How to catch a cheater with AI (and what it actually does)

It is late, and you are three tabs deep into apps that promise to end the guessing for $19.99.

One says it analyzes message tone. One says it detects micro-expressions on video. One just says "AI" over a stock photo of a phone.

Maybe you have already pasted a screenshot into a chatbot and asked what it thinks. Maybe you got an answer that sounded clinical and certain, and you felt relief for about four minutes. Maybe you got the opposite answer and decided the model did not have enough context.

Most of what you will find tonight is either a listicle of AI cheating apps or a lecture about how you should just trust your partner.

Both skip the interesting part, which is what these tools actually compute.

Here is the frame. AI does not detect deception. It detects patterns in whatever you feed it, and what you feed it is already shaped by your suspicion.

The output feels like evidence. It is closer to an echo with good posture.

Below: what AI can and cannot do here, why no detector in 3,000 years has found a reliable tell, what an algorithmic verdict does to the person holding it, and a way to use these tools for the next week that leaves you better off instead of further in.

How to catch a cheater with AI: the honest version of what it can do

Start by splitting the claims into what is computation and what is marketing.

An AI cheating app is doing one of four things, and only one of them is difficult.

• Summarizing text you supply. Genuinely useful. A model can compress a long thread into themes and timestamps faster than you can read it.
• Scoring sentiment or "tone." Real technique, weak claim. It tells you a message reads as warm or distant. It does not tell you why.
• Searching public profiles. Sometimes accurate, often stale, usually sold with a monthly subscription you will have to cancel twice.
• Returning a deception verdict. This is the part that does not work, and it is the part the pricing page is built around.

The gap matters, so hold both halves at once.

Read the right-hand column twice. Every one of those limits is structural, not a bug that the next version fixes.

The tools that search for public profiles are a different category, and they have their own failure modes. If that is the thread you are actually pulling, how to find out if your partner is cheating online covers what those lookups return, and is Cheaterbuster legit walks through what happens when you pay one.

No lie detector has ever found a tell, and that is why the AI version is stuck

The failure is older than software.

People have been building deception detectors for about three thousand years, and the design has barely changed. Rice in the mouth, a pulse under a thumb, a needle on a chart, a neural network on a video. Every one of them measures stress and calls it guilt.

Rebecca Wilcoxson and Emma Turley, who study this at CQUniversity Australia, put the core problem plainly: these methods have a high false-positive rate, and there is no evidence that specific nonverbal behaviours reliably accompany deception. The hard part is not spotting fear. It is separating fearful innocent people from fearful guilty ones.

Sit with what that means for a model.

Machine learning needs a signal to learn. If no behavior reliably marks a lie, there is no stable thing in the data for the system to find. It will find something, because that is what these systems do. What it finds is usually the accent, the phrasing or the mood of whoever was in the training videos.

The everyday version of this is just as bad. Psychology Today's overview of deception notes that most people are terrible at detecting deception, performing no better than chance, and that there is no single behavior that indicates it. The eye-contact rule that everyone believes runs backwards: people tend to make more eye contact when lying, not less.

So the honest summary is uncomfortable and freeing at the same time.

Nobody can look at a person and know. Not you, not a therapist, not a polygraph examiner, not a model trained on ten thousand interviews.

What is available is the pattern across weeks, which is why a structured relationship assessment gets you further than any single moment ever will.

"Then what am I supposed to do, just live with not knowing?"

No. You are supposed to stop trying to read the person and start reading the pattern. One is unavailable to you. The other has been sitting in front of you for weeks.

The Confidence Laundry: what happens when a machine hands you a verdict

Here is the part almost nobody writes about, and it is the reason I would keep you away from these apps even if they worked better.

An algorithmic verdict changes what you are willing to do.

Researchers at the University of Würzburg, with colleagues in Duisburg, Berlin and Toulouse, built a machine-learning lie detector and ran a large incentivized experiment with it. The model hit 81% accuracy on their statements. People with no help managed 46.5%, and people given the algorithm's assessment reached 60.1%.

That is the part the headlines cover. The next part is the one that should change your evening.

Without the algorithm, fewer than 20% of people were willing to accuse someone of lying. Among the people who asked the algorithm for a prediction, the accusation rate ran to around 58%. Only about a third of participants requested a prediction at all, and roughly 88% of those who did followed whatever it said.

Alicia von Schenk, one of the authors, warned that this kind of tool could significantly disrupt social harmony by feeding mistrust.

Call the mechanism The Confidence Laundry.

You walk in holding a suspicion, which is an uncomfortable thing to hold because you know you might be wrong and you know what a wrong accusation costs. You hand it to a system. It hands you back the same suspicion with a percentage on it. Nothing new entered the building. What changed is that you now feel entitled to act.

That hesitation the machine washed out of you was not weakness. It was the part of you that knows the cost of being wrong about someone you love.

And notice the asymmetry in your own use of it. If the tool says "no signs of deception", almost nobody closes the tab and sleeps well. They decide it needs more context. A detector you only believe when it agrees with you is not a detector. It is a permission slip.

If the loop of checking, re-reading and searching has become the main event of your nights, obsessive thoughts about your spouse cheating is the more useful read, because at that point the tool is feeding the loop rather than ending it.

Three ways people use AI on a relationship

You will recognize yourself in one of these, probably the first.

The Prosecutor. Feeds screenshots, asks for a verdict, argues with the answer when it comes back hedged. The prompt always contains the conclusion, so the output always contains it too. Nothing is learned, and the case file grows.

The Interpreter. Asks the model what a message "really means". Gets four plausible readings. Picks the worst one, because at 1am the worst one feels like the honest one. This is the most expensive habit of the three, in hours.

The Organizer. Uses the model as a clerk. What changed, when, in what order, and what would I need to say out loud tomorrow. No private material, no verdict requested. This one actually ends somewhere.

The difference between them is not the tool. It is whether you are asking the machine to decide or asking it to arrange.

A machine can arrange. It has never once decided anything about a marriage.

How to use AI for the next seven days without making it worse

So what do you actually do tonight?

Use The Five Prompt Rule. Five things AI is genuinely good for in this situation, none of which require your partner's private data, and a hard stop after them.

Ground rule first, because it saves people real trouble. Do not put your partner's messages, photos, location history or account contents into any app. Beyond the ethics, accessing another adult's device or accounts without permission can break computer access and wiretapping laws depending on where you live, and installing monitoring software on a phone you do not own is treated seriously in a lot of places. It also hands your partner the subject of every conversation that follows. You will spend a month defending how you found out.

1. Build the timeline, from your own memory. Open a blank note. Write what you noticed and roughly when, with no interpretation. "Phone face-down since late July. Gym four nights a week starting August. Stopped telling me about work." Then ask a model to sort it by date and name the clusters. You are outsourcing organization, not judgment.

2. Separate observation from meaning. Two columns. Left: what you saw. Right: what you made it mean. Ask the model to move anything from the left column that is actually an interpretation. Most people find half their evidence migrates. That is not a defeat. It is the first honest inventory you have had.

3. Write the opening two sentences of the conversation. Not a script for an interrogation. One observation, one feeling, one question. Ask for three versions and pick the one you could say without your voice changing. "Something has felt different since the summer and I have been sitting on it. Tell me what is going on with you."

4. Rehearse the answers you are dreading. Ask the model to play the three most likely responses: defensiveness, a flat denial, and a real disclosure. Reading them in advance is the cheapest way to stop the conversation from ending in the first ninety seconds.

5. Get the structure of your own week. Sleep, food, one person you can call, one thing on the calendar that is yours. This sounds off topic. It is not. Decisions made on four hours of sleep and a search history are not decisions.

Then stop. Close the tab.

Two things sit outside the rule and are worth more than all five.

Take the assessment. The relationship assessment does not detect anything and does not claim to. It sorts what you have noticed across five areas, so you can see whether you are looking at one odd habit or a cluster that moved together in the same month. That distinction is the actual question underneath "how do I catch them", and it is one a paragraph from a chatbot cannot give you because the chatbot has only ever seen your worry.

Bring in a person. A couples therapist is not a verdict machine either, and that is the point. Their job is to make the second conversation possible after the first one goes badly. If your partner will not come, go alone. Individual counseling is the standard route for exactly this situation, and it does more for the 3am loop than any app on the market.

One line that is not optional. If any part of this involves fear, threats, or being monitored yourself, it has stopped being a question about trust. The National Domestic Violence Hotline is 1-800-799-7233. Tools that track a partner get used against people far more often than they get used to catch anyone.

What you were actually asking the machine

Strip away the apps and the prompts.

You were not asking for a probability. You were asking for permission to stop carrying a question alone, because carrying it alone is unbearable and saying it out loud makes it real.

No model can take that from you. A person can.

The tools promise to end the uncertainty. What ends uncertainty is a conversation, and occasionally an answer you did not want.

If you want the non-AI version of the same problem, how to catch a cheater covers what people actually do and what it costs them, and whether you can trust your gut about cheating deals with the feeling that started all of this.

Certainty is not a product. It never went on sale.

Sources

• Lie detection tests have worked the same way for 3,000 years and they're still hopelessly inaccurate, The Conversation
• AI in Lie Detection: Social Harmony at Risk?, Julius-Maximilians-Universität Würzburg
• Deception, Psychology Today Basics
• National Domestic Violence Hotline