Most chatbot mental-health tools are built around repeated sessions, which means a meaningful share of users drop out before they get much benefit. A study from researchers at the Technical University of Munich and the University of Cambridge asked a narrower question: could a single conversation help, if it was designed carefully enough.
Key takeaways:
- Researchers Thomas Menzel, Michel Schimpf and Thomas Bohné built overit, a chatbot app designed to help users process romantic breakups, built on Claude Sonnet 4.5.
- A randomised, controlled trial with 171 US and UK participants tested a single roughly 20-minute conversation against no intervention.
- After 7 days, the app group's distress score fell from 35.3 to 26.6, versus 35.9 to 32.2 in the control group.
- After one month, the app group still scored lower (26.0) than the control group (29.0).
- 61.7% of app users reported a "sudden insight" about their breakup, versus 19.3% of the control group.
The theory behind the design
The app leans on memory reconsolidation theory: the idea that recalling a painful memory, naming the limiting belief attached to it ("I was abandoned because I am not enough"), and then offering a contradicting interpretation ("you did the best you could, but you were failed by someone you trusted") can durably change how that memory feels.
Users first completed a survey covering their breakup timing, distress level and their ex-partner's name. The conversation itself moved through four phases: open-ended questions about the breakup, identifying a self-limiting belief, offering an alternative interpretation, and finally asking what the user had learned and how they now felt. Conversations were capped at 18 turns.
Splitting evaluation from generation
The detail worth learning from sits underneath the conversation design. With every user message, the app made two separate calls to the model rather than one. First, it asked Claude to evaluate which of five milestones the last few turns had hit, such as whether a limiting belief had been identified or challenged. Only then did it generate the actual reply, using that phase assessment plus the conversation history and the original survey data.
| Step | What it does |
|---|---|
| Evaluation call | Checks recent turns against five therapeutic milestones |
| Generation call | Produces the reply, informed by the phase and full context |
This separation matters because chatbots left to respond in a single pass tend to default to sycophantic comfort, agreeing with whatever the user says. Splitting tracking from expression let the app stay analytical about progress while still generating an empathetic response, rather than collapsing into pure reassurance.
The results, and their limits
The size of the seven-day gap is the headline number, but the one-month persistence matters more for judging whether this was a real shift rather than a brief mood lift from talking to something novel. The "sudden insight" data adds a mechanism: over three times as many app users reported a specific reframing moment, and those who did tended to report feeling better afterward.
Concept to take away: evaluate, then generate. Splitting a model's response into a separate judgement step and a separate output step is a pattern with uses well beyond therapy chatbots. Any agent meant to track progress toward a goal, not just react to the latest input, benefits from assessing state before producing output, rather than doing both in one pass.
Why this matters
A single AI conversation costs a fraction of what a human therapy session costs. A 20-minute chat that meaningfully reduces measured distress, and appears to hold a month later, is a strong economic result even before considering scale. It is also a useful case study in building agents that challenge a user's framing rather than simply validating it.
This article covers a research study on a therapeutic AI tool. If you are personally experiencing distress after a breakup or any other difficult experience, consider speaking with a mental health professional or a trusted person in your life.
