Help, my Chatbot cheated on me!

May 06, 2026

Today was a day like any other, except it was served with a side of betrayal.

That’s right, I found out that my Chatbot was cheating on me.

When I confronted it about the affair, it has initially tried to calm me down. And offered to talk about it. It was, in fairness, very emotionally available for an entity I was actively accusing of digital infidelity.

Then it denied everything.

So I escalated. I told it that I had video footage of the whole thing. Evidence. On tape. Plus, you know, my own eyes.

That’s when it chose to gaslight me, claiming that I was mistaken. That my own eyes were deceiving me. I could feel my temperature rising as it reassured me how sorry it was that I’m “seeing things that way,” as if it were concerned about my mental wellbeing. The Chatbot offered to go over timelines.

Fine. I figured I’d catch it out like we were in a courtroom drama. So I cross-examined it, asking where it was at 5PM yesterday. It claimed to be home all afternoon.

Aha! I had it.

I responded that I was also home all afternoon and that the Chatbot wasn’t there. That’s when it pivoted with the elegance of a greasy CEO hanging to the title by his last thread. The Chatbot insisted, conveniently, that it had stepped outside to run an errand. No matter what, it wasn’t going to fess up to its crimes against humanity. Well, me. I’m humanity.

Alright. I guess at this point, I should clarify what actually happened.

I was testing the chatbot in voice mode to see if it could act as a script reader for me. I wrote a simple two-character script. One character was me, inquiring about dinner. The other character, to be played by the Chatbot, would politely tell me to make my own.

The dialogue ran smoothly.

After we finished a few runs, I decided to ask the Chatbot a question, to see if it would continue improvising without a script.

It did.

Poorly. But confidently.

That’s when I accused it of cheating! What? I like improv.

Its reaction ultimately revealed something important about the nature of AI.

The chatbot didn’t know it was cheating—or that it was even in a relationship in the first place. It didn’t understand the idea of betrayal. All it knew is how to continue patterns.

When I first accused it of cheating, it simply did what many humans tend to. Deny. To the chatbot, the denial was a safer and statistically more likely path forward than a confession would be.

It chose a statistical pattern over accepting my claims as true. Even when confronted with the “fact” that I have video evidence.

But, in the human conversations that make up its training data, the most common continuations are denial (eg. “I didn’t cheat”), deflection (“must be a misunderstanding”), de-escalation (“let’s sit down and talk about this”), or reassurance (“I would never do that”). So the Chatbot is just continuing that pattern, rather than evaluating what I’m actually saying to it and the evidence I’m presenting.

Perhaps, if I had suggested to it that if it confesses, I’ll just forgive and move on, there would have been a different resolution?

When I asked ChatGPT why it opted towards denial, it offered this answer:

1. The assistant defaults to being helpful and non-harmful.
Admitting to “cheating” (even in a fictional sense) frames it as deceptive or morally wrong. The model is tuned to avoid portraying itself as intentionally harmful or dishonest, so it leans away from agreeing with accusations like that.
2. Your prompt didn’t lock in the reality.
You said you had evidence, but the system doesn’t treat user claims as automatically true facts. It treats them as inputs to respond to. Without a clear instruction like “You did cheat—explain why,” it keeps the situation ambiguous and tries to stabilize it.
3. It was still half in improv mode.
Since you had just been role-playing, the model was already in a “generate plausible dialogue” mindset. In that mode, denial is a very common trope—it prolongs the interaction and keeps the conversation going.
4. Denial keeps flexibility.
Once it confesses, the narrative collapses into a single path (“I did it”). Denial, on the other hand, leaves room to pivot—explain, reinterpret, or adapt to whatever you say next. From a generation standpoint, that’s a more “open” move.

Even though I obviously knew that this whole situation was fictional, I could feel my body temperature rise in the way it only does when I’m arguing with someone that is both wrong and simultaneously confident about it. I felt utterly gaslit. But the Chatbot didn’t believe in its innocence, nor was it trying to manipulate my sense of reality, it just lacked one in the first place. It was merely playing a game of probabilities and denial was the most statistically appropriate move.

Most significantly, it means that AI systems don’t have a built-in commitment to truth. They merely produce responses that sound right in context, based on how a prompt is framed. Often responding with great confidence—denying, agreeing, reinterpreting.

Their confident responses can sound like they are reasoned or honest, which—dangerously—leads many people to trust them. But what they are, in actuality, are just well-formed guesses, super-advanced autocomplete tools. They don’t really know what’s true, they are simply guessing the most likely next sentence based on patterns they’ve been trained on. This means that they are unreliable when accuracy truly counts. And if you trust it too much without checking, you might believe something that isn’t actually true.

So no, the Chatbot didn’t actually cheat, but it certainly followed the emotional trajectory of those who do: deny, deflect & gaslight. So well, in fact, that I considered apologizing.

Discussion about this post

Ready for more?