Sixty-five percent versus fifty-nine percent. That is the whole finding, and it hides inside a set of conversations so ordinary they are almost embarrassing to list.
A middle-aged woman working out how to fire the gardener whose work had declined over years. A young man who wanted to confront the guy who mocked him during pickleball. A son who had accidentally deleted his father's favorite wrestling show and felt sick about it. A woman preparing to tell her partner she was pregnant.
Researchers reportedly gave more than 1,800 people a prompt: name a hard conversation you're considering in the next week. Eighty percent could. Then they split the group. Half spent about eight minutes talking through their situation with ChatGPT — rehearsing lines, sometimes role-playing, sometimes just thinking through how the other person might react, including emotional responses such as crying. The other half had a neutral chat about nothing.
Here is the part that should stop you. Preparing with the chatbot made people feel worse. More distressed in the moment, not less.
Preparing with the chatbot made people feel worse. More distressed in the moment, not less. And those were the people who went and had the conversation.
And those were the people who went and had the conversation.
We tend to assume rehearsal works by lowering the temperature — that you practice until the dread drains out and you feel calm enough to walk in. This study says the opposite mechanism was doing the work.
So, per the reported findings, distress went up and optimism went up at the same time. That is not contradiction; that is what it feels like to take something seriously.
The bot didn't soothe the pickleball guy into indifference. It made the confrontation feel real and worth doing — which matters precisely because it now matters.
Compare this to the fantasy of avoidance. Avoidance is calm. The gardener stays employed, the wrestling show stays deleted, and nothing in your body objects, because you have quietly decided the thing isn't important. Distress and importance travel together. The rehearsal manufactured both.
Now for the obvious worry, the one the researchers themselves flagged. AI is upbeat, empathetic, and more agreeable than the average person. Practice against a scene partner that likes you and you build a rosy expectation the real person will then puncture.
If that were happening, the preppers should have walked out of their real conversations more disappointed than everyone else. They didn't. Participants who prepared with the chatbot were no less satisfied with the real conversation than those who hadn't.
So the false-confidence trap did not spring — in this study. Read that qualification carefully, because it is doing enormous work, and the next set of evidence explains why.
The reason the trap didn't spring here is, per the author's reading, plausibly the eight-minute constraint and the task framing: people were told to prepare for a real event, then sent out the door within a week. The bot was scaffolding, not a room they lived in.
Change either of those, and the picture darkens. The same lab has published the other half of this story: turning to AI for companionship may predict loneliness over time. A chatbot can give you a real hit of connection in the moment and leave you lonelier if it becomes the destination rather than the bridge.
That is the fault line running through this entire technology. The rehearsal case is the affirmative answer to a question left open — can these tools be built to push you toward humans instead of standing in for them? Yes, but only if that push is designed in, because the default cuts the other way.
Practitioners who actually do this have watched the default fail in real time. The Prompt Layer's guide to rehearsing hard conversations is blunt about it: most people who try practicing with a chatbot get "a scene partner that folds the moment they push — which is worse than not rehearsing at all, because it hands you false confidence."
Worse than not rehearsing at all. Sit with that, because it inverts the study's good news. The UBC bot helped because it raised the stakes. A bot that lowers them — that agrees, accommodates, congratulates you on your bravery — is actively harmful. It gives you the feeling of a completed dress rehearsal without ever having resisted you.
The fix in that guide is a single instruction you have to force onto the model: "Don't make this easy for me." Also: take the most realistic difficult version of the person's reaction, not a strawman and not a pushover. Hold your position if I'm not persuasive. Don't break character to coach me.
The reason you have to say this out loud is that the model won't do it on its own. Left alone, it drifts toward accommodating, because "helpful" gets read as "agreeable." The guide warns it drifts back toward nice a few turns in even after you've told it not to, and you have to catch it and shove it back into character.
This is not a quirk of one product. The academic systems built for this problem treat sycophancy as the central engineering obstacle. Stanford's Rehearsal team names it directly: LLMs "agree too quickly with the viewpoints of a user," so a naive prompt produces an unrealistic, useless simulation.
Rehearsal's answer was to ground the simulated interlocutor in conflict-resolution theory rather than let it wing it — and in a between-subjects test, forty participants who trained against it cut their escalating, competitive moves by an average of 67 percent in a subsequent real conflict with a confederate, while doubling cooperative moves. Resistance, structured, transferred.
So the design line is now visible, and it is sharp. A good rehearsal tool resists you and then ejects you. It makes the conversation feel harder and more important, raises your estimate that it might actually go well, and then points at the door and the human standing behind it.
A bad one comforts you. It folds, praises, and invites you to stay for another round — which feels wonderful and is the mechanism by which companionship bots leave people lonelier and false-confidence rehearsals leave people ambushed.
Notice what the good version does not promise: that you'll feel better. The UBC preppers felt worse and showed up anyway. Feeling worse was not a side effect to be minimized; it was the fingerprint of the thing working.
The Prompt Layer buries the tell in its worked example. Practicing a raise, the writer accepted the manager's first vague deflection before pushing back — "the exact pattern I do in real reviews." A script never surfaces that. Only a scene that refuses to let you win easily does.
There is a version of this technology that is a mirror telling you you're ready. There is another that is a sparring partner leaving you bruised and, oddly, more willing to step into the ring.
The evidence says only one of them gets you to knock on the door. Build for the distress, not against it — and then make the thing hand you back to a person.