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.
> Reasoning that if he's not good it would show up in patients thinking he's bad, and not feeling any better.
This is like a questionnaire on how much stronger you feel after working out at a gym: you often don't, you feel tired.
Both gym and talking therapy (when done correctly) will push you slightly out of your comfort zone, and aim to let you safely deal with moderate amounts of something that you find really hard. So as to expand your capabilities.
"I feel good" immediately after is utterly the wrong metric.
Being more capable / feeling better some time later is the more reliable indicator, like progress at a gym.
And also this is why an agreeable statistical word generator LLM is not the correct tool for the job.
It's not the past anymore, we don't need to debate, we can watch and listen to actual recordings of therapy sessions and the patients going from feeling variously bad to better. Here's Dr David Burns channel with a 4hr video of a session with a woman who is obsessively anxious about her college-age daughter's safety: https://www.youtube.com/watch?v=on2N5DsKHRk
Here's a 2.5 hour session (split into several videos) with a doctor who has a bad relationship with his son and felt like a failure for it:
Here's a couple of hour session with Marilyn who was diagnosed with lung cancer and spiraling with depression, anxiety, shame, loneliness, hopelessness, demoralization, and anger, despite her successful career:
It's like saying "it is still debated if debugging even works" as if all languages, all debuggers, all programmers, all systems, are the same and if you can find lots of people who can't debug then "debugging doesn't work". But no, you only need a few examples of "therapy working" to believe that it works, and see the whole session to see that it isn't just luck or just the relief of talking, but is a skill and a technique and a debugging of the mind.
Dilettante_|2 months ago
(And I'm not being theoretical here, I have quite a bit of experience getting incredibly inadequate mental health care.)
chemotaxis|2 months ago
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.
knollimar|2 months ago
It's better not to degrade the close friend, and "life coach focused on healthy self awareness" is probably indistinguishable from most good therapy.
jodrellblank|2 months ago
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.
SideburnsOfDoom|2 months ago
This is like a questionnaire on how much stronger you feel after working out at a gym: you often don't, you feel tired.
Both gym and talking therapy (when done correctly) will push you slightly out of your comfort zone, and aim to let you safely deal with moderate amounts of something that you find really hard. So as to expand your capabilities.
"I feel good" immediately after is utterly the wrong metric.
Being more capable / feeling better some time later is the more reliable indicator, like progress at a gym.
And also this is why an agreeable statistical word generator LLM is not the correct tool for the job.
airstrike|2 months ago
"Good" is too broad and subjective to be a useful metric.
ffuxlpff|2 months ago
jfindper|2 months ago
jodrellblank|2 months ago
Here's a 2.5 hour session (split into several videos) with a doctor who has a bad relationship with his son and felt like a failure for it:
https://www.youtube.com/watch?v=42JDnrD106w
https://www.youtube.com/watch?v=S5H2YGljhqQ
https://www.youtube.com/watch?v=bZ9_0j_fmeg
https://www.youtube.com/watch?v=eiCrdGVa8Q0
https://www.youtube.com/watch?v=cARvhlTckaM
Here's a couple of hour session with Marilyn who was diagnosed with lung cancer and spiraling with depression, anxiety, shame, loneliness, hopelessness, demoralization, and anger, despite her successful career:
https://www.youtube.com/watch?v=S7sQ_zDGsY8
https://www.youtube.com/watch?v=tyuFN4mbGZQ (there's probably more parts to find through YouTube somehow)
And a session with Lee with loneliness and marriage relationship problems:
https://www.youtube.com/watch?v=imEMM3r6XL8 (probably more parts as well)
It's like saying "it is still debated if debugging even works" as if all languages, all debuggers, all programmers, all systems, are the same and if you can find lots of people who can't debug then "debugging doesn't work". But no, you only need a few examples of "therapy working" to believe that it works, and see the whole session to see that it isn't just luck or just the relief of talking, but is a skill and a technique and a debugging of the mind.
ajuc|2 months ago
Arsenik1|2 months ago