What the Research Actually Shows About AI Therapy with Dr. Nick Jacobson

welcome to Mental Health Evolution, a podcast about what's changing in mental health and why it matters. I'm your host, Rachel Harrison, inviting you into honest conversations with people from all perspectives in the field. Clinicians, tech founders, investors, insurance companies, and all the folks in between. Let's explore what's working, what's not, and what's next. Welcome back everyone, to the Mental Health Evolution Podcast. I'm excited to introduce to you Today our guest, Dr. Nick Jacobson, Associate professor at Dartmouth Geisel School of Medicine and Director of the AIM High Laboratory, which stands for AI and Mental Health and Innovation in Technology Guided Healthcare. Nick is a clinical psychologist and researcher who has spent years working at the intersection of artificial intelligence and mental health treatment. His lab develops AI tools designed to deliver scalable, personalized care for people with anxiety and depression. He recently led the first randomized controlled trial of a generative AI therapy chatbot called Therabot, and the results have sparked serious attention and debate across the mental health field. Participants all previously diagnosed with a mental health condition, engaged with Therabot for an average of six hours over eight weeks, and many reported improvements in their symptoms comparable to what you would expect from cognitive therapy. Just as striking, many participants appeared to form a genuine sense of connection and trust with the software. Nick has also taken his findings directly to policymakers, testifying before the FDA and the New Hampshire State House about how the current wave of AI regulation could do more harm than good if it is not written carefully. Today we are going to dig into the research, the policy debate, and what all of it means for clinicians and practice owners working on the ground. As always, before we dive into a conversation with our guest, we'd like to bring up some relevant articles related to our topic today to kind of help our listeners get a baseline for what's happening and what this conversation is based on. So the first article I want to mention is from WBUR and it's called Many People Now Trust AI With Their Feelings and Therapists Want to Talk About it. This is from May 2026 and it looks at how clinicians are starting to ask their patients directly about their use of AI chatbots for emotional support and what they are finding. It also covers the Massachusetts legislation that would restrict AI from delivering therapy independently and features Nick and Therabot as a case study in what carefully developed AI therapy can look like. It is the most current snapshot of where this conversation stands right now. The next article is from our guest Nick Jacobson and Michael Hines. And it's called this Bill Doesn't Protect New Hampshire from AI Harm, It Assures It. Written by Union Leader. And that's from January 2026. And this is an op ed written by Nick and his colleague Michael Hines in direct response to the New Hampshire legislation that would restrict AI therapy tools. Their argument is that blanket bans would not protect people from bad AI. Instead, they would push vulnerable people away from clinically developed platforms and toward unregulated consumer apps with no safety oversight at all. It is Nick's own voice making the case that he will be making today. And we'll dig into that further in just a moment. But first, I want to talk about Dart from Dartmouth. The article First Therapy Chatbot Trial Yields Mental Health Benefits. This is the Dartmouth news story covering the Therabot trial results. It walks through what the trial tested, who participated, and what the eight weeks of data actually showed, including symptom improvements participants experienced and the unexpected depth of trust they developed with the software. It is written for a general audience and gives a clear picture of what this research is and why it matters. These three pieces frame the central tension of today's conversation. We have early evidence that clinically developed AI tools can genuinely help people, and we have a policy environment that may shut that research down before it has a chance to mature. Nick is one of the few people in the country whose data on who has data on both sides of that question. So without further ado, Nick, I'm excited to dig into this conversation with you. Thanks for being here.

Absolutely. Thanks for having me.

Yeah. So let's start with the basics. When we talk about AI therapy, what are we actually talking about? Maybe even what problem is it trying to solve and who is this for sure?

So I'll start with the background of why and then I'll. I'll talk about what it is. So my perspective on the state of the care system of mental health especially, and in particular a major focus of us is within the United States. But the same arguments, frankly happen and occur globally where there is a massive difference between the amount of mental health that folks experience. So about one in three folks will experience mental health disorder in any given year and the number of providers that we have to try to treat those folks in the United States, the numbers are marginally higher than globally, but a higher resource setting within the United States would be like 35 mental health providers per a hundred thousand people. And they are essentially, if you start to then put that Provence number that I shared just a second ago, 35 people trying to treat 33,000 people every year. Those numbers don't make sense. They aren't in any place where they, they are close to making sense. And that's in the most high resource settings within the United States. It gets worse in the United States where there's many low resource settings. If you look at maps of resource mental health resource settings across the United States, it doesn't look particularly good either. So that's the best case scenario that results in many major access barriers to care. Most folks that have a mental health disorder don't receive any minimally adequate treatment for it. So a single session with a mental health provider of any kind in any given year. And so that's, yeah, kind of where the state of the care system is. That's the focus of the work that we do, is really trying to create different types of scalable solutions that can be scaled to population levels in large part to try to enable access to care. That perspective will shape what the intended development of the work that we do is. So this, because of these access barriers and challenges in accessing care, we really try to develop things that could be potentially developed outside of the traditional care system in terms of their deployment. And so a lot of folks are very interested in how we can use similar styles of technology to deliver it within the care system, which is great and I think could improve care there too. But I think where we really think that the, the biggest slice of the pie in terms of the potential for impact and improving what, what receipt looks like is actually outside of the care system because of that. So in terms of what we've developed and the type of work that we do, for the past about seven years now, we've been focused on using generative AI to deliver psychotherapy that is both evidence based and safe. And the idea of this, that's been the goal from the outset. It hasn't always been started out in a way that looked like we were going to go on to achieve that goal. We had some major early failures in our attempts to develop technology and things that didn't work well. It's required an inordinate amount of time from myself, but also from my research team. So we've had hundreds of folks that have spent hundreds of thousands of human hours developing this with us. And that's been the time required to really do this well. And it's not something that's easy to do. It's not something that is kind of happens like organically out of a foundation model. So a lot of folks are their exposure to this type of technology is through the major foundation models that are attempting to essentially do everything and that any specialization they, they tend to have is within math and coding. So this is of your major tech companies at this point. So OpenAI and ChatGPT anthropic Claude and then Google's Gemini are examples of those types of foundation models that are really attempting to do quite a lot of things and being generalist in many different forms and settings, but don't have any specialization in mental health rerun. In fact they actually specifically advise folks to try to not use it for that in their policy. Um, so this, our efforts are really kind of narrowly defined on the goals of trying to do this and actually providing psychotherapy, trying to repurpose general model to try to just through a prompt or something along those lines to. To try to provide treatment. Because our experience is that that really doesn't work well. So that's the backdrop. I'm happy to talk further about the journey and the outcomes and things like that, but not. Yeah, we've been doing this for a long time and had some. It's been a journey, a winding one both with ups and downs and we are at a spot though where things are looking pretty good.

I'm curious about this safety piece. I love that you talked about an evidence based approach which obviously a general AI might not have. Right? And then also the safeguard piece. I know there have been lots of things in the news, etc. There's lots of concern about, you know, what happens if someone presents with some suicidality or other things like that and how AI responds. So if you can speak a little bit to that safety piece and what's different about the robot in that way,

I think a lot of it is, is really how it's. It's both been developed and deployed. So we have made the decision from the beginning which has resulted in us going in a very different path from how the major model providers are actually operating. The first is that we have really trained Therabot to actually deliver crisis care directly. So rather than just a referral out, it actually will try to actually trigger treat imminent crises as they occur. In large part because we don't want to withhold a resource that folks are actually willing to engage with. Instead of relying on that though, we have additional safeguards in place to try to escalate the issues in the moment. One is a crisis button that within the app flashes and tries to get folks to 988 through the UI in addition to that, we monitor conversations with Arabot and we have crisis paging systems that enact and page folks on the research team as to when things are happening. And that also allows us to reach out to participants saying, you know, really with folks on the team trying to also escalate and provide further assessment of the situation and when warranted, further escalation to. To crisis services. So that's kind of the major areas that we focused on. We've spent years on the actual model itself and delivering and responding to crisis care in large part through a bunch of situations where we do systematic testing called purple teaming, where we try to get the model to actually respond in a way that would be harmful and that would fail to deliver appropriate care and these types of contexts. And then seeing if in these types of responses we can understand when anything is suboptimal, what actually is the reason why they're perform underperforming in those areas and fix those. So we have folks within the research team that are systematically trying to probe that. And then when we find any area we want to, we need to improve on improving upon it directly. So that's the high level, just on the safe.

Okay. And I want to make sure we touch on the policy piece because you have definitely spoken against blanket policies that say AI is harmful in mental health. Can you talk a little bit about your perspective on that?

Yeah, a lot of. I think. One I'll start with. I think that there absolutely needs to be regulation in the space. So I have no, no opposition to the idea that this space is regulated. I think, unfortunately, a very large amount of the regulation that is happening is happening on the state level and is copying a single bill that was introduced in Illinois and repurposing and modifying that bill across many different states. That bill is unfortunate in many ways. And a lot of the derivative bills are extremely similar in nature and that they put this as unlicensed practice. The idea surrounding a lot of these arguments treats AI as if it is an entity, a human entity in many ways, with a lot of the implications not making any sense, frankly. So, like it's kind of out the gate that it doesn't really make sense. It's not a human, but it's regulating as if it is a human, and then will essentially require any AI that the allowance of AI within most of these bills is surrounding what is overseen by a clinician. So not only that, the clinician is the one that is taking on the liability within these bills. So not only they not allow for something that isn't an effective strategy to be deployed independently of a clinician, which of course I think introduced within the care system, we're not doing well in those kinds of scenarios. We have a mental health crisis and we're not meeting that. And these bills actually try to make sure that in some ways ensure that that mental health crisis goes unabated despite advancements in this, in this area. But on the further side is like it doesn't actually make it in a let from a liability perspective something that clinicians would ever want. It doesn't make sense from a liability perspective because clinicians own what the responses are from these models that they aren't the ones actually developing or training. And these would put it in their lap when something goes wrong, which the folks that are developing this technology are the ones that should be accountable for what happens. And the other aspect of this is all of these bills really very commonly have like a lot of caveats around what this would not cover. And those gaps are profound. They really often would not cover any of the impacts of the broader foundation models in delivering this. So it's like they target the folks that are specializing in this, which are often the folks that are trying to do this in much more thoughtful ways and then leaving unaddressed the broader issues of about half of psychiatric populations at this point already accessing these tools that weren't meant for it, but are providing this type of intervention. And so I think in almost always a lot of the state legislature has come at this in a misinformed way, in a way that is actually likely to massively stifle innovation and thereby access to care and actually will continue to enable the major companies that are providing this type of interventions but aren't developed for it to go unabated. And so there are some states that are coming at this with a different angle under consumer protection law. And we're really about, for example, the AI entities not being misrepresenting themselves as to what they're claiming and are based on actualized harms. And those types of things I think are actually great. I think that is a strategy especially when really within the idea of post market surveillance of when there's harms, they're accountable for those harms. I think those types of things are ways of addressing this that really do allow for innovation but really also allow for technologies that are dangerous to end up being held accountable and for folks to be disincentivized from carrying out and moving on a dangerous product.

Yeah, yeah. So I know that we are running out of time. I Want to ask you one last question. If there's a thought that you could leave our audience with about this idea of AI and therapy, what would be the one thing you would want people to really be thinking about or consider when looking at this?

I think that the, the big picture is that the technology has really the early signs of, of the evidence of things that are developed for this purpose can be both safe and highly effective. And so you know, in terms of thinking about this within these settings, I think that this technology will only get better over time and has gotten better over time. So the idea of this I think is really potentially really enable access to care if essentially we don't have legislatures that try to go on to prevent this. And so I do hope that we have improved access to care. I think the other thing that I will say around this is some folks get the feeling that oh gosh, this is coming from my job. And I really don't think that clinicians should fear for their jobs in large part because we are in a very one sided relationship right now. If you look at the economics of mental health care, clinicians are often really profoundly long wait lists. It's part of that really large amount of people that really need treatment and aren't able to access it. They're benefiting. From an employment perspective from right now we cut more than half of those cases down by treating them effectively with AI there would still be long wait lists. So the idea that this is coming for their job I think they really don't need to be worried about because my clients current read of the room is about one third of folks or so would never be interested in accessing AI for mental healthcare. And if that's all clinicians ever treated, they would still have plenty, plenty of folks to go around and treat. So I think the I, I really would encourage folks to really think about the, the broader landscape of their employment markets and where the, where the picture actually looks like is although this AI is disrupting many sectors, I really don't expect that it will actually impact clinicians livelihoods or their ability to, to go on and see folks. I think there'll be plenty of folks that will need and benefit from human care, but it will enable for care to those that would otherwise not have access to it. Which is I think some of the major areas that we need as a system to improve upon.

Well, thank you so much Nick for your perspective and for being here and for your work developing hopefully a safe a product. And for our listeners, we will be back next week with more on the evolving landscape in the mental health industry. Bye for now.