In the quiet space where human intimacy meets technological possibility, researchers at the University of Lausanne asked whether a machine could hold the weight of a troubled relationship—even briefly. A randomized trial of 258 participants found that Amanda, a GPT-4o chatbot trained as a relationship therapist, improved satisfaction, communication, and well-being just as reliably as a structured writing exercise, with gains persisting two weeks later. The finding is less a triumph of artificial intelligence than a reminder that reflection itself—however it is prompted—may be the true agent of
GPT-4 Chatbot Matches Writing Task in Relationship Support Trial
Both interventions fostered the same kind of emotional and cognitive reflection.
Mark
So both the chatbot and the writing task worked equally well. That's surprising, right? I thought the whole point of a chatbot was that it could do better.
Mimi
The researchers thought so too. But what they found is that both interventions created space for reflection. The chatbot asked questions and listened; the writing task asked people to step back and see a conflict from a neutral angle. Both approaches helped people think differently about their relationship problem. In a single session, that seems to be enough.
Luke
But we should be careful here. The sample was 81% women, mostly college-educated, mostly in stable relationships. These weren't people in crisis. The average relationship confidence was nearly 6 out of 7. There's not much room for improvement when you're already doing pretty well. The effect sizes might look different in a clinical population.
Mimi
That's fair. The researchers acknowledge that. They say the non-clinical nature of the sample is a real limitation. But it also means the chatbot didn't make things worse, which is something.
Mark
What about the safety issue? That sounds serious.
Mimi
It is. Amanda missed some warning signs. When someone said they felt like "just giving up," the chatbot didn't follow up. The researchers had screened everyone for safety beforehand, so no one was actually at risk. But Amanda didn't know that. She should have asked.
Luke
And that's the core problem with deploying this in the real world. You can't screen everyone. If Amanda goes live, she'll encounter people with genuine safety concerns, and the current version isn't reliable at catching them. The researchers suggest pairing the chatbot with an "AI supervisor" that watches for risk. That's a workaround, but it's not a solution.
Mark
So what's the takeaway? Is Amanda ready to use?
Mimi
Not yet. This was a single session. The researchers want to test multi-session interventions, where the chatbot might build a stronger therapeutic relationship. They want to test it with people who actually need help. And they need to fix the safety detection.
Luke
The study is well-designed—randomized, controlled, good follow-up rate. But it's a proof of concept, not a green light. Amanda works as well as a writing task for people who are already fairly healthy and willing to engage. That's useful information. It's not yet evidence that she's ready for clinical practice.
Le Pouls
- Researchers expected the conversational warmth of a chatbot to outperform a solitary writing task, but both interventions produced nearly identical improvements across 13 of 14 measured outcomes—a result that quietly upended the study's own hypothesis.
- Amanda earned strong marks from users for usability, empathy, and the sense of genuine collaboration, yet the technical review exposed a troubling gap: the chatbot occasionally missed safety signals, including a participant who mentioned wanting to 'just give up.'
- The trial deliberately excluded anyone with active suicidal ideation or domestic violence risk, meaning Amanda has only been tested on the most stable slice of the population—leaving her readiness for real-world distress largely unproven.
- Researchers are now pressing toward multi-session formats, clinical populations, and AI safety supervisors that could monitor conversations for danger, treating this study not as an arrival but as a carefully bounded first step.
In the quiet space where human intimacy meets technological possibility, researchers at the University of Lausanne asked whether a machine could hold the weight of a troubled relationship—even briefly. A randomized trial of 258 participants found that Amanda, a GPT-4o chatbot trained as a relationship therapist, improved satisfaction, communication, and well-being just as reliably as a structured writing exercise, with gains persisting two weeks later. The finding is less a triumph of artificial intelligence than a reminder that reflection itself—however it is prompted—may be the true agent of change. What remains unresolved is whether a chatbot can sustain that gift safely, and for those who need it most.
In July 2024, researchers at the University of Lausanne enrolled 258 people in romantic relationships and gave them two ways to work through a conflict: half spoke with Amanda, a chatbot built on GPT-4o and trained to behave like a relationship therapist, while the other half completed a structured writing exercise in which they reframed a disagreement from a neutral third party's perspective. Both groups spent 20 to 30 minutes on their task. Two weeks later, the researchers returned to measure what had changed.
The results were remarkably consistent. Thirteen of fourteen outcome measures improved significantly in both groups—relationship satisfaction, communication patterns, confidence in handling future conflicts, and personal well-being. The gains appeared immediately and held at follow-up. The chatbot group and the writing group looked nearly identical on the scorecard, which contradicted the researchers' original expectation that Amanda's interactive, responsive nature would give her a meaningful edge over a solitary exercise. What seemed to matter most was the act of reflection itself—the turning over of a problem, the reframing of it—regardless of the medium.
Amanda did show one narrow advantage: participants who spoke with her reported lower levels of the partner-demand/self-withdraw dynamic immediately after the session, while the writing group reported lower overall distress. Both were meaningful, just different. On feasibility, Amanda performed well—users rated her 4.19 out of 5 for usability and 4.75 out of 6 for working alliance, and independent reviewers confirmed high empathy and contextual coherence in the transcripts.
Limitations, however, were real. Amanda sometimes fell into repetitive patterns, and more seriously, she occasionally failed to flag potential safety concerns. When one participant mentioned feeling like 'just giving up,' the chatbot did not probe further. The study had screened out high-risk individuals beforehand, so no one was in danger—but Amanda had no way of knowing that. The sample itself was also notably stable: mostly women, college-educated, living in the UK, with relationship confidence near the top of the scale and distress well below clinical thresholds.
The researchers are candid about what comes next. Amanda must be tested across multiple sessions, where her capacity to build a sustained therapeutic alliance might finally distinguish her from a writing prompt. She must be tested with people in genuine crisis, not paid participants who are already relatively well. And the safety gap must be closed—through better training, or through an AI supervisor capable of recognizing risk and intervening. For now, Amanda stands as a proof of concept: a large language model can deliver a single session of relationship support as effectively as a well-designed writing task. Whether it can do more, and do it safely, is the question that remains.
In July 2024, researchers at the University of Lausanne ran a straightforward experiment: they gave 258 people in romantic relationships a choice between two ways to think through a conflict. Half talked to Amanda, a chatbot built on OpenAI's GPT-4o language model and trained to act like a relationship therapist. The other half completed a structured writing exercise—a well-established technique where people reframe a disagreement from a neutral third party's perspective. Both groups spent 20 to 30 minutes on their task. Two weeks later, the researchers checked in again.
The results were striking in their consistency. Thirteen of fourteen measured outcomes improved significantly in both groups. People reported higher relationship satisfaction, better communication patterns, stronger confidence in handling future conflicts, and improved personal well-being. These gains appeared immediately after the intervention and held steady or continued to improve at the two-week follow-up. The chatbot group and the writing group looked nearly identical on the scorecard.
This finding contradicted the researchers' initial hypothesis. They had expected Amanda to outperform the writing task, reasoning that the interactive, responsive nature of a conversation would engage people more deeply than a solitary writing exercise. But it didn't work out that way. Both interventions fostered the same kind of emotional and cognitive reflection—the turning over of a problem in one's mind, the reframing of it—and that reflection alone seemed to be what mattered. The single-session format may have been too brief to reveal the chatbot's potential advantage: building a therapeutic alliance over time, the kind of sustained relationship that might keep people engaged through longer treatment.
Where Amanda did show a measurable edge was narrow and specific. Participants who spoke with the chatbot reported lower levels of one particular destructive communication pattern—the partner-demand/self-withdraw dynamic, where one person pushes for discussion while the other retreats—immediately after the session. The writing group, by contrast, reported lower overall distress about their issue. Both were wins, just different ones.
On the feasibility side, Amanda performed well. Users rated the chatbot 4.19 out of 5 for usability, 3.99 out of 5 for therapeutic skills, and 4.75 out of 6 for working alliance—the sense that they and the chatbot were collaborating toward shared goals. Researchers who coded the actual conversation transcripts confirmed these ratings. They found high levels of empathy, good contextual awareness, and coherent dialogue. Amanda asked relevant follow-up questions, reflected what people said back to them, and offered validation. The chatbot stayed focused on the person's specific situation rather than veering into generic advice.
But limitations emerged in the technical review. Amanda sometimes repeated herself or fell into a formulaic pattern: reflect, validate, ask a question. Some users noticed this mechanical quality. More concerning, the chatbot occasionally failed to flag potential safety issues. In one case, a participant mentioned feeling like "just giving up," and Amanda did not probe deeper or express concern. The researchers had screened out anyone with active suicidal ideation or domestic violence before enrollment, so no one was actually in danger. But Amanda didn't know that. She should have checked in anyway.
The study enrolled 615 people initially. Researchers excluded 191 who reported risk factors too high for a self-guided intervention—suicidal thoughts, domestic violence, frequent emotional abuse. Another 166 dropped out or failed to complete the study. The final 258 were mostly women (81%), white (79%), married (51%), living in the UK (88%), and college-educated (76%). Most had used chatbots or virtual assistants before. The sample was well-resourced and relatively healthy; average relationship confidence was nearly 6 out of 7, and distress levels were below clinical thresholds. This matters. A tool that works for people who are already fairly stable may not work the same way for couples in crisis.
The researchers acknowledge this gap. They note that future work must test Amanda in multi-session formats, where the chatbot's ability to build alliance over time could matter more. They must test it with people who are actually seeking help for serious relationship problems, not paid research participants. And they must solve the safety problem—either by training Amanda to better recognize and respond to risk, or by pairing her with an "AI supervisor" that monitors for danger and takes over the conversation when needed. Until then, Amanda is a proof of concept: a large language model can deliver a single session of relationship support as effectively as a well-designed writing task. Whether it can do more, and do it safely, remains to be seen.
Citations marquantes
Both interventions fostered emotional and cognitive reflection, which may be equally impactful in short-term, single-session formats.— Study discussion
The chatbot's ability to identify and respond to safety risks remains a critical concern, as Amanda did not consistently flag potential warning signs even when participants expressed concerning statements.— Study implications
Envie de l'histoire complète ?
Lire l'original sur PLOS (Public Library of Science) ↗
Science & Space