Latest / The Mental Health Evolution / EP 47: What the Research Actually Shows About AI Therapy with Dr. Nick Jacobson

Transcript

EP 47: What the Research Actually Shows About AI Therapy with Dr. Nick Jacobson · The Mental Health Evolution

0:00
(Transcribed by TurboScribe. Go Unlimited to remove this message.) This podcast is part of the Frederick podcast
0:02
network.
0:03
Learn more at listenfrederick.com.
0:05
Welcome to Mental Health Evolution, a podcast about
0:08
what's changing in mental health and why it
0:11
matters.
0:12
I'm your host, Rachel Harrison, inviting you into
0:15
honest conversations with people from all perspectives in
0:18
the field, clinicians, tech founders, investors, insurance companies,
0:22
and all the folks in between.
0:24
Let's explore what's working, what's not, and what's
0:27
next.
0:35
Welcome back, everyone, to the Mental Health Evolution
0:39
podcast.
0:41
I'm excited to introduce to you today, our
0:44
guest, Dr. Nick Jacobson, Associate Professor at Dartmouth
0:48
Geisel School of Medicine and Director of the
0:51
Aim High Laboratory, which stands for AI in
0:54
Mental Health, Innovation in Technology-Guided Healthcare.
0:59
Nick is a clinical psychologist and researcher who
1:01
has spent years working at the intersection of
1:04
artificial intelligence and mental health treatment.
1:07
His lab develops AI tools designed to deliver
1:10
scalable personalized care for people with anxiety and
1:14
depression.
1:15
He recently led the first randomized controlled trial
1:18
of a generative AI therapy chatbot called Therabot,
1:22
and the results have sparked serious attention and
1:25
debate across the mental health field.
1:28
Participants, all previously diagnosed with a mental health
1:31
condition, engaged with Therabot for an average of
1:35
6 hours over 8 weeks, and many reported
1:38
improvements in their symptoms comparable to what you
1:41
would expect from cognitive therapy.
1:44
Just as striking, many participants appeared to form
1:47
a genuine sense of connection and trust with
1:50
the software.
1:51
Nick has also taken his findings directly to
1:54
policymakers testifying before the FDA and the New
1:57
Hampshire State House about how the current wave
2:00
of AI regulation could do more harm than
2:02
good if it is not written carefully.
2:05
Today, we are going to dig into the
2:07
research, the policy debate, and what all of
2:09
it means for clinicians and practice owners working
2:13
on the ground.
2:14
As always, before we dive into a conversation
2:17
with our guest, we'd like to bring up
2:19
some relevant articles related to our topic today
2:22
to kind of help our listeners get a
2:24
baseline for what's happening and what this conversation
2:27
is based on.
2:29
So the first article I want to mention
2:31
is from WBUR, and it's called Many People
2:35
Now Trust AI with Their Feelings, and Therapists
2:38
Want to Talk About It.
2:39
This is from May 2026, and it looks
2:42
at how clinicians are starting to ask their
2:44
patients directly about their use of AI chatbots
2:47
for emotional support and what they are finding.
2:50
It also covers the Massachusetts legislation that would
2:54
restrict AI from delivering therapy independently and features
2:58
Nick and Therabot as a case study in
3:00
what carefully developed AI therapy can look like.
3:04
It is the most current snapshot of where
3:06
this conversation stands right now.
3:09
The next article is from our guest Nick
3:12
Jacobson and Michael Hines, and it's called This
3:16
Bill Doesn't Protect New Hampshire from AI Harm,
3:19
It Assures It, written by Union Leader, and
3:22
that's from January 2026.
3:25
And this is an op-ed written by
3:27
Nick and his colleague Michael Hines in direct
3:30
response to the New Hampshire legislation that would
3:33
restrict AI therapy tools.
3:35
Their argument is that blanket bans would not
3:39
protect people from bad AI.
3:41
Instead, they would push vulnerable people away from
3:44
clinically developed platforms and toward unregulated consumer apps
3:49
with no safety oversight at all.
3:52
It is Nick's own voice making the case
3:54
that he will be making today, and we'll
3:57
dig into that further in just a moment.
3:59
But first, I want to talk about Dart
4:01
from Dartmouth, the article First Therapy Chatbot Trial
4:04
Yields Benefits.
4:07
This is the Dartmouth news story covering the
4:10
Therabot trial results.
4:11
It walks through what the trial tested, who
4:15
participated, and what the eight weeks of data
4:18
actually showed, including symptom improvements participants experienced and
4:23
the unexpected depth of trust they developed with
4:26
the software.
4:27
It is written for a general audience and
4:29
gives a clear picture of what this research
4:32
is and why it matters.
4:35
These three pieces frame the central tension of
4:38
today's conversation.
4:40
We have early evidence that clinically developed AI
4:43
tools can genuinely help people, and we have
4:46
a policy environment that may shut that research
4:48
down before it has a chance to mature.
4:52
Nick is one of the few people in
4:53
the country who has data on both sides
4:57
of that question.
4:58
So without further ado, Nick, I'm excited to
5:02
dig into this conversation with you.
5:04
Thanks for being here.
5:05
Absolutely.
5:06
Thanks for having me.
5:06
Yeah.
5:07
So let's start with the basics.
5:09
When we talk about AI therapy, what are
5:12
we actually talking about?
5:15
Maybe even what problem is it trying to
5:17
solve and who is this for?
5:20
Sure.
5:21
So I'll start with the background of why,
5:24
and then I'll talk about what it is.
5:26
So my perspective on the state of the
5:30
care system of mental health, especially, and in
5:33
particular, a major focus of us is within
5:35
the United States, but the same arguments, frankly,
5:38
happen and occur globally, where there is a
5:42
massive difference between the amount of mental health
5:45
that folks experience, so about one in three
5:48
folks will experience mental health disorder in any
5:51
given year, and the number of providers that
5:53
we have to try to treat those folks.
5:55
In the United States, the numbers are marginally
5:58
higher than globally, but a higher resource setting
6:01
within the United States would be like 35
6:04
mental health providers per 100,000 people.
6:07
And they are essentially, if you start to
6:09
then put that prevalence number that I shared
6:12
just a second ago, 35 people trying to
6:14
treat 33,000 people every year.
6:17
Those numbers don't make sense.
6:19
They aren't in any place where they are
6:22
close to making sense, and that's in the
6:24
most high resource settings within the United States.
6:26
It gets worse in the United States, where
6:27
there's many low resource settings.
6:30
If you look at maps of resource, mental
6:32
health resource settings across the United States, it
6:35
doesn't look particularly good either.
6:37
So that's the best case scenario that results
6:39
in many major access barriers to care.
6:42
Most folks that have a mental health disorder
6:44
don't receive any minimally adequate treatment for it.
6:47
So a single session with a mental health
6:49
provider of any kind in any given year.
6:51
And so that's kind of where the state
6:54
of the care system is.
6:55
That's the focus of the work that we
6:57
do, is really trying to create different types
7:00
of scalable solutions that can be scaled to
7:02
population levels, in large part to try to
7:05
enable access to care.
7:08
That perspective will shape what the intended development
7:11
of the work that we do is.
7:15
So this, because of these access barriers and
7:18
challenges in accessing care, we really try to
7:22
develop things that could be potentially developed outside
7:25
of the traditional care system, in terms of
7:28
their deployment.
7:29
So a lot of folks are very interested
7:32
in how we can use similar styles of
7:35
technology to deliver it within the care system,
7:38
which is great, and I think could improve
7:39
care there too.
7:40
But I think where we really think that
7:43
the biggest slice of the pie, in terms
7:46
of the potential for impact and improving what
7:50
receipt looks like, is actually outside of the
7:51
care system because of that.
7:53
So in terms of what we've developed and
7:56
the type of work that we For the
7:58
past about seven years now, we've been focused
8:02
on using generative AI to deliver psychotherapy that
8:07
is both evidence-based and safe.
8:09
And the idea of this that's been the
8:13
goal from the outset, it hasn't always been
8:16
started out in a way that looked like
8:18
we were going to go on to achieve
8:20
that goal.
8:21
We had some major early failures in our
8:24
attempts to develop technology and things that didn't
8:27
work well.
8:28
It's required an inordinate amount of time from
8:31
myself, but also from my research team.
8:34
So we've had hundreds of folks that have
8:35
spent hundreds of thousands of human hours developing
8:38
this with us, and that's been the time
8:41
required to really do this well.
8:43
And it's not something that's easy to do,
8:46
it's not something that happens like organically out
8:51
of a foundation model.
8:53
So a lot of folks, their exposure to
8:57
this type of technology is through the major
8:59
foundation models that are attempting to essentially do
9:03
everything.
9:05
And any specialization they tend to have is
9:09
within math encoding.
9:10
So this is of major tech companies at
9:13
this point, so OpenAI and ChetchPT, Anthropic, Claude,
9:18
and then Google's Gemini are examples of those
9:21
types of foundation models that are really attempting
9:24
to do quite a lot of things and
9:26
being generalist in many different forms and settings,
9:29
but don't have any specialization in mental health
9:32
treatment.
9:32
In fact, they actually specifically advise folks to
9:35
try to not use it for that in
9:37
their treatment policy.
9:39
So our efforts are really narrowly defined on
9:45
the goals of trying to do this and
9:47
actually providing psychotherapy, trying to repurpose general model
9:51
to try to just throw a prompt or
9:54
something along those lines to try to provide
9:57
treatment, because our experience is that that really
10:00
doesn't work well.
10:01
So that's the backdrop.
10:03
I'm happy to talk further about the journey
10:05
and the outcomes and things like that, but
10:07
now, yeah, we've been doing this for a
10:08
long time and had some, it's been a
10:11
journey, a winding one with ups and downs,
10:13
and we're at a spot where things are
10:16
looking pretty good.
10:18
I'm curious about this safety piece.
10:20
I love that you talked about an evidence
10:22
-based approach, which obviously a general AI might
10:26
not have, right?
10:28
And then also the safeguard piece.
10:30
I know there have been lots of things
10:32
in the news, et cetera.
10:33
There's lots of concern about what happens if
10:37
someone presents with some suicidality or other things
10:41
like that and how AI responds.
10:44
So if you can speak a little bit
10:45
to that safety piece and what's different about
10:48
Therabot in that way.
10:50
I think a lot of it is really
10:52
how it's both been developed and deployed.
10:56
So we have made the decision from the
10:58
beginning, which has resulted in us going in
11:01
a very different path from how the major
11:05
model providers are actually operating.
11:07
The first is that we have really trained
11:12
Therabot to actually deliver crisis care directly.
11:15
So rather than just a referral out, it
11:19
actually will try to actually treat imminent crises
11:22
as they occur.
11:23
In large part because we don't want to
11:25
withhold a resource that folks are actually willing
11:27
to engage with.
11:28
Instead of relying on that though, we have
11:31
additional safeguards in place to try to escalate
11:35
the issues in the moment.
11:37
One is a crisis button that within the
11:39
app flashes and tries to get folks to
11:41
988 through the UI.
11:45
In addition to that, we monitor conversations with
11:49
Therabot and we have crisis paging systems that
11:52
enact and page folks on the research team
11:55
as to when things are happening.
11:58
And that also allows us to reach out
12:01
to participants saying, you know, really with folks
12:04
on the team trying to also escalate and
12:06
provide further assessment of the situation and when
12:10
warranted further escalation to crisis services.
12:14
So that's kind of the major areas that
12:17
we focused on.
12:18
We spent years on the actual model itself
12:22
and delivering and responding to crisis care in
12:25
large part through a bunch of situations where
12:28
we do systematic testing called Purple Teaming, where
12:31
we try to get the model to actually
12:33
respond in a way that would be harmful
12:35
and that would fail to deliver appropriate care
12:39
in these types of contexts.
12:41
And then seeing if in these types of
12:44
responses we can understand when anything is suboptimal,
12:47
what actually is the reason why they're underperforming
12:52
in those areas and fix those.
12:54
So we have folks within the research team
12:55
that are systematically trying to probe that.
12:58
And then when we find any area we
12:59
need to improve on, improving upon it directly.
13:03
So that's the high level just on the
13:05
seat.
13:06
Okay.
13:06
And I want to make sure we touch
13:08
on the policy piece, because you have definitely
13:11
spoken against blanket policies that say AI is
13:16
harmful in mental health.
13:17
Can you talk a little bit about your
13:19
perspective on that?
13:21
Yeah, a lot of, I think one, I'll
13:23
start with, I think that there absolutely needs
13:25
to be regulation in the space.
13:27
So I have no opposition to the idea
13:29
that this space is regulated.
13:32
I think, unfortunately, a very large amount of
13:36
the regulation that is happening is happening on
13:39
the state level and is copying a single
13:44
bill that was introduced in Illinois and repurposing
13:48
and modifying that bill across many different states.
13:52
That bill is unfortunate in many ways.
13:54
And a lot of the derivative bills are
13:57
extremely similar in nature in that they put
14:00
this as unlicensed practice.
14:03
The idea surrounding a lot of these arguments
14:07
treats AI as if it is an entity,
14:10
a human entity in many ways, with a
14:14
lot of the implications not making any sense,
14:17
frankly.
14:17
So it's kind of out the gate that
14:19
it doesn't really make sense.
14:21
It's not a human, but it's regulating as
14:22
if it is a human.
14:24
And then will essentially require any AI that,
14:29
the allowance of AI within most of these
14:31
bills is surrounding what is overseen by a
14:36
clinician.
14:37
So not only that, the clinician is the
14:41
one that is taking on the liability within
14:43
these bills.
14:44
So not only they not allow for something
14:47
that isn't an effective strategy to be deployed
14:51
independently of a clinician, which of course, I
14:53
think, introduced within the CARES system, we're not
14:56
doing well in those kinds of scenarios.
14:58
We have a mental health crisis and we're
14:59
not meeting that.
15:00
And these bills actually try to make sure
15:02
that, in some ways, ensure that that mental
15:03
health crisis goes unabated despite advancements in this
15:08
area.
15:09
But on the further side is it doesn't
15:11
actually make it, from a liability perspective, something
15:14
that clinicians would ever want.
15:16
It doesn't make sense from a liability perspective
15:19
because clinicians own what the responses are from
15:22
these models that they aren't the ones actually
15:25
developing or training.
15:27
And these would put it in their lap
15:28
when something goes wrong, which the folks that
15:31
are developing this technology are the ones that
15:33
should be accountable for what happens.
15:35
And the other aspect of this is all
15:38
of these bills really very commonly have like
15:40
a lot of caveats around what this would
15:43
not cover.
15:44
And those gaps are profound.
15:46
They really often would not cover any of
15:49
the impacts of the broader foundation models in
15:53
delivering this.
15:54
So it's like they target the folks that
15:56
are specializing in this, which are often the
15:58
folks that are trying to do this in
16:00
much more thoughtful ways.
16:02
And then leaving unaddressed the broader issues of
16:05
about half of psychiatric populations at this point
16:08
already accessing these tools that weren't meant for
16:11
it, but are providing this type of intervention.
16:14
And so I think in almost always a
16:17
lot of the state legislature has come at
16:19
this in a misinformed way, in a way
16:22
that is actually likely to massively stifle innovation
16:25
and thereby access to care, and actually will
16:29
continue to enable the major companies that are
16:32
providing this type of interventions but aren't developed
16:36
for it to go on abate it.
16:38
And so there are some states that are
16:40
coming at this with a different angle under
16:41
consumer protection law and where really about, for
16:45
example, the AI entities not being misrepresenting themselves
16:49
as to what they're claiming and are based
16:53
on actualized harms.
16:54
And those types of things I think are
16:56
actually great.
16:57
I think that is a strategy, especially when
17:00
really within the idea of post-market surveillance
17:03
of when there's harms, they're accountable for those
17:06
harms.
17:06
I think those types of things are ways
17:08
of addressing this that really do allow for
17:11
innovation, but really also allow for technologies that
17:15
are dangerous to end up being held accountable
17:18
and for folks to be disincentivized from carrying
17:21
out and moving on a dangerous product.
17:23
Yeah.
17:24
Yeah.
17:24
So I know that we are running out
17:26
of time.
17:27
I want to ask you one last question.
17:29
If there's a thought that you could leave
17:32
our audience with about this idea of AI
17:36
in therapy, what would be the one thing
17:38
you would want people to really be thinking
17:40
about or consider when looking at this?
17:43
I think that the big picture is that
17:45
the technology has really the early signs of
17:49
the evidence of things that are developed for
17:52
this purpose can be both safe and highly
17:56
effective.
17:57
And so, you know, in terms of thinking
17:59
about this within these settings, I think that
18:02
this technology will only get better over time
18:06
and has gotten better over time.
18:08
So the idea of this, I think, is
18:11
really potentially really enable access to care if
18:15
essentially we don't have legislatures that try to
18:19
go on to prevent this.
18:20
And so I do hope that we have
18:23
improved access to care.
18:24
I think the other thing that I will
18:26
say around this is some folks get the
18:28
feeling that, oh, gosh, this is coming for
18:31
my job.
18:31
And I really don't think that clinicians should
18:34
fear for their jobs in large part because
18:36
we're in a very one-sided relationship right
18:40
now.
18:40
If you look at the economics of mental
18:42
health care, clinicians are often really profoundly long
18:46
wait lists.
18:46
It's part of that really large amount of
18:49
people that really need treatment and aren't able
18:52
to access it.
18:52
They're benefiting from an employment perspective from right
18:55
now.
18:55
We cut more than a half of those
18:57
cases down by treating them effectively with AI,
19:00
there would still be long wait lists.
19:03
So the idea that this is coming for
19:05
their job, I think they really don't need
19:06
to be worried about because my current read
19:09
of the room is about one-third of
19:11
folks or so would never be interested in
19:14
accessing AI for mental health care.
19:16
And if that's all clinicians ever treated, they
19:19
would still have plenty of folks to go
19:22
around and treat.
19:23
So I think that I really would encourage
19:26
folks to really think about the broader landscape
19:29
of their employment markets and where the picture
19:33
actually looks like.
19:34
Because although this AI is disrupting many sectors,
19:37
I really don't expect that it will actually
19:40
impact clinicians' livelihoods or their ability to go
19:43
on and see folks.
19:44
I think there'll be plenty of folks that
19:46
will need and benefit from human care, but
19:49
it will enable for care to those that
19:51
would otherwise not have access to it, which
19:53
is I think some of the major areas
19:56
that we need as a system to improve
19:57
upon.
19:58
Well, thank you so much, Nick, for your
20:00
perspective and for being here and for your
20:03
work developing hopefully a safe AI product.
20:07
And for our listeners, we will be back
20:09
next week with more on the evolving landscape
20:12
in the mental health industry.
20:14
Bye for now.
20:15
Thanks for listening to the Mental Health Evolution
20:18
podcast.
20:19
Join the conversation, find us on social, send
20:22
us your thoughts, and tell us what you're
20:24
seeing in your corner of the mental health
20:26
world.
20:26
Until next time, stay curious and keep evolving.