Hacker News Reader

Comment by jes | Hacker News Reader

I am hopeful that we will soon see AI-based psychotherapy available at scale at a price such that anyone who wants it will be able to get it. I can imagine that it would be more or less like having a telephone conversation with a real therapist, except in this case the therapist is an AI application running in a datacenter.

I think we will see something like this in the next five to ten years.

Thoughts?

View on HN

It would be great if that were possible, but a lot of the value of therapy comes from the therapeutic relationship rather than a specific type of therapy. Most types of psychotherapy tend to have the same level of effect in studies, suggesting there is no "recipe" for the type of therapy that works. This phenomenon also suggests (but does not prove) that the human relationship, rather than therapeutic program, accounts for the measured improvement in outcomes

Not to say an AI tool wouldn't be helpful for some people, but it isn't clear it would be that useful. And i don't know of any evidence or plausible hypotheses that an AI tool could work for anything other than mild anxiety or depression

AI + human therapy could potentially be better than human therapy alone however

I thought Cognitive Behavioral Therapy was actually the only kind of talk therapy that showed greater effect than placebo.

Cognitive behavioral therapy is a more standardized and measurable type of therapy than things like interpersonal therapy or problem solving therapy or psychodynamic therapy, and thus it is easier to study in a controlled setting. So there are more studies of CBT than other types of therapy just because it is easier to study

I don't have references offhand but there are controlled studies that show benefit of other types of therapy

Overall all though, all psychotherapy studies are messy

That makes sense. If you want to go beyond self reporting, you have to use very long term measures like employment, incidence of suicide, etc. And of course it's very hard to guarantee that an individual pursued and sustained a specific type of therapy for a substantial amount of a long term study period.

The idea you're raising, the differential efficacy of CBT, is a myth originating in the politics of psychiatry and psychology, especially in the 70s and 80s. There is abundant evidence for psychotherapy, but no evidence that any type is better than any other, especially when you control for publication bias. It's the "common factors" phenomenon, where therapies that have some baseline characteristics all do about equal. It's well documented and known, but people in the field often pretend it doesn't exist.

There may be some tendency for some therapies to work better for some people more than others, but identifying how to match people to therapies hasn't been demonstrated well or replicably.

Interesting. Is this based on your personal observations of the published literature or are there systematic reviews that concur?

I think it would be helpful in a different way. To put it simple, it would give people in suffering, more time to talk.

I don’t think it can replace the human relationship, but it could act as helper, a listener, or something between the patient and the therapist. It could listen for hours very cheap, provide feedback, replay things previously said by the patient helping in auto reflection, create a report for the therapist, etc.

Yes, we do need the therapeutic relationship, but for most that is too expensive. Take for example someone that can only afford 2 therapies a month. That’s simply not enough in times of crisis, when the thoughts in your head can be compared with a bull stampede.

That's a really good point and I agree that those tools could really help people. Anonymous expression of suffering can be therapeutic; people say things to Siri that they've never told anyone

I just have an adverse intellectual and emotional reaction when I hear people suggest AI / some app is the solution to mental illness. It devalues the hard and thankless work that therapists do, reduces people's deep suffering to a trivial engineering problem, naively presumes superiority to the centuries of research humankind has done on the mind and suffering, and is just logically and factually incorrect

There are plenty of ways to improve access to care that would have more cost-effective benefit than AI. If people started investing in those then maybe we could actually see some improvement

/endrant

Yeah, a trivial engineering problem, what could possibly go wrong?.

I wonder which are those other cost-effective ways to improve access to care you refer to...

Was probably aggressive to characterize it as "trivial"

Some examples of cost effective interventions include better mental health and substance abuse treatment in primary care (methadone tx for example), integration of medical and behavioral health reimbursement and relaxing some onerous rules like not being able to bull for medical and behavioral services on the same day (though I think this is improving), better training and supervision for mental health professionals, more rigorous research into group therapy, Better education on psychiatric meds for PCPs, more follow up etc

This seems like a case where you have downgraded the meaning of "AI" to match its current commercial connotations. A relationship with an AGI >= a relationship with a human, at least in intellectual terms.

The big hope, that AFAIK we have not abandoned, is that AGIs will be able to walk into a messy human situation, diagnose exactly what the dynamics are, and recommend the optimal thing to do to fix things. This could happen via an ELIZA-like interface or through a broader perceptual engagement. The point, though, as the GP suggested, is that AGI could quickly learn to far exceed the perceptiveness and skill of a human therapist, and do so on an inconceivable scale.

It's the idea of developing an "ecology of mind", as Gregory Bateson put it. To be able to diagnose quickly and accurately disorders (individual and collective) that currently require lucky individual insights and experience to identify.

My broader point was that much of the benefits of psychotherapy can be attributed to social or emotional factors, not intellectual factors. While there is considerable debate about this, a lot of people are of the mind that therapy is not about precisely diagnosing some issue and then saying something to a patient to optimally fix things. If AI becomes emotionally and socially intelligent, maybe they can do good therapy. But if you think that mental illness can be solved through deterministic problem solving, logic and recommending the optimal thing to do to fix things, i would view that perspective as not consistent with our current understanding of neuroscience and mental illness

I'd question whether you are trying to reduce the problem of mental illness to fit the capabilities of the tool you are working with, rather than trying to earnestly understand the problem

I'm a mental health service user who has experienced psychosis and multi-week hospitalisation. I know what I'm talking about when I express hope for a more objective and -- ideally -- by today's standards seemingly omniscient AGI for treatment/therapy. Human therapists are unfortunately variable in quality, and they often participate in a system of casual oppression of the mentally ill, even if you discount whatever they charge for their services.

So, I don't lack earnestness. On the contrary, I strongly believe in a reductionist model of the brain. Who would deny that a sufficiently advanced AGI should be able to diagnose and treat mental illness, enlisting whatever social and emotional (or biopsychosocial) interventions are necessary or economical.

You seem to be waving your hands and saying that physical reality is not susceptible to deterministic modelling and interventions; or maybe you just mean that the brain is too complex for AGI to grasp. Either way, I think you overestimate the diversity and functional complexity of mental disorders, and underestimate the capability of a focused diagnostic intelligence. The promise of AGI is that it will get to the bottom of the dynamics of the patient's disorder. Once you have that picture, you can deploy appropriate social/emotional resources (and much more economically, which means it could actually happen in dysfunctional health systems.)

Even if it's just "good enough" and in the right direction. That would be a huge improvement on the existing situation.

As for AI being "the tool I am working with", I'm not involved in AI/ML in any way. AGI to me means human-equivalent intelligence, however implemented. The hope that it could be massively augmented relative to a human expert is uncontroversial (for example, it is anticipated in the way that self-driving car systems will learn from parallel acquisition of data by many vehicles.)

So, as a patient, it seems to me that better insight into the dynamics of the illness (via AI) is a huge potential benefit. Human generosity and empathy is often helpless in the face of severe mental illness -- relatives suffer for years without any hope of improvement. So I'm not sympathetic to the idea that what is needed is just more uninformed but kind and generous front-line human cannon fodder. Mental health is a very difficult profession.

Thank you for the detailed response, and I apologize for the assumptions I made about your background. I just get a bit upset and irrational sometimes about this subject, because it is one I care a lot about

I understand that human therapists can do just as much harm as good and that a lot of patients are treated terribly by the system. I think the best way to have near term positive impact on patients with mental illness is to expand access to current evidence based care and high quality practitioners. I am not talking about generous and kind front line human cannon fodder, but service models that expand access to high quality evidence based care and support and empower the mental health professionals that deal with severe, complex patients so these patients don't have to deal with terrible destructive therapists

For serious mental illness and psychosis, I understand that expanding access to services is not always enough, although we do have existing pharmaceutical interventions that can help certain patients, and that improving access to some of these medications could be helpful.

I appreciate your comment that I am "hand waving", but I believe that proponents of an AI solving mental health are likewise hand-waving and assuming that mental illness can 1) be reduced to a problem that is tractable with current or near term AI tools or 2) that the singularity and godlike AI is inevitable. As far as I know, both of those are beliefs, not foregone conclusions. I am open to being convinced, but i have not seen any solid evidence or specific hypotheses for how AI would help mental illness beyond chatbots, and while chatbots can certainly help some people they will not be a cure all

I would love to see evidence that I am overestimating the diversity and complexity of mental illness. From what I understand we don't know enough about these conditions to know how complex and diverse they are, but our current tools have failed to bound or characterize the complexity or diversity of mental illness in a way that make them tractable problems. Decades of study of the genetics and mechanisms of depression have failed to yield any breakthroughs, and we couldn't even predict which genes are associated with the disease (no serotonin / dopamine genes are that strongly correlated with disease). Fmri and eeg have horrible resolution and do very little to elucidate biochemical pathways. No new drugs with novel mechanisms have been developed for psychiatric disorders in decades. I'm not an AI expert by any means, but without tools to collect useful data, I can't see how AI could deliver much that is useful in terms of characterizing the biology of disease and intervening

Rather than invest in AI, I would like to see more investment in higher resolution, dynamic tools to analyze molecular pathways in the brain in Vivo. That would probably generate enough low hanging fruit that you could develop better treatments without AI tools

I went to a talk by Tom Insel, the former head of NIMH who was at Verily and now at a MH startup, and he told a story about how he gave a talk on the research of genetics in mental illness to a group of patients and their families. A parent of a child with sever mental illness stood up and said "you don't get it. The house is on fire and you're talking about the chemistry of the paint". Maybe I'm missing something, but I sort of feel like this when people talk about AI or apps for mental illness

Anyway I appreciate you sharing your thoughts so objectively, fairly and openly and hope that mental health care will improve through AI or other means

Are you familiar with: https://woebot.io/

It acts as a cognitive behavioral therapy tool.

I was not. Thank you for that link.

Thank you for sharing this!

Reply on news.ycombinator.com