Latest / THE DAILY BRIEF / How AI is replacing mental health guidance

Transcript

How AI is replacing mental health guidance · THE DAILY BRIEF

• 0:04
  Speaker 1: All right, now to the latest part of daily life
• 0:06
  Speaker 1: where AI is making its mark. We're talking about therapy.
• 0:10
  Speaker 1: More people, including teenagers, say they turn to online chat
• 0:13
  Speaker 1: pots like chat GPT for mental health guidance. Just this week,
• 0:17
  Speaker 1: open ai announced they're adding new mental health related guardrails
• 0:21
  Speaker 1: to chat GPT. It comes in the wake of a
• 0:23
  Speaker 1: lawsuit from a California family who claims the service played
• 0:26
  Speaker 1: a role in their teenage son's suicide. Open ai responded
• 0:30
  Speaker 1: in a statement saying, recent heartbreaking cases of people using
• 0:33
  Speaker 1: chat GPT in the midst of a q crises waste
• 0:36
  Speaker 1: heavily on us. There have been moments when our systems
• 0:39
  Speaker 1: did not behave as intended in sensitive situations. So joining
• 0:42
  Speaker 1: us now is Nick Jacobsin. He's an associate professor of
• 0:45
  Speaker 1: biomedical data science, psychiatry, and computer science at Dartmouth. So
• 0:48
  Speaker 1: let's start with whether or not people should be trusting
• 0:51
  Speaker 1: AI for mental health guidance.
• 0:54
  Speaker 2: Excellent question, Perry.
• 0:55
  Speaker 3: I think one of the things that's really challenging about
• 0:58
  Speaker 3: this space is there's not enough providers to go around,
• 1:01
  Speaker 3: so a lot of folks are really seeking their care
• 1:04
  Speaker 3: in ways that are accessible, and we just our system
• 1:08
  Speaker 3: is actually failing folks, and because of that, they're trying
• 1:11
  Speaker 3: to reach out to things that are these personalized and
• 1:13
  Speaker 3: accessible means. Unfortunately, I think one of the things that's
• 1:17
  Speaker 3: really challenging about this space is it really takes a
• 1:20
  Speaker 3: long time and a lot of effort to actually design
• 1:23
  Speaker 3: something that is both safe and effective. Our team has
• 1:25
  Speaker 3: been working on this with over one hundred people for
• 1:28
  Speaker 3: over one hundred thousand human hours and designing what was
• 1:31
  Speaker 3: the first random and doing the first randomized control trial
• 1:35
  Speaker 3: of an AI used for mental health treatment.
• 1:39
  Speaker 2: And this work is intensive.
• 1:41
  Speaker 3: It has taken us many years to do both safely
• 1:44
  Speaker 3: and effectively, and so I do think that this care
• 1:48
  Speaker 3: can be delivered in AI in a safe and effective way.
• 1:52
  Speaker 3: We have evidence of that in our first randomized control
• 1:54
  Speaker 3: trial that there were really large symptom reductions and depression, anxiety,
• 1:58
  Speaker 3: and eating disorder symptoms in our and our randomized control trial.
• 2:02
  Speaker 3: But one of the things that's really challenging about this
• 2:05
  Speaker 3: is the types of things that we're developing are quite
• 2:08
  Speaker 3: different from what is accessible through these open large model providers,
• 2:14
  Speaker 3: and those types of models act in ways that often
• 2:18
  Speaker 3: feel helpful but are actually harmful. And so that's one
• 2:21
  Speaker 3: of these these really space, these spaces where they can
• 2:25
  Speaker 3: act like they're really giving the user what they want
• 2:28
  Speaker 3: in a lot of ways, but in a lot of
• 2:30
  Speaker 3: ways that the best thing within therapy is often actually
• 2:35
  Speaker 3: undergoing and dealing with the intense discomfort that might be
• 2:38
  Speaker 3: associated with these and these models are often trying to
• 2:41
  Speaker 3: reduce that down and that often results in a lot
• 2:44
  Speaker 3: of adverse.
• 2:45
  Speaker 2: Effects in the moment.
• 2:47
  Speaker 3: But it also these models can act in ways that
• 2:49
  Speaker 3: are seriously harmful, as you see, and in some of
• 2:53
  Speaker 3: the cases that you're mentioning, but also in the cases
• 2:56
  Speaker 3: related to psychotic kind of hallucinations and delusions kind of
• 3:02
  Speaker 3: being amplified and encouraged. And so I think these it
• 3:06
  Speaker 3: is something that I think folks should really appreciate the
• 3:10
  Speaker 3: degree of harm that they could experience and sometimes really unknowingly,
• 3:15
  Speaker 3: And I think that's part of the case of what
• 3:18
  Speaker 3: makes this space particularly difficult. It can feel helpful, but
• 3:22
  Speaker 3: they have some of the same levels of challenges of
• 3:26
  Speaker 3: something like the danger of the open Internet, but they
• 3:29
  Speaker 3: have a lot of polish in them.
• 3:31
  Speaker 1: So open Aye says that they're going to be adding
• 3:34
  Speaker 1: new mental health guardrails themselves. Why should they be trusted
• 3:38
  Speaker 1: to police themselves on something like this.
• 3:42
  Speaker 2: I think that's an excellent question.
• 3:44
  Speaker 3: I think right now there isn't enough legislation surrounding how
• 3:49
  Speaker 3: to actually govern what is safe and appropriate care that
• 3:53
  Speaker 3: is provided.
• 3:55
  Speaker 2: A lot of this on the federal level is enacted.
• 3:57
  Speaker 3: Based on claims that folks are making, and a lot
• 3:59
  Speaker 3: of these large model provider companies are aware that they're
• 4:02
  Speaker 3: treatment or their models are being used for treatment, but
• 4:06
  Speaker 3: they aren't advertising.
• 4:07
  Speaker 2: Them that way.
• 4:08
  Speaker 3: And so things like the FDA, which is the long
• 4:11
  Speaker 3: long standing way that something like this would be regulated
• 4:14
  Speaker 3: in the case they were making claims, is really not
• 4:18
  Speaker 3: actually governing that because they're not making the claims that
• 4:21
  Speaker 3: it's safe and effective or something along those lines for
• 4:23
  Speaker 3: different types of conditions. So I do think that there's
• 4:27
  Speaker 3: need for legislation and oversight of these types of these
• 4:31
  Speaker 3: types of models, and I think there isn't enough incentive
• 4:35
  Speaker 3: alone for model provider major model providers to do it themselves.
• 4:39
  Speaker 3: In a lot of ways, you're seeing that a lot
• 4:42
  Speaker 3: of these model providers are being reactionary as opposed to proactive,
• 4:45
  Speaker 3: and it takes a long time to do this. Well,
• 4:48
  Speaker 3: it's not something that you can do in a month
• 4:51
  Speaker 3: or two and really learn how to how to understand
• 4:54
  Speaker 3: the space and how to actually provide effective character regenerative AI.
• 4:58
  Speaker 1: Nick Jacobson, thank you for that very much important story.
• 5:00
  Speaker 1: Appreciate your time.