Multi-agent AI vs chatbot: what the MIND study found
Editorial correction, 24 September 2026. The earlier headline and text incorrectly presented 42% as an advantage of multi-agent AI over a regular chatbot. We have corrected the metric, experiment description and conclusions. This is an editorial correction, not a new expert or clinical review.
What 42% actually means in MIND
In MIND, version 2, Chen and colleagues, 2025, the authors report an average 42% decline in ablation evaluations that removed the memory mechanism, Guide or Strategist from their system. Four clinical psychology researchers scored generated examples on six content metrics, including coherence (whether the story stays logical and consistent). Section 4.2 and Appendix E describe these evaluations.
The comparison is between the complete MIND system and modified versions of it. It does not mean that MIND is 42% better than every chatbot, outperforming a full course of psychotherapy, or demonstrating a benefit from Nearby.
An ablation answers a narrow question: what changes in this implementation when a selected component is removed? It cannot establish a universal rule that more agents produce better support. Different tasks, models, prompts or evaluation methods could produce different results.
How the research prototype works
MIND has four main roles: Trigger creates a situation, Devil voices a distorted thought, Guide supports reframing, and Strategist directs the story. A simulated patient is used in evaluation; it is not a fifth required agent in the user's dialogue.
Dividing tasks gives researchers a way to change one component and inspect its contribution. The roles are software instructions, rather than professionals participating in the conversation.
What the study allows us to conclude
A separate experiment involved eight students aged 18–21, but only two used MIND. Two used EmoLLM, two used CACTUS and two formed a no-conversation control group. Each dialogue participant completed five rounds of conversation; PANAS questionnaires captured mood before and after. Participants needed a concern that had bothered them for one day to one week. Appendix C.1 gives the allocation and procedure.
The study excluded people currently experiencing mental illness and those at risk of self-harm or suicide. That leaves a small experiment about immediate mood and user experience in a selected student group, rather than evidence about treatment for a clinical population. Ethics Statement.
Three questions need to stay separate:
- Response quality: does a conversation seem coherent, understandable and supportive?
- Change in emotional state: did a participant's self-report change after a conversation?
- Clinical effectiveness: does a particular intervention help a defined group against predetermined criteria over sufficient time?
An answer to the first or second question cannot replace investigation of the third. This is why this article has no table of effectiveness percentages comparing AI with a human therapist: the MIND results do not support that table.
What this means when choosing an app
Evaluate the specific product: is it clear that AI is responding, are its limits explained, and can you read its data policy and payment conditions? Conversation history and structure depend on the app's implementation. It would be inaccurate to claim that all ordinary chatbots lack memory or always lose focus.
Nearby was not evaluated in the MIND study. Shared ideas or CBT-informed features do not transfer another system's results to our product. Our Evidence and safety page collects the available information about access, data handling and limits.
Frequently asked questions
Is a multi-agent system always better than a single chatbot?
The study does not establish that general conclusion. The system, task and evaluation criteria matter. Agent count is not a measure of support quality.
Can 42% be read as an improvement in wellbeing?
No. It describes evaluation results for MIND variants with components removed. It is not a recovery rate, symptom reduction or a measure of Nearby's effectiveness.
What would a stronger outcome study need?
A larger, clearly defined participant group, an appropriate comparison and follow-up beyond the conversation. The MIND authors themselves call for further clinical trials with real mental-health patients.
Source
Chen, Y., Li, C., Wang, Y., et al. (2025). MIND: Towards Immersive Psychological Healing with Multi-Agent Inner Dialogue. Primary publication, arXiv v2. Checked on October 9, 2026.