New benchmark shows top LLMs struggle in real mental health care

Did you use the same prompts for all the models, or individualized prompts per model? Did you try a range of prompts that were very different from each other, if you used more than a base prompt?

I'm sure it's somewhere in the details somewhere, but after a quick skim I didn't find anything outlined about how you managed and used the prompts, and if it was per model or not.

Thanks a bunch for being open to answering questions here, and thanks for trying to attack this particular problem with scientific rigor, even if it's really difficult to do so.

Have you seen the feeling great app? It's not an official therapy app but it's based in TEAM CBT, made by David Burns and team.

Burns is really into data gathering and his app is LLM based on the rails of the TEAM process and it seems to be very well received.

I found it simple and very well done - and quite effective.

A top level comment says that therapists aren't good either - Burns would argue that mainly no one tests before and after and so no measuring effect is done.

And of people I know who see a therapist, practically none can tell me what exactly they are doing or what methods they are doing or how anything is structured.

Did the real clincians get all 6's in this test?

Thanks for open sourcing this.

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.

People have killed themselves because of therapy provided by humans, as well as psychiatric medication prescribed by humans. Why are we doing those?

Because they help more than they hurt and yield a net benefit to society.

From that wikipedia link, it looks like most of the deaths were of people who already had severe mental illness.

Playing devil's advocate, many people die using all kinds of tools. It doesn't make the tools any less useful for people who use them responsibly.

That said, the idea that a pattern recognition and generation tool can be used for helping people with emotional problems is deeply unsettling and dangerous. This technology needs to be strictly regulated yesterday.

Same reason we’ve always done this -

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

> 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.

Exactly. We don't do claims about humans. But there is room for improvement on current LLMs... For researchers to be able to improve LLMs we first need to know how to evaluate them. We can only improve what we can measure so we studied how to measure them :)

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.

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.

Sure, uniformly zero as far as anyone knows.

Forget about calling it mental healthcare or not: Most people end up dealing with people in significant distress at one point or another. Many do it all the time even when they aren't trained or getting paid as mental health professionals, just because of circumstances. You don't need a clinical setting for someone to tell you that they have suicidal ideation, or to be stuck interacting with someone in a crisis situation. We don't train every adult in this, but the more you have to do it, the more you have to learn some tools for at least doing little harm.

We can see an LLM as someone that talks with more people, for more time, than anyone on earth talks in their lifetime. So they are due to be in constant contact with people in mental distress. At that point, you might as well consider the importance of giving them the skills of a mental health professonal, because they are going to be facing more of this than a priest in a confessional. And this is true whether someone says "Gemini, pretend that you are a psychologist" or not. You or I don't need a prompt to know we need to notice when someone is in a severe psychotic episode: Some level of mental health awareness is built in, if just to protect ourselves. So an LLM needs quite a bit of this by default to avoid being really harmful. And once you give it that, you might as well evaluate it against professionals: Not because it must be as good, but because it'd be really nice if it was, even when it's not trying to act as one.

> It's self-evidently a terrible idea

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.

The mental health version of "AI is here to stay, like it or not you have to use it" that some people keep trying to tell me in software.

Do you have some better alternatives for a country where private mental health care costs €150/hr, while the government/insurance paid care have 3-6M+ waiting lists?

I wonder why people use LLMs as a mental health provider replacement.

It’s a trivial claim that people are going to use AI as a therapist. No grumbling is going to stop that.

So it’s sensible that someone out there is evaluating its competence and thinking about a better alternative for these folks than yoloing their worst thoughts into chatgpt.com’s default LLM.

Everyone's hand is being forced by the major AI providers existing.

Even if you were a perfect altruist with a crusade against the idea of people using LLMs for mental health, you could still be forced to dash towards figuring out how to build LLM tools for mental health in your consideration for others.

I'll argue the opposite.

(1) The demand for mental health services is an order of magnitude vs the supply, but the demand we see is a fraction of the demand that exists because a lot of people, especially men, aren't believers in the "therapeutic culture"

In the days of Freud you could get a few hours of intensive therapy a week but today you're lucky to get an hour a week. An AI therapist can be with you constantly.

(2) I believe psychodiagnosis based on text analysis could greatly outperform mainstream methods. Give an AI someone's social media feed and I think depression, mania, schizo-* spectrum, disordered narcissism and many other states and traits will be immediately visible.

(3) Despite the CBT revolution and various attempts to intensify CBT, a large part of the effectiveness of therapy comes from the patient feeling mirrored by the therapist [1] and the LLM can accomplish this, in fact, this could be accomplished by the old ELIZA program.

(4) The self of the therapist can be both an obstacle and an instrument to progress. See [2] On one level the reactions that a therapist feels are useful, but they also get in the way of the therapist providing perfect mirroring [3] and letting optimal frustration unfold in the patient instead of providing "corrective emoptional experiences." I'm going to argue that the AI therapist can be trained to "perceive" the things a human therapist perceives but that it does not have its own reactions that will make the patient feel judged and get in the way of that unfolding.

[1] https://en.wikipedia.org/wiki/Carl_Rogers

[2] https://en.wikipedia.org/wiki/Countertransference

[3] why settle for less?

[4] https://www.sciencedirect.com/science/article/pii/S0010440X6...

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.

Most likely future is that babies born right now during their upbringing will hear and see more words produced by LLMs than by humans.

We really need to get the psychology right with LLMs.

[flagged]

People constantly amazed that a machine can outperform a 24 year old charging $250/hour. Especially when the 24 year old seems incapable of calculating compound interest on their student loan deferrals. Surely this 24 year old who cannot use a formula 14 year old can will have wisdom to share. Iona Potapov talks to horse, modern man talks to machine, man with more money than sense talks to young graduate with no life experience about his struggles. All do equally well: 4 on LLM benchmark for mental health.

humans can be sued. what about AI? or even a commercial software?

Why?

An association algorithm can also never replace free thought.

Is that really true? Dr. Patric Gagne is a diagnosed sociopath and also a successful clinical psychologist. She claims her lack of empathy for patients allows her to help them solve their problems in an objective manner. I don't have any personal experience in that area and don't know if she's correct but it seems plausible.

https://patricgagne.com/

yep yep. Its something we have to study and its likely we can improve the LLM as a Judge further.

Same thing for the patient LLM. We can probably fine-tune an LLM to do a better job at simulating patients.

Those two components of our framework have space for improvement

Cite the ArXiv paper for now.

This is the real question. Not to rehash the self-driving cars arguments that have been had to death, but with potential LLM mental healthcare the question "but what if it causes harm in some interactions" is asked much, much more than with human mental healthcare professionals.

(And I'm not being theoretical here, I have quite a bit of experience getting incredibly inadequate mental health care.)

I've known quite a few people who went to therapy and I'm not sure that's even the right question to ask. I don't think they were paying to get helped as much as they were just paying to have someone to talk to. To be clear, there are people who genuinely need help, but for most, a therapist is probably just a substitute for a close friend / life coach.

And say what you will about this, a paid professional is, at the very least, unlikely to let you wind yourself up or go down weird rabbit holes... something that LLMs seem to excel at.

As I sometimes repeat on HN, Dr David Burns started giving his patients a survey at the start and end of every session, to rate how he was doing as the therapist and to rate their feelings, on a scale of 1-5.

Reasoning that if he's not good it would show up in patients thinking he's bad, and not feeling any better. And then he could tune his therapy approaches towards the ones which make people feel better and rate him as more understanding and listening and caring. And he criticises therapists who won't do that, therapists who say patients have been seeing them for years with only incremental improvements or no improvements.

Yes there's no objective way to measure how angry or suicidal or anxious someone is and compare two people, but if someone is subjectively reporting 5/5 sadness about X at the start of a session and wants help with X, then at some point in the future they should be reporting that number going down or they aren't being helped. And more effective help could mean that it goes down to 1/5 in three sessions instead of down to 4/5 in three years, and that's a feedback loop which (he says) has got him to be able to help people in a single two-hour therapy session, where most therapists and insurance companies will only do a too-short session with no feedback loop.

I hear you, but I feel it's also important to differentiate between the kinds in which humans and LLMs can be quote-unquote "bad".

"Good" is too broad and subjective to be a useful metric.

It is still debated if therapies even work. The evidence is moving to the direction that they don't.

They are high-pay LLM wrappers dressed as humans

Less bad at least.

mostly because they were trained to say yes