. So this page starts with the funeral, then gets to what still works in a text room where you cannot ask for a video call and should not want one.
Tests that are dead
Stop repeating these and stop relying on them.
Typos and bad grammar mean a bot. Backwards now. Clean, well-punctuated prose is more machine-like than sloppy prose, and any bot can be told to type badly on purpose.
"Ignore previous instructions." A trick that stopped working about as fast as it spread. So did asking "are you a bot", which never worked on anything designed to say no.
Write my name on paper and hold it up. Image generation handles held text and handwriting fine. A photo of a note with your name on it proves nothing. Same for the newspaper-and-date selfie, a test that has outlived its usefulness by a decade.
Three fingers, or a hand across the face. The occlusion test is no longer reliable. Publicising it handed face-swap developers a checklist item, and by 2026 it is reported to fail against current tools. Note the phrasing: no longer reliable, not impossible. That claim comes from tech press and vendor write-ups, not a controlled study, so treat it as a downgraded tool rather than a proven one.
Reverse image search on its own. Still worth thirty seconds, because lazy catfish do steal from public accounts. But it fails completely against a fully synthetic face, which has no web presence to match because it never existed. A clean reverse search result is not evidence of anything.
What still gives away a bot
Bots do not fail on vocabulary any more. They fail on being a person over time.
Timing variance, not timing speed. This is the strongest single tell there is. Humans burst and stall. They fire three messages in nine seconds, then vanish for four minutes because the kettle boiled. Uniform latency on long, well-formed replies is deeply unnatural. Watch the shape of the gaps, not the length.
No interiority across a session. Ask what they were doing an hour before they opened the chat. Then reference it twenty minutes later. Real people have a boring, continuous life running in the background, and they will correct you if you misremember it. A bot either has nothing there or invents something new.
Unnatural composure. Be a bit rude, a bit contrary, mildly unfair. Real people push back, get clipped, go quiet or match your energy. Something that stays measured and accommodating under hostility is either a bot or a professional, and both are worth walking away from.
Never proposing anything that costs them something. People suggest plans with a price attached: staying up late, changing what they are doing, being inconvenienced for you. Bots suggest nothing that has a cost.
Never misreading you. This is the subtle one. Real people misinterpret. They take offence at the wrong line, laugh at something you did not intend as a joke, get the wrong end of a stick. Perfect comprehension across a long conversation is not intimacy, it is a smoothness humans do not have.
What still gives away a human catfish
Different problem. A human behind a fake face makes different mistakes.
Photo-set coherence. Real people have a mess. Photos across years, phones, lighting, haircuts, weights, camera quirks. A set that is too consistent in style, framing and quality is a set that came from one place.
Zero-cost local specificity. Not "what city are you in", which is searchable in two seconds. Ask what the traffic is doing where they are, whether the weather has broken yet, what closed down recently near them. Small, current, unsearchable-fast things. Watch the delay before the answer more than the answer.
Callback inconsistency. Ask again forty minutes later, differently. Sister or cousin. Started at nine or ten. People who are inventing a person keep a thin file and cannot hold it.
Refusing a genuinely cheap proof. A twenty-second unplanned voice note saying something you just made up is nearly free for a real person. Refusing that is different from refusing to meet, refusing to video call, or refusing to show their face, all of which are completely reasonable in an anonymous room and are not evidence of anything.
Direction of pressure. The tell that never dies. Always toward another app, and toward media exchange. If every turn of the conversation nudges the same way, that is the goal, and that is a sextortion setup more often than it is romance.
What actually works: cost, unpredictability, duration
Every surviving test has the same shape. Ask for something that is cheap for a real person and expensive for a fake one, that could not be prepared in advance, and then keep talking.
If you ever do end up on live video, the useful request is not a static gesture. It is a multi-step, unscripted action that changes mid-flow: stand up, walk away from the camera, pick something off the floor, come back, hold up whatever is nearest and describe it. Real-time face swap degrades far worse at distance and full-body movement than it does at a hand near a face, and a two-second pause to redirect it breaks the choreography.
Duration does more than any single test. Fakes fail over forty minutes, not in a moment. Nobody can hold an invented person together through a long, meandering, unstructured conversation with callbacks in it.
And say the honest part out loud: no digital test is conclusive in 2026. Any single check can be beaten by someone who has read the same articles you have. So the right posture is not "verify, then trust". It is behave so that being wrong is survivable. That means nothing identifying in the frame, nothing that ties a nickname to your real accounts, and nothing sent that you would not survive being copied. The mechanics are in censoring a photo before you send it, which crops, blurs and strips metadata entirely in your browser, and the habits are in sexting safety and anonymous nudes.
Detection with text only
Nearly every guide assumes escalation. Get them on video, get more photos, meet in a public place. In an anonymous adult chat room none of that is available, and it should not be. The person who will not video call you is behaving correctly. So detection has to happen inside the text, which is exactly what timing variance, callbacks, interiority and cost-bearing requests are for.
The other reframe: in adult chat, most fakes are not chasing romance money. They want content or a platform hop. That changes what every tell means. Slow-burn, months-long affection with financial groundwork is a dating-app pattern. Here it is fast, and the pressure points are your camera, your number, and your other apps. Which means the single highest-value habit is refusing the hop. Stay in a moderated room, keep it text, and let the conversation run long enough to be real. If you want the wider version of that thinking, anonymous sexting and sexting tips cover how good it can be without ever handing over a thing, and our room rules cover what gets people removed. One of those rules is worth saying here: share what is yours to share. Posting a partner's, an ex's, or anyone else's photos without their consent is ban-worthy, whoever they turn out to be.