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.
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.
I heard someone say that LLMs don't need to be as good as an expert to be useful, they just need to be better than your best available expert. A lot of people don't have access to mental health care, and will ask their chatbot to ask like a psychologist.
This mostly makes sense.
The problem is that people will take what you've said to mean "If I have no access to a therapist, at least I can access an LLM", with a default assumption that something is better than nothing. But this quickly breaks down when the sycophantic LLM encourages you to commit suicide, or reinforces your emerging psychosis, etc. Speaking to nobody is better than speaking to something that is actively harmful.
I've heard people say the same thing ("LLMs don't need to be as good as an expert to be useful, they just need to be better than your best available expert"), and I also know that some people assume that LLMs are, by default, better than nothing. Hence my comment.
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.
I haven't been so unlucky myself, but I know many who've had terrible first experiences with therapists and psychologists, where I'm wondering why those people even are in the job they are, but some of them got so turned off they stopped trying to find anyone else to help them, because they think most mental health professionals would be the same as the first person they sought help from.
It's also less pressure, a more comfortable environment (home vs. stranger's office), no commitments to a next session, and less embarrassing (sharing your personal issues to a computer via text is less anxiety-inducing than saying them to a person's face).
With that all said, I'm strongly opposed to people using LLMs as therapists.
In their mind, most of the times, if there is no one standing behind them when they chat with a LLM, then the conversation for most intents and purposes is private.
Obviously, us who are born with a keyboard in front of our hands, know this to not be true, and know we're being tracked constantly with our data being sold to the highest bidder. But the typical person have more or less zero concerns about this, which is why it's not a priority issue to be solved.
Should be banned. Average people have no basis to know whether drug X is appropriate for them. If your doctor thinks you need it, he'll tell you. These ads also perpetuate the harmful idea that there's a pill for everything.
However, unless we have a measure on how helpful self-help books actually are, we still don't know if they help or not.
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.
My other stance, which I suspect is probably more controversial, is that I'm not convinced that mental health care is nearly as effective as people think. In general, mental health outcomes for teens are getting markedly worse, and it's not for lack of access. We have more mental health access than we've had previously -- it just doesn't feel like it because the demand has risen even more sharply.
On a personal level, I've been quite depressed lately, and also feeling quite isolated. As part of an attempt to get out of my own shell I mentioned this to a friend. Now, my friend is totally well-intended, and I don't begrudge him whatsoever. But, the first response out of his mouth was whether I'd sought professional mental health care. His response really hurt. I need meaningful social connection. I don't need a licensed professional to charge me money to talk about my childhood. I think a lot of people are lost and lonely, and for many people mental health care is a band-aid over a real crisis of isolation and despair.
I'm not recommending against people seeking mental health care, of course. And, despite my claims there are many people who truly need it, and truly benefit from it. But I don't think it's the unalloyed good that many people seem to believe it to be.
Which is to say, your stance might not be as controversial as you think, since it was the adult take in a children's cartoon almost 60 years ago.
Peanuts is funny, but it may not be the source of wisdom you think it is.
Lucy isn't actually a psychologist which is part of the reason the "gag" is funny.
There's also the elephant in the room that mental healthcare, in particular for teens will probably just be compensating for the disease that is social media addiction. Australia has the right idea, banning social media for all goods.
Professional mental health care cannot scale to the population that needs it. The best option, like you mention, is talking to friends about our feelings and problems. I think there has been an erosion (or it never existed) of these social mental health mechanisms. There is a learned helplessness that has developed that people have lost their capacity to just be with someone that is hurting. There needs to be a framework for providing mental health therapy to loved ones that can exist without licensed professionals, otherwise LLm's are the only scalable option for people to talk about their issues and work on finding solutions.
This might be controversial but mental health care is largely a bandaid when the causes of people's declining mental health is due to factor's far outside the individual's control: loneliness epidemics, declining optimism towards the future, climate change, the rise of global fascism, online dating, addictiveness of social media and the war on our attention, etc.
We really need to get the psychology right with LLMs.
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.
The age-old problem is how to prevent that disaster and save those lives. That's not trivial. Creating an Oncological Rhubarb Authority could easily make the problem much worse, not better.
If you try to merely stop people from using LLMs as therapists (could you elaborate on what that looks like?) and call it a day, your consideration isn't extending to all the people who will do it anyways.
That's what I mean by forcing your hand into doing the work of figuring out how to make LLM therapists work even if you were vehemently against the idea.
I think you're assuming my proposed solution is to take rhubarb away from people? It's not.
Maybe you want to found the "Oncological Rhubarb Validation Association" or something. If so, that has nothing to do with me? That's just snake oil marketing. Not my field.
(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...