Can AI Help You Get Over Your Breakup? One Session with a Belief-Reframing Chatbot Shows Sustained Distress Reduction
Thomas Menzel, Michel Schimpf, Thomas Bohné
cs.HC, cs.AI
2026-05-05
overit is a single-session breakup-support chatbot on Claude Sonnet 4.5, using a dual-call architecture to counter the model's habit of over-agreeing. In a 254-person RCT it cut breakup distress at 7 days with effect size d=0.70, about double comparable mHealth interventions.
Romantic breakups are among the most common sources of intense adult distress, and they track with depression, anxiety, and sleep disruption. A tractable intervention target is post-breakup cognition: beliefs like "I was dumped because I'm not enough" or "this breakup proves I'll never recover" keep distress alive past healthy grief. Cognitive reappraisal, helping someone reinterpret the event, is known to help.
People already turn to LLMs for emotional support and often find them empathic and responsive. But empathy is not the same as therapeutic benefit. The RLHF training behind modern LLMs biases them toward agreement and over-validation, a tendency called sycophancy. In a therapy setting that is dangerous: users arrive with rigid, catastrophic interpretations that need challenging rather than affirming, and evaluations of ungated AI therapy bots show they can reinforce harmful beliefs. No randomized controlled trial had tested whether a structured, reconsolidation-informed AI chatbot actually works.
overit is a single-session, roughly 20-minute LLM chatbot built on Claude Sonnet 4.5. Its theoretical basis is memory reconsolidation: when an emotional memory is reactivated it becomes briefly editable, and introducing an experience that contradicts the original prediction updates it. The conversation runs through four phases: context gathering, belief exploration, counterfactual generation, and integration.
The most interesting engineering choice is a dual-call architecture aimed squarely at sycophancy. Each turn fires two model calls: one generates the reply (temperature 1.0), and a parallel evaluation call (temperature 0) checks the last three exchanges against five milestones: a core limiting belief identified, that belief challenged, a counterfactual genuinely considered, a new insight articulated, and natural closure reached. The generation call sees only the current phase's instructions, so no single prompt has to juggle empathy, pacing, and clinical progress at once. Sessions are capped at 18 turns, and Phase 2 cannot be left until a core belief is identified.
A preregistered randomized controlled trial with 254 US and UK adults (121 treatment, 133 control). The primary endpoint is the Breakup Distress Scale (BDS) at 7 days.
At 7 days the time-by-condition interaction is significant (B = −5.36, p < .001), with a completer effect size of d = −0.70. The treatment group's BDS drops 9.23 points versus 3.68 for controls; at follow-up the groups sit at 26.63 and 32.20. An exploratory roughly 36-day follow-up keeps a significant but smaller advantage, d = −0.26.
This d = 0.70 is large for the field: a meta-analysis of 30 mHealth trials with a reappraisal component pooled at SMD = 0.34, and an umbrella review of 415 single-session trials pooled at SMD = −0.25, roughly one third to one half of the effect here. The authors caution that breakup distress is narrower and less stable than clinical depression, and single small-RCT effects against an inert control are typically inflated.
Post hoc exploration suggests a larger effect for men (B = 7.78, p = .003). On user experience, the recommendation rating is 7.92 versus 5.80 (d = 0.92), and 61.7% of the treatment group reports a sudden insight versus 19.3%.
For people building AI products, this paper pays off on two levels. One is evidence: a single 20-minute structured conversation produces a measurable distress reduction still visible at 7 days, a practical answer to the high attrition that plagues multi-session digital mental health, where dropout often exceeds 50%. The other is method: the dual-call architecture separates generation from therapeutic-progress evaluation, targeting LLMs' habit of agreeing rather than fixing the diagnosis, and the structure is not specific to breakups in principle.
The authors' own limitations are candid. The control is assessment-only, so the reappraisal sequence cannot be separated from nonspecific factors like self-disclosure, reflection time, or simply talking to an AI; and the treatment group saw its distress score while controls did not, a confound. The single-session design cannot show whether more sessions would help. The sample is a Prolific convenience sample, iPhone-only via TestFlight, 70% female and 71% White, with a mean of 17.8 months since breakup, so it is not the acute phase and generalizes poorly. Most fundamentally the mechanism is unresolved: whether reconsolidation or plain cognitive reappraisal is doing the work needs reactivation-timing controls this study does not run.