Hi HN - I’m the Head of AI Research at Sword Health and one of the authors of this benchmark (posting from my personal account).
We built MindEval because existing benchmarks don’t capture real therapy dynamics or common clinical failure modes. The framework simulates multi-turn patient–clinician interactions and scores the full conversation using evaluation criteria designed with licensed clinical psychologists.
We validated both patient realism and the automated judge against human clinicians, then benchmarked 12 frontier models (including GPT-5, Claude 4.5, and Gemini 2.5). Across all models, average clinical performance stayed below 4 on a 1–6 scale. Performance degraded further in severe symptom scenarios and in longer conversations (40 turns vs 20). We also found that larger or reasoning-heavy models did not reliably outperform smaller ones in therapeutic quality.
We open-sourced all prompts, code, scoring logic, and human validation data because we believe clinical AI evaluation shouldn’t be proprietary.
Happy to answer technical questions on methodology, validation, known limitations, or the failure modes we observed.
I'm skeptical of the value of this benchmark, and I'm curious for your thoughts - self play / reinforcement tasks can be useful in a variety of arenas, but I'm not a priori convinced they are useful when the intent is to help humans in situations where theories of mind matter.
That is, we're using the same underlying model(s) to simulate both a patient and a judgment as to how patient-like that patient is -- this seems like an area where I'd really want to feel confident that my judge LLM is accurate; otherwise the training data I'm generating is at risk of converging on a theory of mind / patients that's completely untethered from, you know, patients.
Any thoughts on this? Feel like we want a human in the loop somewhere here, probably on scoring the judge LLMs determinations until we feel that the judge LLM is human or superhuman. Until then, this risks building up a self-consistent, but ultimately just totally wrong, set of data that will be used in future RL tasks.
I saw there was another benchmark where top LLMs also struggle in real patient diagnostic scenarios in a way that isn't revealed when testing in e.g. medical exams. I wonder if this also applies to law, too...
I heard a story on NPR the other day, and the attitude seems to be that it's totally inevitable that LLMs _will_ be providing mental health care, so our task must be to apply the right guardrails.
I'm not even sure what to say. It's self-evidently a terrible idea, but we all just seem to be charging full-steam ahead like so many awful ideas in the past couple of decades.
Maybe you’re comparing it to some idealized view of what human therapy is like? There’s no benchmark for it, but humans struggle in real mental health care. They make terrible mistakes all the time. And human therapy doesn’t scale to the level needed. Millions of people simply go without help. And therapy is generally one hour a week. You’re supposed to sort out your entire life in that window? Impossible. It sets people up for failure.
So, if we had some perfect system for getting every person that needs help the exact therapist they need, meeting as often as they need, then maybe AI therapy would be a bad idea, but that’s not what we have, and we never will.
Personally, I think the best way to scale mental healthcare is through group therapy and communities. Having a community of people all coming together over common issues has always been far more helpful than one on one therapy for me. But getting some assistance from an AI therapist on off hours can also be useful.
It's not inevitable that LLMs will be providing mental health care; it's already happening.
Terrible idea or not, it's probably helpful to think of LLMs not as "AI mental healthcare" but rather as another form of potentially bad advice. From a therapeutic perspective, Claude is not all that different from the patient having a friend who is sometimes counterproductive. Or the patient reading a self-help book that doesn't align with your therapeutic perspective.
Grok 3 and 4 scored at the bottom, only above gpt-4o, which I find interesting, because there was such big pushback on reddit when they got rid of 4o due to people having emotional attachments to the model. Interestingly the newest models (like gemini 2.5 and gpt 5 did the best.
It doesn't show that they "struggle". It shows that they don't behave according to modern standards. I wouldn't put much weight into an industry without sensible scientific base that used to classify homosexuality as a disease not so long ago. The external validity of the study is dubious, let's see comparison to no therapy, alternative therapy, standard therapy; and then compare success rates.
No surprises here. Its long been known that humans cannot improve their own mental health with machines - there have to be other humans involved in the process, helping.
This will become more and more of an issue as people look for a quick fix for their life problems, but I don't think AI/ML is ever going to be an effective mechanism for life improvement on the mental health issue.
It'll instead be used as a tool of oppression like in THX1138, where the apparency of assistance is going to be provided in lieu of actual assistance.
Whether we like it or not, humans are a hive species. We need each other to improve our lives as individuals. Nobody ever climbed the mountain to live alone who didn't come back down, realizing how much the rest of humanity is actually essential to human life.
This'll be received as an unpopular opinion, but I remain suspicious of any and all attempts to replace modern health practitioners with machines. This will be subverted and usurped for nefarious purposes, mark my words.
Full disclosure: after leaving tech, I’m back in grad school to get my LMHC so I’m obviously biased.
First, I just don’t see a world where therapy can be replaced by LLMs, at least in the realistic future. I think humans have been social creatures since the dawn of our species and in these most intimate conversations are going to want to be having them with an actual human. One of my mentors has talked about how after years of virtual sessions dominating, the demand for in-person sessions is spiking back up. The power of being in the same physical room with someone who is offering a nonjudgmental space to exist isn’t going to be replaced.
That being said, given the shortage of licensed mental health counselors, and the prohibitive cost especially for many who need a therapist most, I truly hope LLMs develop to offer an accessible and cheap alternative that can at least offer some relief. It does have the potential to save lives and I fully support ethically-focused progress toward developing that sort of option.
> I think humans have been social creatures since the dawn of our species and in these most intimate conversations are going to want to be having them with an actual human. One of my mentors has talked about how after years of virtual sessions dominating, the demand for in-person sessions is spiking back up.
Agreed. I used to frequent a coworking space in my area that eventually went fully automated and got rid of their daytime front desk folks. I stopped going shortly thereafter because one of the highlights of my day was catching up with them. Instead of paying $300/mo to go sit in a nice office, I could just use that money to renovate my home office.
A business trying to cultivate community loses the plot when they rely completely on automation.
It's also important to understand how bad LLMs actually are.
It's very easy to imagine that LLMs are smart, because they can program or solve hard maths problems, but even a very short attempt to have them generate fiction will demonstrate an incredible level of confusion and even an inability to understand basic sentences.
I think the problem may have to do with the fact that there are really many classes, and in fiction you actually use them. They simply can't follow complex conversations.
Do you have plans to improve the quality of the LLM as judge, in order to achieve better parity with human clinician annotators? For example, fine-tuning models?
Thinking that the comparative clinician judgements themselves would make useful fine-tuning material.
Everything in this research is simulated and judged by LLMs.
It might be hard to prove which of those LLMs struggles with exactly what.
The grounding this had was that texts produced by role-playing humans (not even actual patients) were closer to texts produced by the patient simulations prompt they decided to end up with than others they tried.
Human therapists are often quite bad as well. It took me around 12 before I found a decent one. Not saying that LLMs are better but they do theoretically have more uniform quality.
Quality is variable, but did any of those 12 encourage you to kill yourself?
If a therapist found to encourage any of their patients to self-harm would lose their license to practice and would likely face prosecution. The plagiarism machine should face the same level of scrutiny.
How many were "therapists" or "Counselor" vs actually credentialed professionals?
There's also a lot of credentialed professionals who got their credential decades ago and haven't at all kept up with the significant changes or new data over that time. This is a pretty big problem in all of medical care.
This is a 1250 word judging prompt - likely AI generated
Along with 10 scored conversation samples - all also AI generated
No verification in the field, no real data
In other words, AI scores on AI conversations - disguised as a means of gauging clinical competence / quality?
This is not an eval - this is a one-shotted product spec!
39 comments
[ 5.2 ms ] story [ 88.2 ms ] threadWe built MindEval because existing benchmarks don’t capture real therapy dynamics or common clinical failure modes. The framework simulates multi-turn patient–clinician interactions and scores the full conversation using evaluation criteria designed with licensed clinical psychologists.
We validated both patient realism and the automated judge against human clinicians, then benchmarked 12 frontier models (including GPT-5, Claude 4.5, and Gemini 2.5). Across all models, average clinical performance stayed below 4 on a 1–6 scale. Performance degraded further in severe symptom scenarios and in longer conversations (40 turns vs 20). We also found that larger or reasoning-heavy models did not reliably outperform smaller ones in therapeutic quality.
We open-sourced all prompts, code, scoring logic, and human validation data because we believe clinical AI evaluation shouldn’t be proprietary.
Happy to answer technical questions on methodology, validation, known limitations, or the failure modes we observed.
I'm skeptical of the value of this benchmark, and I'm curious for your thoughts - self play / reinforcement tasks can be useful in a variety of arenas, but I'm not a priori convinced they are useful when the intent is to help humans in situations where theories of mind matter.
That is, we're using the same underlying model(s) to simulate both a patient and a judgment as to how patient-like that patient is -- this seems like an area where I'd really want to feel confident that my judge LLM is accurate; otherwise the training data I'm generating is at risk of converging on a theory of mind / patients that's completely untethered from, you know, patients.
Any thoughts on this? Feel like we want a human in the loop somewhere here, probably on scoring the judge LLMs determinations until we feel that the judge LLM is human or superhuman. Until then, this risks building up a self-consistent, but ultimately just totally wrong, set of data that will be used in future RL tasks.
I'm not even sure what to say. It's self-evidently a terrible idea, but we all just seem to be charging full-steam ahead like so many awful ideas in the past couple of decades.
Maybe you’re comparing it to some idealized view of what human therapy is like? There’s no benchmark for it, but humans struggle in real mental health care. They make terrible mistakes all the time. And human therapy doesn’t scale to the level needed. Millions of people simply go without help. And therapy is generally one hour a week. You’re supposed to sort out your entire life in that window? Impossible. It sets people up for failure.
So, if we had some perfect system for getting every person that needs help the exact therapist they need, meeting as often as they need, then maybe AI therapy would be a bad idea, but that’s not what we have, and we never will.
Personally, I think the best way to scale mental healthcare is through group therapy and communities. Having a community of people all coming together over common issues has always been far more helpful than one on one therapy for me. But getting some assistance from an AI therapist on off hours can also be useful.
Terrible idea or not, it's probably helpful to think of LLMs not as "AI mental healthcare" but rather as another form of potentially bad advice. From a therapeutic perspective, Claude is not all that different from the patient having a friend who is sometimes counterproductive. Or the patient reading a self-help book that doesn't align with your therapeutic perspective.
You trust humans to do it. Trust has little to do with what actually happens.
Edit: Thank you!
https://www.forbes.com/sites/johnkoetsier/2025/11/10/grok-le...
Grok 3 and 4 scored at the bottom, only above gpt-4o, which I find interesting, because there was such big pushback on reddit when they got rid of 4o due to people having emotional attachments to the model. Interestingly the newest models (like gemini 2.5 and gpt 5 did the best.
This will become more and more of an issue as people look for a quick fix for their life problems, but I don't think AI/ML is ever going to be an effective mechanism for life improvement on the mental health issue.
It'll instead be used as a tool of oppression like in THX1138, where the apparency of assistance is going to be provided in lieu of actual assistance.
Whether we like it or not, humans are a hive species. We need each other to improve our lives as individuals. Nobody ever climbed the mountain to live alone who didn't come back down, realizing how much the rest of humanity is actually essential to human life.
This'll be received as an unpopular opinion, but I remain suspicious of any and all attempts to replace modern health practitioners with machines. This will be subverted and usurped for nefarious purposes, mark my words.
First, I just don’t see a world where therapy can be replaced by LLMs, at least in the realistic future. I think humans have been social creatures since the dawn of our species and in these most intimate conversations are going to want to be having them with an actual human. One of my mentors has talked about how after years of virtual sessions dominating, the demand for in-person sessions is spiking back up. The power of being in the same physical room with someone who is offering a nonjudgmental space to exist isn’t going to be replaced.
That being said, given the shortage of licensed mental health counselors, and the prohibitive cost especially for many who need a therapist most, I truly hope LLMs develop to offer an accessible and cheap alternative that can at least offer some relief. It does have the potential to save lives and I fully support ethically-focused progress toward developing that sort of option.
Agreed. I used to frequent a coworking space in my area that eventually went fully automated and got rid of their daytime front desk folks. I stopped going shortly thereafter because one of the highlights of my day was catching up with them. Instead of paying $300/mo to go sit in a nice office, I could just use that money to renovate my home office.
A business trying to cultivate community loses the plot when they rely completely on automation.
It's very easy to imagine that LLMs are smart, because they can program or solve hard maths problems, but even a very short attempt to have them generate fiction will demonstrate an incredible level of confusion and even an inability to understand basic sentences.
I think the problem may have to do with the fact that there are really many classes, and in fiction you actually use them. They simply can't follow complex conversations.
The grounding this had was that texts produced by role-playing humans (not even actual patients) were closer to texts produced by the patient simulations prompt they decided to end up with than others they tried.
If a therapist found to encourage any of their patients to self-harm would lose their license to practice and would likely face prosecution. The plagiarism machine should face the same level of scrutiny.
There's also a lot of credentialed professionals who got their credential decades ago and haven't at all kept up with the significant changes or new data over that time. This is a pretty big problem in all of medical care.
https://en.wikipedia.org/wiki/Deaths_linked_to_chatbots
Because they help more than they hurt and yield a net benefit to society.
because we feel trapped, and either don’t see a way out, or feel like we prefer death to the prospect of continuing to live a life of torture.
Also the other weirder “I’m going to reincarnate as jesus” or “a comet will carry me to heaven” hallucinatory delusions I guess
Seems to me that benchmarking a thing has an interesting relationship with acceptance of the thing.
I'm interested to see human thoughts on either of these.
The architecture and evaluation approach seem broadly similar.
In other words, AI scores on AI conversations - disguised as a means of gauging clinical competence / quality?
This is not an eval - this is a one-shotted product spec!