Youth Mental Health and Conversational AI: AI Use for Emotional Support "In the Wild"

0:04 Speaker 1: This program contains mature discussion and language.

0:08 Viewer discretion is advised.

0:28 Cinnamon Bless, Ph.D: In particular,

0:30 I'll be focusing on the phenomenon of AI use for emotional support,

0:37 for companionship out in the wild.

0:40 What I'm calling out in the wild and what I mean

0:42 by that is not under the auspices or guidance of a clinician.

0:47 I look forward to talking about that this morning.

0:52 By almost all accounts, the US is experiencing a youth mental health crisis

1:00 that we've been in for a number of years now.

1:04 In 2023, just two years ago,

1:08 18% of adolescents aged 12-17 who had a major depressive episode,

1:17 but only half received mental health care.

1:21 Many of you are clinicians may be familiar with that, but I

1:26 invite you to take a moment and let that statistic sink in.

1:31 These are some additional facts and figures

1:35 from the National Alliance on Mental Illness,

1:39 suggesting high distress levels and low access to care.

1:44 Forty percent of US high school students had persistent sadness or hopelessness.

1:50 In 2023, this number is up at 65% for LGBTQ+ students.

1:56 Twenty percent of US high school

1:59 students had serious thoughts of suicide tragically.

2:02 Suicide is the second leading cause of death

2:05 among people ages 10-24 years in the United States.

2:09 When we think about young adults, individuals aged 18-25,

2:16 only about 50% of individuals in that age

2:20 group with a mental illness received treatment in 2024, so about a year ago.

2:29 Refocusing on teens and like all people,

2:35 we all experience episodes of feeling distressed, and we all use,

2:40 as do teens, many sources of support out in the wild, outside of clinical care.

2:47 Things like self reflection, turning to family and trusted adults.

2:53 For teens, in particular,

2:55 they may look towards institutional supports through their school,

3:00 relational help seeking from professionals.

3:02 For teens who feel comfortable and can have access, can afford access,

3:09 they may seek help from professionals in person

3:13 or seek informational help from credible experts or from Dr.

3:17 Google, as we heard about earlier today,

3:20 they may look towards their friends and peers,

3:22 community resources like sports teams and other communities

3:28 or cultural and spiritual practices through their churches, for instance.

3:34 What I want to talk today about, however,

3:38 is the fact that we've been learning more and more that teens

3:44 are turning to AI based conversational agents for emotional support, really.

3:50 I wouldn't argue, certainly,

3:53 there's not a lot of rigorous evidence in this space yet,

3:57 but I wouldn't argue that it's

3:59 necessarily replaced these other sources of support,

4:02 but I would say that I've mentioned the gaps,

4:06 and I think that AI is being leveraged to fill the gaps.

4:10 You can see here Harvard Business Review made the case

4:15 early in 2025 that therapy and companionship moved from being

4:20 much lower on the use case list in 2024

4:24 to becoming the top use case for generative AI in 2025.

4:31 You can see here all of these headlines are

4:33 from 2025 from a number of different news sources,

4:36 suggesting that this is certainly becoming a phenomenon that is receiving

4:44 a lot of attention in our society for both adults and teens alike.

4:52 But what I want to point out, too, is that we are starting to see evidence

4:57 that there's high use among youth in particular.

5:02 You can see some headlines up here that are particularly focused on youth.

5:06 The future is emotional AI, and Gen Z offers an early glimpse.

5:10 I myself have a Gen Zier, and what's I think notable too,

5:15 is that I think one of the first research studies came

5:19 out late last year in November this JAMA Network open paper,

5:23 really looking at the use of generative AI

5:25 for mental health advice among US adolescents and young adults.

5:29 This was a nationally representative survey of over 1,000 youth,

5:33 12-21 years and found that 13% reported

5:37 using generative AI for mental health advice,

5:40 and higher rates were noted from individuals

5:44 at the older end of that age range, 22%.

5:49 Among users, 65% sought advice monthly or more often than monthly.

5:55 Among users, over 90% found the advice somewhat or very helpful,

6:02 which I think is pretty interesting.

6:05 Then of course, continuing my review of headlines, tragically,

6:11 a number of very difficult and concerning

6:13 outcomes have been reported in the media,

6:15 including suicides which have been linked to the use

6:19 of generative AI for mental health purposes.

6:25 In fact, these are in many ways,

6:28 reflect lawsuits that are ongoing with major AI companies, most commonly OpenAI.

6:34 Then finally, anybody in here look to Common Sense Media

6:39 for advice about what to allow their kids to look at.

6:44 I'm somebody who does that.

6:46 I see a couple of hands out there.

6:48 In fact, Common Sense Media published a report recently.

6:53 You can look at the range of their findings,

6:56 but I think the top line one that I emphasize here is that the source that many

7:01 parents turn to found that mental health support is

7:04 one of the most common and in their perspective,

7:07 most dangerous ways that teens use AI.

7:09 I would argue that whether you see this trend as a natural disaster:

7:19 fire, flood, drought, or tornado, perhaps,

7:24 or you see it as just a ripple in the march forward of technology,

7:32 I would argue that the ecosystem has already changed.

7:35 One thing I have been thinking quite a lot about

7:39 with a number of my colleagues is what steps are needed

7:43 at this point to really guide this space and foster

7:48 the healthy pro social evolution of this landscape that now includes AI.

7:53 What I'm going to do next for the middle part of my talk is convey some ideas,

8:00 that's really the best word I have for them,

8:02 ideas for promoting youth mental health in the age of AI,

8:05 and I'm going to link each one of the five ideas that I present

8:09 to you to some work that I've been involved with in various parts of my work.

8:15 Those ideas are as follows.

8:18 I've listed them up here.

8:19 Listen to and reflect on the lived experiences of youth.

8:25 Two, develop research agendas that align with youth interests.

8:30 Three, assess shifting preferences among youth and adults while

8:36 seeking to understand fundamental values around mental health support.

8:42 Four, monitor harms, and five prepare the workforces needed to support

8:48 youth mental health in the age of AI.

8:51 We'll start with the first one,

8:52 listen to and reflect on the lived experiences of youth.

8:55 What I want to talk a little bit about is some work

8:58 that I know I've had a chance to speak with some of you about.

9:03 Early last year, I was fortunate to receive some funding through our Vice

9:08 Chancellor for research through what's called

9:11 the Convene and Influence Program here at UCSD.

9:14 This program is really focused on having

9:18 projects that advance the university's strategic themes.

9:21 In this case, it was the themes of health

9:23 and technology through work across campus and across departments.

9:30 We formed a cross campus working group with faculty from public health,

9:36 cognitive science, neuroscience, psychiatry.

9:39 You can see several of my colleagues up here, philosophy,

9:41 data science to pursue a number of things over the course of the year.

9:47 What I want to talk most about here this morning is actually a public symposium,

9:52 much like the one we're here at today

9:54 that we held late last year in December of 2025.

9:59 We were so fortunate,

10:01 and I just want to give a shout out to my incredible colleague, Dr.

10:05 Desiree Shapiro, who's an adolescent psychiatrist,

10:09 and this was really her brainchild.

10:11 But we were able in this meeting to host a youth panel on lived experience,

10:18 asking youth what is appealing and concerning about AI for emotional support.

10:23 We had five youth on the panel I think between the ages of, I want to say 14-19.

10:30 We asked what factors might make using AI for emotional support appealing.

10:35 What do you think are factors that might make young people

10:38 avoid it or features that would make AI feel safe and supportive?

10:44 This is not a research study, more just conveying to you what we

10:49 heard from this panel of lived experience youth.

10:54 We heard two things in response to the first question.

10:58 Overwhelmingly, the accessibility and availability of AI.

11:02 You can see here, when you said all you have to do is open

11:06 up your laptop and you have something that you can use to vent your feelings.

11:09 The availability and accessibility.

11:12 Second, the ability to express feelings without judgment.

11:16 This idea that you can say whatever you want, and nobody's going to judge you.

11:22 That was an overwhelming thing.

11:24 Sorry, this is cut off.

11:26 When we asked about concerns,

11:28 I think some of the concerns I heard from them were quite interesting.

11:32 One I'll just amplify here is a concern around substitution of human

11:39 connection with AI interaction and risk

11:42 of echo chambers and sycophantic responses,

11:45 as well as the last one that you can see here, number 5,

11:48 that AI may provide temporary relief but may undermine

11:52 the development of critical interpersonal and emotional regulation skills.

11:55 Really thinking about how might repeated or high use of AI

12:02 in this manner really affect the developing trajectory of youth.

12:07 You can see here, one youth reflecting on that, saying,

12:11 a real person is just so much more valuable because they can look you

12:14 in the face and call you out on what you're lying to them about.

12:18 Youth also mentioned they want safe and approachable

12:22 adults in addition to dialogue about AI safety.

12:25 These were the three areas that they mentioned

12:28 in terms of what they wanted to see from AI.

12:31 Improve AI credibility and transparency.

12:34 I'll let you take a look at the quotes.

12:35 Second, encourage adults to validate emotions and foster safe relationships.

12:40 Like if you don't want us to turn to AI, then adults,

12:44 you better get your stuff together and be the resource that we need.

12:53 Then, third, increase dialogue about AI safety and privacy into schools,

12:57 and I'll talk a little bit more about that in a moment.

13:00 I really wanted to emphasize that first point,

13:04 and then I'll go a little bit faster through the remaining four points,

13:08 but want to mention them nonetheless.

13:09 Second, develop research agendas.

13:11 In particular, this is focused on third

13:14 party researchers that align with youth interests.

13:17 This is a commentary that was actually developed by a postdoc in my lab, Dr.

13:22 Melissa Karnaze and really tries to articulate this case for the need

13:27 for third party research on human sought emotional support from AI.

13:31 This is currently in press at Nature Human Behavior.

13:35 The argument is like research researchers

13:38 vantage point complements that of companies

13:41 that are innovating and competing for market share and profits.

13:44 Very likely underestimating risks,

13:47 as well as governing bodies seeking to protect public,

13:50 possibly overestimating risks as well as shaped by politics.

13:55 Emphasizing the role for third party researchers.

13:59 This is really just scratching the surface

14:02 of what's included in this commentary.

14:03 In fact, I think there's over 30

14:06 next steps offered for researchers in this paper.

14:09 But you can see here the six reasons that are outlined,

14:13 and I'll emphasize number 2,

14:15 which is conversational AI tools offer novel emotional experiences.

14:20 I think that's likely particularly relevant for teens and one

14:24 key next step is to assess whether human relationships are undermined.

14:30 Really, as well, I think there are many articulated

14:33 reasons that align with the expressed concerns of youth.

14:37 The third idea, assess shifting preferences

14:40 while seeking to understand fundamental values.

14:43 I would say this is for youth as well as for all people.

14:46 This is a project that we pursued in the Center

14:50 for Empathy and Technology about a year ago,

14:53 and there's been a lot of actually data that's come out on this in the interim,

14:58 trying to get a sense of attitudes towards AI therapy.

15:02 We affectionately called this study the confident AI study.

15:08 We surveyed a convenience sample of US Internet

15:11 users about 130 in the fall of 2024, assessed uses of AI in the past year.

15:17 Then attitudes towards a hypothetical scenario of a conversational

15:22 AI model that was licensed to provide online therapy.

15:25 You can see here, these numbers, I think,

15:28 are fairly well aligned with what we're seeing

15:30 in news reports and other surveys that are coming

15:34 out in the gray literature suggesting about 10-20%

15:37 of people use AI for some type of social,

15:41 emotional support, and this was within the past year.

15:45 Then we asked, what would you want

15:47 from a hypothetical AI chatbot license to provide online therapy?

15:51 You can see some of the thematic areas that emerged.

15:54 I think adults, which is what we surveyed here,

15:59 but I would add youth to this want insight, advice, non judgment.

16:05 That's a big one and to feel better.

16:08 I think they're looking for resources that will accomplish that.

16:14 Then finally, we gave people a hypothetical

16:18 scenario that involved the experience of burnout

16:21 of a person named Alex and asked whether

16:24 they would choose a human or AI therapist, and you can see the percentages here.

16:28 It's about 60/40, 40% choosing a mental health chatbot.

16:34 Then we also asked about what are your concerns with AI therapist?

16:38 I think one of the interesting things for me,

16:40 seeing the blue box and the things that are

16:42 listed over there is you could easily ask somebody,

16:45 what are you concerned about with a human

16:47 therapist and you'd probably get a pretty similar list.

16:50 I would say here, preferences may shift,

16:55 but findings suggest that adults and youth

16:58 want to feel understood, safe, not judged.

17:00 I would say this is really the priority,

17:03 not necessarily whether the source of support is human or AI.

17:07 I would say, as somebody who's trained as a clinical psychologist and really

17:12 taught that the therapeutic relationship is

17:15 the cornerstone of the mechanism of intervention,

17:20 it's quite interesting to see these findings emerge.

17:24 I want to talk a little bit about monitoring harms, as well.

17:29 Here, I'm going to amplify the work

17:31 of one of my doctoral students Deepa Manjanatha.

17:34 Just convey that both looking to the literature to understand

17:40 the effectiveness of chatbot based

17:43 interventions is obviously critically important,

17:46 but monitoring of harms is critical, as well.

17:50 This was a systematic review of studies

17:54 on AI chatbots in adolescent mental health.

17:57 We had seen some reviews of the literature for interventions,

18:02 but none that were really focused

18:03 on adolescents and none that had systematically evaluated

18:07 each one of the studies for whether

18:09 or not they systematically evaluated the harms.

18:13 This project, after screening over 1,200 articles,

18:18 included 14 studies and included youth as well as college students.

18:24 They were mostly app based interventions,

18:26 looking at the effectiveness in treating depression,

18:29 anxiety, mood, affect, stress, well being, and body image.

18:34 What's interesting here is that not for every

18:38 outcome that they looked at, but roughly half

18:41 of the studies reported did report statistically significant

18:45 improvements in at least one target mental health domain.

18:48 But only two out of the 14 studies monitored and then discussed harms,

18:55 and those that did included worsening of depressive symptoms,

19:00 mental health related emergency department visits.

19:02 These were some of the harms that were reported

19:04 in these very few studies that looked at this.

19:06 I think really underscores the need to look at this.

19:13 Then finally, I'll make some comments about,

19:16 and I think this could be a much longer discussion,

19:21 certainly among this group of professionals,

19:24 but the need to prepare as well as care

19:27 for the workforces needed to support youth AI in the AI age.

19:31 I'm going to talk about they're a little bit disconnected,

19:34 but two different things.

19:35 One is the need to prepare future clinicians for patients who are using.

19:42 We've heard about many of them this morning,

19:45 ways AI is leveraged within the health system.

19:50 But I think there's also a need to prepare

19:52 clinicians for patients who are utilizing AI for mental

19:57 health purposes outside of the healthcare system and how

20:01 that interplays with clinicians who may be seeing patients.

20:05 This is a new elective that my colleague

20:08 and I are offering just this quarter to really

20:12 hopefully help prepare future health professionals for these discussions

20:17 about LLMs with their patients and their colleagues.

20:21 We try to do a variety of things in the course, and it will, of course,

20:25 undoubtedly be updated next year when we've had a chance to offer it this year.

20:31 Then the other thing I want to mention in terms of workforce, very briefly,

20:35 the role of our online safety workforce,

20:42 the trust and safety professionals, the people who moderate content online.

20:46 It's a little bit of a different project,

20:50 but I think this is work again, from my doctoral student Deepa Manjanatha.

20:55 We surveyed well over 600 individuals who are professionally working

21:00 as content moderators and tried to get a sense of various

21:08 associations between mental health outcomes and different elements of their work

21:13 and different ways in which they might be supported at work,

21:16 by their coworkers, by their bosses.

21:18 Do they work in person, do they work online, things like that.

21:21 I'm just showing you here a snippet of this work.

21:25 But as you can see here,

21:27 we assessed, depression, anxiety, burnout, post traumatic stress.

21:31 These individuals are exposed to the worst elements of what is on the Internet.

21:38 These outcomes are very real risks for them.

21:43 I think the point I want to make here is that with the acceleration of AI,

21:48 it becomes much easier to generate harmful content.

21:51 It can be done much faster in much larger volumes.

21:54 I think this digital mental health workforce,

21:59 becomes significantly more at risk, which I think puts everybody at risk,

22:03 including youth who are actively engaged in online spaces.

22:08 I'm going to take the minutes that I have left

22:11 and just make some comments about broader trends and some takeaways.

22:16 First, I'll just say that I do think there are some things from history

22:22 that can inform our thinking about the future

22:24 of generative AI for youth mental health.

22:27 One of them is that there's been

22:31 a longstanding interest in automating mental health treatment.

22:35 Some of you may recall the development of Eliza, one of the very first chat bots

22:43 that was developed to essentially simulate a Rogerian psychotherapist,

22:50 who for those of you who may not know,

22:52 Carl Rogers was incredible thought leader in the field

22:57 of clinical psychology and really the founder

23:00 of the idea of giving unconditional positive regard

23:03 to patients as a way to offer therapy.

23:07 You can imagine how AI chatbots might be very

23:12 well situated to take up that unconditional positive regard,

23:18 a framework of psychology versus,

23:22 say there maybe other areas of psychology that are not as well suited.

23:27 But I think this long history of interest in automating treatment,

23:31 this is another step down that road.

23:34 Similarly, there's the direct to consumer movement in medicine,

23:39 and I think I put up here direct to consumer genomics.

23:43 This was an area that I was quite involved in a decade or so ago,

23:46 but really was controversial in terms of the idea that people could have

23:51 direct access to their genetic information

23:53 without the involvement of a healthcare provider.

23:56 One way of thinking about the phenomenon that I've been talking about here

23:59 this morning is this idea of direct

24:03 access to this form of mental health support.

24:09 I think some of the things that we were talking

24:12 about in the late 2000 around direct to consumer health technologies,

24:16 we might be able to use some

24:17 of those frameworks to think about this phenomenon as well.

24:20 Then I think in addition many people more recently that I talked to about

24:26 this phenomenon often talk about the analogy

24:28 with social media and youth mental health,

24:31 which has been a very difficult space to think

24:36 through and also been very intractable from a policy standpoint.

24:40 In fact, in many ways,

24:41 I'm certainly not an expert on the policy space around AI,

24:48 but I think in many ways,

24:49 it seems to me from my vantage point that we've actually had a lot more policy

24:55 activity in the generative AI space than we

24:59 really saw go forward in the social media space.

25:03 You can see here our surgeon general, I think it was in 2003,

25:07 that finally a report came out noting that there should really be a public

25:16 health consideration for social media after over a decade of concern about this.

25:22 Then I'll also just say,

25:24 and this ties back into, I think some of the comments that we heard from youth.

25:28 I think the impact of local

25:32 level actions are perhaps sometimes underappreciated.

25:36 We heard from youth that having more discourse in their schools,

25:42 thinking about how they might engage

25:45 with mental health or emotional support through

25:47 schools and I had to have that discussion around AI within the school setting,

25:54 I think could be fruitful.

25:56 I'm just showing you this is actually in my group,

25:59 we also do some work on social media.

26:02 This is a paper that was published

26:04 by with my graduate student Alli O'Daffer where we offered

26:08 some recommendations to include schools a bit more thoughtfully

26:12 and intentionally when thinking about steps for social media.

26:16 I think these were our recommendations for social media,

26:19 but I think they could easily be

26:21 very straightforwardly repurposed for this AI space.

26:26 Then, finally, I think I don't need to tell this room that there's

26:32 a need to balance AI safety

26:34 with our ethical obligation to increase access to treatment.

26:38 We could all, I'm sure, think of many ethical considerations related to the use

26:43 of AI in the ways that I've been describing.

26:45 But we also, in my opinion, have an obligation to leverage what I think could be

26:51 a very promising technology for addressing the gaps that we've discussed.

26:55 I think being able to really hold both of these things

27:02 at once and determine a way forward is our ethical obligation.

27:07 I'll just close with some high level takeaways for you all

27:14 about even more meta than the five ideas that I presented.

27:20 I think this is a very difficult space,

27:23 and it's very fast moving, and our youth are changing.

27:28 My kids are changing right before my eyes,

27:30 and I'm sure that's the same for all of you.

27:33 I think work needs to go across sectors,

27:36 disciplines, professions, stakeholders, and generations.

27:40 This touches on the need to listen

27:42 to and reflect on the lived experiences of youth.

27:45 I also think there's a need, certainly from a research standpoint to be agile,

27:51 consider networked hubs in this space

27:53 that bring in all the different stakeholders that I just mentioned versus more

27:57 traditional structures that are perhaps more linear.

28:00 I think these hubs may be what's needed to figure

28:04 out how to have impact and then to amplify that impact.

28:09 Then also, as I mentioned, balance safety with the ethical obligation

28:13 to increase access to mental healthcare.

28:16 Then finally, I would just say to stay curious and strive

28:20 to understand what will likely be a constantly changing ecosystem.

28:27 That is one of the things I strive for in the work that I do.

28:31 I'd like to thank all of these folks and thank you for your time and attention.