WEBVTT
00:00:01.199 --> 00:00:16.079
So we're really thinking about this and you know, particularly in the way that we have built Tee, as certainly not a replacement, but also a way to kind of create that access point and help people get into care with a person when they need it.
00:00:16.079 --> 00:00:19.920
If they're more comfortable talking to, you know, kind of an AI solution.
00:00:19.920 --> 00:00:21.280
There's a lot of reasons for that.
00:00:21.280 --> 00:00:21.920
I get it.
00:00:21.920 --> 00:00:30.160
We're not we're not trying to kind of fight against that, but we also want to say, hey, actually at this point, you know, this is kind of clinically significant.
00:00:30.160 --> 00:00:31.519
There may be a safety concern.
00:00:31.519 --> 00:00:37.200
We want to help you now get into therapy and not just kind of send a referral.
00:00:37.200 --> 00:00:42.159
Um we want we want to kind of take that next step and kind of give people a really clear pathway into care.
00:00:43.520 --> 00:00:44.960
This is CrisisTalk.
00:00:44.960 --> 00:00:46.719
I'm your host, Stephanie Hepburn.
00:00:46.719 --> 00:00:48.000
A quick announcement.
00:00:48.000 --> 00:00:54.079
In its inaugural season, CrisisTalk has won gold at the fifth annual Signal Awards in the mental health genre.
00:00:54.079 --> 00:00:56.159
I'll put a link in the show notes.
00:00:56.159 --> 00:01:05.920
Today I'm speaking with Monika Dargis, head of clinical operations and quality at Talkspace, who is also leading Tee, Talkspace's new AI chatbot.
00:01:05.920 --> 00:01:10.560
It's designed to be a wellness tool alternative to general use large language models.
00:01:10.560 --> 00:01:15.680
Unlike broad use LLMs, Tee has clinicians on staff who can provide access to care.
00:01:15.680 --> 00:01:21.760
We discuss what happens when conversations escalate, AI safety, privacy, and user data.
00:01:21.760 --> 00:01:23.200
Let's jump in.
00:01:24.400 --> 00:01:26.239
My name is Monika Dargis.
00:01:26.239 --> 00:01:35.599
I am the head of clinical operations and quality at Talkspace, Chief Clinical Officer of Tee, and a licensed clinical psychologist by training.
00:01:36.319 --> 00:01:45.760
So the recent news is that Universal Health Services, which is a major hospital and healthcare provider, just acquired Talkspace.
00:01:45.760 --> 00:01:48.640
What should people be looking out for and what's going to be different?
00:01:48.959 --> 00:02:00.000
Yeah, this is something we're really excited about because I think there's a real opportunity here for us to close some of the known gaps that exist in our current mental health care system.
00:02:00.000 --> 00:02:09.520
So one of the things that we're really excited about that will come from this acquisition is really bringing that entire continuum of care in-house.
00:02:09.520 --> 00:02:30.000
So thinking about talks, offering outpatient-level telehealth care all the way up to UHS having the highest level of acuity care inpatient facilities, because we're now part of a much larger system that has a lot of different entry points already, it's offering a lot of different opportunities to get connected.
00:02:30.000 --> 00:02:38.400
So if someone, for example, presents to an intensive outpatient facility, but actually they could benefit from outpatient care.
00:02:38.400 --> 00:02:40.639
Right now UHS and Talkspace are connected.
00:02:40.639 --> 00:02:43.599
So it makes that referral into us a lot more smooth.
00:02:43.599 --> 00:03:08.560
The other thing that we're really thinking about, and I think this is really where Tee comes in, which is how can we use safe, purpose-driven AI tools to really expand the continuum of care so that folks who maybe don't need therapy, maybe they they kind of have a couple of stressors, some things they just kind of want to talk through, have a really easy, accessible resource to do that, particularly if they're not ready or don't yet need therapy.
00:03:08.560 --> 00:03:11.360
That's also a new access point that we're building.
00:03:11.680 --> 00:03:22.479
Being in the digital mental health space, how do you navigate as a company being abridged, but not usurping human care?
00:03:22.479 --> 00:03:31.039
It's a delicate balance, I realize, because the world of large language models has entered the space, whether we wanted it or not.
00:03:31.039 --> 00:03:32.719
People are turning to them.
00:03:32.719 --> 00:03:46.240
On one hand, you can make absolute statements and say people should not be turning to large language models, but we have to also just acknowledge that this is happening and that illustrates that there's potentially a need or a gap.
00:03:46.240 --> 00:03:48.560
I think that's a big discussion point too.
00:03:48.560 --> 00:04:01.280
I mean, somebody could push back and say not necessarily, just because people are turning to it doesn't necessarily mean that there are gaps in the system, but we have to find better ways to give people the resources they need through outreach.
00:04:01.280 --> 00:04:13.840
But from your perspective, how do you make sure the human element doesn't disappear as we are increasingly using digital tools?
00:04:14.159 --> 00:04:19.199
It's a really important question and one that I spend essentially 23 hours a day thinking about.
00:04:19.199 --> 00:04:21.600
Now, there's a couple of thoughts on this.
00:04:21.600 --> 00:04:25.519
One is we have now built Tee, which is this LLM.
00:04:25.519 --> 00:04:29.199
So we're able to reach a lot more people, which is really important.
00:04:29.199 --> 00:04:36.000
Um, and at no point is it our intention that this becomes a replacement for human care.
00:04:36.000 --> 00:04:37.519
I'm a psychologist myself.
00:04:37.519 --> 00:04:40.160
My role in Talkspace, I'm really leading our clinical network.
00:04:40.160 --> 00:04:42.879
I work with clinicians every single day.
00:04:42.879 --> 00:04:45.519
That is in no way, shape, or form the goal.
00:04:45.519 --> 00:04:48.079
So we're really not looking to build a substitute.
00:04:48.079 --> 00:05:03.759
Some of the ways we're thinking about ensuring that it doesn't unintentionally become that is we have, you know, Tee it is an LLM, but we actually have a pretty intricate back-end structure to Tee that involves clinicians in the loop.
00:05:03.759 --> 00:05:11.759
So we have a team of licensed clinicians working with Tee to ensure that users are kept safe, that escalations are made when necessary.
00:05:11.759 --> 00:05:18.319
And one of those escalations we can make is actually helping people get a referral into therapy when needed.
00:05:18.319 --> 00:05:32.959
So we're really thinking about this in, you know, particularly in the way that we have built Tee, as certainly not a replacement, but also a way to create that access point and help people get into care with a person when they need it.
00:05:32.959 --> 00:05:37.600
If they're more comfortable talking to an AI solution, there's a lot of reasons for that.
00:05:37.600 --> 00:05:38.240
I get it.
00:05:38.240 --> 00:05:40.079
But we're not trying to fight against that.
00:05:40.079 --> 00:05:44.720
But we also want to say, hey, at this point, this is kind of clinically significant.
00:05:44.720 --> 00:05:46.079
There may be a safety concern.
00:05:46.079 --> 00:05:51.279
We want to help you now get into therapy and not just send a referral.
00:05:51.279 --> 00:05:55.040
We want to take that next step and give people a really clear pathway into care.
00:05:55.360 --> 00:06:00.399
So with the general use large language models, there's these challenges, of course, right?
00:06:00.399 --> 00:06:05.759
Because, like you said, they're not humans who, or to our knowledge, right, there's not a whole lot of transparency.
00:06:05.759 --> 00:06:08.240
They are black boxes at this point.
00:06:08.240 --> 00:06:10.240
So we don't really know what's happening.
00:06:10.240 --> 00:06:31.839
But to our knowledge, there are not clinicians who are in the loop who are able to say, okay, this looks like the person is maybe experiencing symptoms of psychosis, or the person is experiencing suicidal ideation, or this conversation is actually exacerbating this person's OCD symptoms.
00:06:31.839 --> 00:06:38.639
How are you thinking about those pitfalls that come with the general large language models?
00:06:38.639 --> 00:06:49.199
Most of the mental health AI chatbots are, they may not be built as a wrapper around existing large language models, but they do utilize them in some part of their architecture.
00:06:49.199 --> 00:06:55.360
What kind of guardrails are you able to put in there to make sure that that's not happening?
00:06:55.839 --> 00:07:05.839
We really think about safety in a couple of ways and have put a lot of effort into making sure that we have several different levels of safety built in.
00:07:05.839 --> 00:07:12.879
So we're not just relying on the model to say the right thing, but I do think it starts with the data that we're training the models on.
00:07:12.879 --> 00:07:18.879
And so that is at the ground level, like talking to the engineers, really getting technical.
00:07:18.879 --> 00:07:30.399
We have licensed clinicians involved with that so that we have people with their relevant expertise to say, hey, this is an appropriate response if someone says X, this is not, and here's why.
00:07:30.399 --> 00:07:39.600
So we're really making some pretty nuanced and at times kind of difficult gray area decisions that are all being led by licensed clinicians on my team.
00:07:39.600 --> 00:07:50.079
So that is, I think, really where it starts is the core fundamental nuts and bolts of the model is really being trained and led by clinical oversight and clinical decisions.
00:07:50.079 --> 00:07:51.680
So that's step one.
00:07:51.680 --> 00:08:02.639
Step two is that we have additional models that essentially run on conversations that are trained to detect potential risks throughout a conversation.
00:08:02.639 --> 00:08:16.160
So if someone starts talking about suicidal ideation, if someone is trying to use Tee to seek reassurance in a way that is potentially problematic and unhelpful, you mentioned OCD.
00:08:16.160 --> 00:08:18.720
That's something we can see a lot with OCD clinically.
00:08:18.720 --> 00:08:20.800
We have models that can detect that.
00:08:20.800 --> 00:08:28.480
And when that is detected, it creates a ticket, which then leads us to our third level of protection, which is the clinician in the loop process.
00:08:28.480 --> 00:08:45.120
So we have licensed clinicians who will review that information and determine if there is indeed a clinically significant risk present, at which case they can, like I mentioned earlier, potentially get that person connected to an actual human therapist for the appropriate level of care.
00:08:45.360 --> 00:08:48.240
So you mentioned you have models that detect that.
00:08:48.240 --> 00:08:53.840
Is the detection done through an algorithm or keywords?
00:08:53.840 --> 00:08:56.399
Or is it more nuanced than that?
00:08:56.399 --> 00:09:06.320
The reason I'm asking is another thing that the large language models are not super great at is detecting the difference between explicit language and implicit language.
00:09:06.320 --> 00:09:08.320
And I'm sure you all have been thinking about that.
00:09:08.320 --> 00:09:11.039
I'm curious, what does that look like?
00:09:11.039 --> 00:09:13.279
Because keywords can be fallible.
00:09:13.279 --> 00:09:14.960
I'm curious how that's layered.
00:09:15.200 --> 00:09:16.879
It's a little bit of everything, frankly.
00:09:16.879 --> 00:09:22.399
What I can say though is that we are not relying solely on keywords, and for exactly the reason that you said, right?
00:09:22.399 --> 00:09:26.320
So much potential risk is not stated overtly.
00:09:26.320 --> 00:09:36.080
And so we really do have to read between the lines, understand the context, and pick up on the full picture of what someone is saying, which only relying on keywords is not going to do a very good job of.
00:09:36.080 --> 00:09:45.840
So we have other LLMs using machine learning, et cetera, so that we have a complex architecture going on on the back end to make sure that we're capturing that present.
00:09:46.159 --> 00:09:50.240
So when a conversation results in a ticket, what happens next?
00:09:50.240 --> 00:09:54.080
On the back end, you all have a team of clinicians.
00:09:54.080 --> 00:09:55.840
What does that look like for you?
00:09:55.840 --> 00:09:58.720
How do you decide which clinician picks that up?
00:09:58.720 --> 00:10:00.879
What does that look like organizationally?
00:10:01.200 --> 00:10:05.759
So big kudos to our engineering and product team because they have made this very easy for us.
00:10:05.759 --> 00:10:12.000
We have basically an integrated system that organizes tickets based on potential acuity.
00:10:12.000 --> 00:10:19.519
So higher risk tickets are prioritized as well as the time that the ticket was created to make sure that we're responding to things quickly.
00:10:19.519 --> 00:10:23.440
And then we have internal SOPs, policies, procedures, et cetera.
00:10:23.440 --> 00:10:25.679
I could spare you the weeds of that.
00:10:25.679 --> 00:10:33.279
But we're using technology to really support our clinicians to ensure that higher risk tickets are escalated and addressed first.
00:10:33.279 --> 00:10:39.360
Everyone sees each other's work so that we're all maintaining the same standards, we're communicating clearly.
00:10:39.360 --> 00:10:42.000
If a decision was made, it's clearly documented.
00:10:42.000 --> 00:10:47.679
And that's truly because we have such a great team that was able to build us a really amazing tool to do this work.
00:10:48.000 --> 00:10:53.120
And so then if a conversation is flagged, does the clinician outreach that individual?
00:10:53.120 --> 00:10:55.759
What does that look like on the user side?
00:10:56.000 --> 00:10:56.240
Yeah.
00:10:56.240 --> 00:10:58.879
So they can outreach the user if indicated.
00:10:58.879 --> 00:11:07.039
It can also, in case of a more acute emergency, we can reach out to local emergency services if indicated.
00:11:07.039 --> 00:11:13.039
We also have, like I mentioned earlier, a really easy way to refer into therapy and get them that referral information.
00:11:13.039 --> 00:11:23.759
So depending on the context, we have a handful of different pathways that we can take to ensure that we're taking one step outside of the system clinically to make the right decision for that individual.
00:11:24.080 --> 00:11:25.600
And then why a chatbot?
00:11:25.600 --> 00:11:27.759
So why did you develop Tee?
00:11:27.759 --> 00:11:30.639
What was the pressure point there?
00:11:30.960 --> 00:11:35.360
A part of it is what you mentioned earlier, which is that we already know people are using chatbots.
00:11:35.360 --> 00:11:41.039
And I think unfortunately, there have been a lot of pretty significant negative outcomes associated with that.
00:11:41.039 --> 00:11:47.840
And rather than right take this stance of, hey, don't do this, we actually thought about it and really thought, well, like, why are people doing this?
00:11:47.840 --> 00:11:51.039
And in my opinion, I think it is because there's a gap.
00:11:51.039 --> 00:11:54.960
Therapy can be still very hard to access for a lot of people.
00:11:54.960 --> 00:11:57.759
It's also stressful, it's scary.
00:11:57.759 --> 00:11:58.159
I don't know.
00:11:58.159 --> 00:12:00.799
Like there's so many reasons why someone might not make that step.
00:12:00.799 --> 00:12:04.320
It's not surprising to me that people are using LLMs in this way.
00:12:04.320 --> 00:12:10.559
So we really wanted to take a stance and say, hey, we're gonna figure out a way to do this safely.
00:12:10.559 --> 00:12:12.559
That was a big motivator for us.
00:12:12.799 --> 00:12:14.320
And then how do people access Tee?
00:12:14.720 --> 00:12:15.840
Right now it's a standalone.
00:12:15.840 --> 00:12:21.840
So they can go to our website to access the tool and create an account, and then they're good to go.
00:12:22.240 --> 00:12:24.879
You mentioned that data is used to train the models.
00:12:24.879 --> 00:12:29.039
If you could tell me a little bit what data is being used to train the models.
00:12:29.039 --> 00:13:02.720
So then, as research comes out, how do you reassess guardrails?
00:13:02.720 --> 00:13:10.480
Let's say there's new research, maybe indicating a conversation that surpasses this amount of time, or maybe it's more nuanced than that.
00:13:10.480 --> 00:13:15.840
This individual seems to be in a loop or this seems to be exacerbating their symptoms.
00:13:15.840 --> 00:13:22.320
You mentioned the clinical side, but as we learn, and the research is always a little bit delayed.
00:13:22.320 --> 00:13:23.919
We're always behind the eight-ball.
00:13:23.919 --> 00:13:29.840
So as research comes out, how do you reassess what those guardrails will look like?
00:13:30.159 --> 00:13:31.600
You know, it takes a village.
00:13:31.600 --> 00:13:42.080
Um, and I think this is one of the strengths that we have is we have an internal research team, we have product, engineer, clinical, and all of these nuanced brains are really working on this.
00:13:42.080 --> 00:13:52.960
And so whenever there is a new study that comes out, that's something that our team is talking about and really being thoughtful about how do we want to integrate this and can move pretty quickly on that internally.
00:13:52.960 --> 00:13:59.039
The other thing is we are actively working on research and really hoping to have some published peer-reviewed studies soon.
00:13:59.039 --> 00:14:08.080
And in the interim, we're able to really look at data internally so that we can get a better understanding of what's happening in real time so that we're able to make quick decisions.
00:14:08.080 --> 00:14:15.679
If we do notice a trend, that's something we can really monitor, track, and better understand rather than waiting for a peer-reviewed research cycle.
00:14:15.679 --> 00:14:20.080
Because speaking as someone who was formerly an academic, I understand the importance of it.
00:14:20.080 --> 00:14:21.600
And it also just takes a really long time.
00:14:21.919 --> 00:14:22.240
It does.
00:14:22.240 --> 00:14:23.360
It takes a really long time.
00:14:23.360 --> 00:14:31.120
So if somebody goes and they start using Tee, do they have to opt into their data being used for machine training?
00:14:31.120 --> 00:14:33.039
The reason I'm asking is with the large language models.
00:14:33.039 --> 00:14:42.000
Let's say I go on to Claude, I can, on the setting side, say, don't use this conversation to further train the model.
00:14:42.000 --> 00:14:43.600
It is a difficult topic.
00:14:43.600 --> 00:14:48.080
On one hand, people should have the right to say, no, I'm not opting in.
00:14:48.080 --> 00:14:54.799
And then I can also see the company desire to have people opt in because it helps to improve the product.
00:14:54.799 --> 00:14:59.200
So it's a difficult topic, but I'm curious at this moment, what can people do?
00:14:59.200 --> 00:15:04.320
Can they say, I do want to use this tool, but I don't want it to further train the model?
00:15:04.639 --> 00:15:06.320
Yeah, it's an important question.
00:15:06.320 --> 00:15:13.039
So when folks sign up, they are prompted to review all of our privacy informed consent policies.
00:15:13.039 --> 00:15:16.240
All of that information is readily accessible for them.
00:15:16.240 --> 00:15:19.919
I will triple check the most recent update on this.
00:15:19.919 --> 00:15:28.159
I think right now our consent process outlines that data can be used for training, but I will triple check and confirm that.
00:15:28.159 --> 00:15:29.519
So maybe I can follow up with you.
00:15:29.840 --> 00:15:49.519
Yeah, I guess the distinction there is whether they can opt out but still use the product, because that's where it becomes one of those things where to use the product, this is what you're agreeing to versus we would prefer if you did, but you can continue to use it and we will not save this data or we will not utilize it to further improve the model.
00:15:49.840 --> 00:15:50.639
Yeah, understood.
00:15:50.639 --> 00:15:53.360
To my knowledge, right now, we don't have a clear opt-out.
00:15:53.360 --> 00:15:56.080
And it's not that we never would.
00:15:56.080 --> 00:15:58.799
It's not necessarily like a hard stance we're taking.
00:15:58.799 --> 00:16:01.600
But as of today, I don't believe that there's an opt-out.
00:16:01.600 --> 00:16:02.720
But let me confirm that.
00:16:03.039 --> 00:16:10.000
So another challenge, again, with the large language models, because people are turning to them, is there is no HIPAA compliance, right?
00:16:10.000 --> 00:16:14.799
So there is no PHI, there is no protected health information.
00:16:14.799 --> 00:16:20.080
But in the behavioral health space and the mental health space, there is a far different bar.
00:16:20.080 --> 00:16:25.600
How do you navigate protecting information and what does that look like?
00:16:25.600 --> 00:16:40.720
And then, of course, you have the challenge of now that you are working with this large healthcare provider, how do you navigate that with also sharing that information with those who need it, like other clinicians or other providers?
00:16:41.039 --> 00:16:43.679
Yeah, Tee was really built as a wellness tool.
00:16:43.679 --> 00:16:47.519
And because it's a wellness tool, it's not considered healthcare.
00:16:47.519 --> 00:16:55.519
So we don't, strictly speaking, fall under the HIPAA guidelines, for example, that our core Talkspace therapy and psychiatry business does.
00:16:55.519 --> 00:16:59.200
With that being said, we take privacy very seriously.
00:16:59.200 --> 00:17:05.279
And so we have built the technical infrastructure of Tee to be HIPAA compliant.
00:17:05.279 --> 00:17:14.640
So we can't really say it's HIPAA compliant because it's not healthcare, but it was built with the same expectations that our sort of core Talkspace product was built.
00:17:14.640 --> 00:17:17.039
So that addresses that piece.
00:17:17.039 --> 00:17:27.039
And right now, because Tee is a standalone, we don't have built-in pathways necessarily for integrating someone's therapist, for example.
00:17:27.039 --> 00:17:32.480
Part of that is prioritizing privacy and the user being able to engage with the tool.
00:17:32.480 --> 00:17:41.200
And I definitely think there's a future in which we have a little bit more integration, maybe not with external providers, but certainly thinking about our own providers.
00:17:41.519 --> 00:17:51.759
With the large language models, the general use ones, if you're experiencing increased distress and it does notice that and responds, because that's not always a certainty.
00:17:51.759 --> 00:17:54.640
The safety nets you all have in place are more robust.
00:17:54.640 --> 00:17:59.359
But what happens when I test it out is it's also not super consistent.
00:17:59.359 --> 00:18:08.000
It's just that's another thing about generative technology is you're getting a customized response, not necessarily a standardized or uniform response.
00:18:08.000 --> 00:18:15.680
With Tee, your space is going to be more challenging because you want it to both be individualized, but also standardized.
00:18:15.680 --> 00:18:21.039
So when somebody's experiencing that, you mentioned the different tiers and then the tickets.
00:18:21.039 --> 00:18:31.119
But could you tell me a little bit on the user side, like if I was navigating Tee, what that would look like for me if I started experiencing increased distress?
00:18:31.119 --> 00:18:38.480
Is there a point where let's say they're a clinician after hours or it's three o'clock in the morning versus daytime?
00:18:38.480 --> 00:18:42.559
What's that layout in terms of, from the user perspective, what they see?
00:18:43.599 --> 00:18:52.480
When someone is using Tee, it's primarily a conversation, which I think for folks who are familiar with interacting with uh general LLMs may look kind of similar.
00:18:52.480 --> 00:18:54.720
Kudos to our product team because it's very pretty.
00:18:54.720 --> 00:19:01.680
So I think it looks very nice, but it's a conversation and the user can text, they can type what they're saying, they can also use voice.
00:19:01.680 --> 00:19:06.240
So it's sort of interactive in probably a way that feels familiar to a lot of folks at this stage.
00:19:06.240 --> 00:19:14.880
If there is a risk concern or clinical concern that's brought up throughout the course of the conversation, what a user would expect is that Tee is going to be responsive to that.
00:19:14.880 --> 00:19:24.240
Depending on what they're saying, Tee may express concern, Tee may provide them with resources, it may recommend calling 988, something of that nature.
00:19:24.240 --> 00:19:26.319
So that's all certainly in place.
00:19:26.319 --> 00:19:34.640
Speaking of the tiers, though, what's happening on the back end is let's say a risk algorithm picks up, hey, this is a risky situation, and a ticket is created.
00:19:34.640 --> 00:19:42.799
Depending on the level of acuity on the user side, what they will then see is essentially a pop-up message that says, hey, there's concern here.
00:19:42.799 --> 00:19:44.079
We're worried about your safety.
00:19:44.079 --> 00:19:48.319
Here are some immediate resources that we would recommend you reach out to.
00:19:48.319 --> 00:19:55.599
So that is really how we're balancing the user experience of Tee will be fairly individualized, but we also need to standardize safety.
00:19:55.599 --> 00:19:57.359
And that's one of the ways we're doing it.
00:19:57.359 --> 00:20:24.960
And one thing I can mention on that front, which is not yet implemented, but is actively in the process of being implemented, is an integration with a crisis resource so that if someone says something that is particularly concerning and acute while they're talking with Tee, they will be essentially prompted to very easily click a button and start a phone call, which will connect them directly with a live clinician who is a crisis counselor.
00:20:24.960 --> 00:20:27.519
So if they're in crisis at 3 a.m.
00:20:27.519 --> 00:20:34.640
and talking to Tee, we're really building out the infrastructure to make sure that they get connected to a live person in real time.
00:20:34.960 --> 00:20:40.160
Is there anything else that is on the horizon or that you hope to do soon with Tee?
00:20:40.799 --> 00:20:42.480
So many things that we hope to do with Tee.
00:20:42.480 --> 00:20:50.319
One of the things that I mentioned earlier, which we're actively working on every day, is the peer-reviewed research process.
00:20:50.319 --> 00:20:54.799
I was a faculty member in the beginning of my career working in clinical trials.
00:20:54.799 --> 00:20:56.960
Research is very near and dear to my heart.
00:20:56.960 --> 00:20:59.279
And it is so critically important in this field.
00:20:59.279 --> 00:21:00.880
Things are moving really quickly.
00:21:00.880 --> 00:21:02.319
There's so much uncertainty.
00:21:02.319 --> 00:21:03.519
There's so much we don't know.
00:21:03.519 --> 00:21:19.119
And so this is something that we are really prioritizing to make sure we're not just proving that Tee is safe and effective and that it's helpful, but that we're also contributing to a larger body of research for the field so that other people who are working in this space are able to learn from us.
00:21:19.119 --> 00:21:23.519
So that we're really setting a standard here for how mental health AI looks.
00:21:23.519 --> 00:21:26.000
So that's one thing that I'm really excited about.
00:21:26.000 --> 00:21:32.240
And then in general, just access, having this be a new entry point, potentially integrating into Talkspace.
00:21:32.240 --> 00:21:43.359
I mentioned the integration with crisis resources, really being creative about how we can ensure that people are staying safe, but that we're also getting them connected to care in a way that really hasn't been possible before.
00:21:43.680 --> 00:21:50.720
Are you all open to outside researchers looking at the data and testing the efficacy?
00:21:51.039 --> 00:21:51.920
Yes, please.
00:21:51.920 --> 00:21:55.200
If anyone is listening to this and is interested, please contact me.
00:21:55.200 --> 00:21:57.680
We're very open to that.
00:21:59.359 --> 00:22:11.759
That was Monika Dargis on Tee, Talkspace's new wellness AI chatbot, which was designed by clinicians as an alternative to broad-use LLMs and has clinicians on staff who can help provide users access to care.
00:22:11.759 --> 00:22:17.599
If you are a researcher interested in studying Tee, I've put Monica's contact information in the show notes.
00:22:17.599 --> 00:22:24.799
I did read Tee's terms of service and privacy policy, and at the time of this recording, there is no data opt-out.
00:22:24.799 --> 00:22:29.119
Meaning, if you use the tool, you can assume your data will be used to train the model.
00:22:29.119 --> 00:22:37.119
However, according to the privacy policy, you may request access to, correction of, or deletion of your personal data.
00:22:37.119 --> 00:22:44.799
As Monica mentioned, because Tee is a wellness tool, but not a mental health one, it does not fall under HIPAA like core talks-based services do.
00:22:44.799 --> 00:22:47.440
At CrisisTalk, we like to hear from our listeners.
00:22:47.440 --> 00:22:55.039
If you've used Tee or another wellness mental health or general use AI chatbot for mental health advice, let us know how it went.
00:22:55.039 --> 00:22:57.039
Our email is in the show notes.
00:22:57.039 --> 00:23:02.559
If you enjoyed this episode, please subscribe and leave us a review wherever you listen to the podcast.
00:23:02.559 --> 00:23:04.400
It helps others find the show.
00:23:04.400 --> 00:23:05.519
Thanks for listening.
00:23:05.519 --> 00:23:13.759
I'm your host and producer, our associate producer is Rin Koenig, Audio Engineering by Michael Keane, Music is Vinyl Couch by Blue Dot Sessions.
00:00:01.199 --> 00:00:16.079
So we're really thinking about this and you know, particularly in the way that we have built Tee, as certainly not a replacement, but also a way to kind of create that access point and help people get into care with a person when they need it.
00:00:16.079 --> 00:00:19.920
If they're more comfortable talking to, you know, kind of an AI solution.
00:00:19.920 --> 00:00:21.280
There's a lot of reasons for that.
00:00:21.280 --> 00:00:21.920
I get it.
00:00:21.920 --> 00:00:30.160
We're not we're not trying to kind of fight against that, but we also want to say, hey, actually at this point, you know, this is kind of clinically significant.
00:00:30.160 --> 00:00:31.519
There may be a safety concern.
00:00:31.519 --> 00:00:37.200
We want to help you now get into therapy and not just kind of send a referral.
00:00:37.200 --> 00:00:42.159
Um we want we want to kind of take that next step and kind of give people a really clear pathway into care.
00:00:43.520 --> 00:00:44.960
This is CrisisTalk.
00:00:44.960 --> 00:00:46.719
I'm your host, Stephanie Hepburn.
00:00:46.719 --> 00:00:48.000
A quick announcement.
00:00:48.000 --> 00:00:54.079
In its inaugural season, CrisisTalk has won gold at the fifth annual Signal Awards in the mental health genre.
00:00:54.079 --> 00:00:56.159
I'll put a link in the show notes.
00:00:56.159 --> 00:01:05.920
Today I'm speaking with Monika Dargis, head of clinical operations and quality at Talkspace, who is also leading Tee, Talkspace's new AI chatbot.
00:01:05.920 --> 00:01:10.560
It's designed to be a wellness tool alternative to general use large language models.
00:01:10.560 --> 00:01:15.680
Unlike broad use LLMs, Tee has clinicians on staff who can provide access to care.
00:01:15.680 --> 00:01:21.760
We discuss what happens when conversations escalate, AI safety, privacy, and user data.
00:01:21.760 --> 00:01:23.200
Let's jump in.
00:01:24.400 --> 00:01:26.239
My name is Monika Dargis.
00:01:26.239 --> 00:01:35.599
I am the head of clinical operations and quality at Talkspace, Chief Clinical Officer of Tee, and a licensed clinical psychologist by training.
00:01:36.319 --> 00:01:45.760
So the recent news is that Universal Health Services, which is a major hospital and healthcare provider, just acquired Talkspace.
00:01:45.760 --> 00:01:48.640
What should people be looking out for and what's going to be different?
00:01:48.959 --> 00:02:00.000
Yeah, this is something we're really excited about because I think there's a real opportunity here for us to close some of the known gaps that exist in our current mental health care system.
00:02:00.000 --> 00:02:09.520
So one of the things that we're really excited about that will come from this acquisition is really bringing that entire continuum of care in-house.
00:02:09.520 --> 00:02:30.000
So thinking about talks, offering outpatient-level telehealth care all the way up to UHS having the highest level of acuity care inpatient facilities, because we're now part of a much larger system that has a lot of different entry points already, it's offering a lot of different opportunities to get connected.
00:02:30.000 --> 00:02:38.400
So if someone, for example, presents to an intensive outpatient facility, but actually they could benefit from outpatient care.
00:02:38.400 --> 00:02:40.639
Right now UHS and Talkspace are connected.
00:02:40.639 --> 00:02:43.599
So it makes that referral into us a lot more smooth.
00:02:43.599 --> 00:03:08.560
The other thing that we're really thinking about, and I think this is really where Tee comes in, which is how can we use safe, purpose-driven AI tools to really expand the continuum of care so that folks who maybe don't need therapy, maybe they they kind of have a couple of stressors, some things they just kind of want to talk through, have a really easy, accessible resource to do that, particularly if they're not ready or don't yet need therapy.
00:03:08.560 --> 00:03:11.360
That's also a new access point that we're building.
00:03:11.680 --> 00:03:22.479
Being in the digital mental health space, how do you navigate as a company being abridged, but not usurping human care?
00:03:22.479 --> 00:03:31.039
It's a delicate balance, I realize, because the world of large language models has entered the space, whether we wanted it or not.
00:03:31.039 --> 00:03:32.719
People are turning to them.
00:03:32.719 --> 00:03:46.240
On one hand, you can make absolute statements and say people should not be turning to large language models, but we have to also just acknowledge that this is happening and that illustrates that there's potentially a need or a gap.
00:03:46.240 --> 00:03:48.560
I think that's a big discussion point too.
00:03:48.560 --> 00:04:01.280
I mean, somebody could push back and say not necessarily, just because people are turning to it doesn't necessarily mean that there are gaps in the system, but we have to find better ways to give people the resources they need through outreach.
00:04:01.280 --> 00:04:13.840
But from your perspective, how do you make sure the human element doesn't disappear as we are increasingly using digital tools?
00:04:14.159 --> 00:04:19.199
It's a really important question and one that I spend essentially 23 hours a day thinking about.
00:04:19.199 --> 00:04:21.600
Now, there's a couple of thoughts on this.
00:04:21.600 --> 00:04:25.519
One is we have now built Tee, which is this LLM.
00:04:25.519 --> 00:04:29.199
So we're able to reach a lot more people, which is really important.
00:04:29.199 --> 00:04:36.000
Um, and at no point is it our intention that this becomes a replacement for human care.
00:04:36.000 --> 00:04:37.519
I'm a psychologist myself.
00:04:37.519 --> 00:04:40.160
My role in Talkspace, I'm really leading our clinical network.
00:04:40.160 --> 00:04:42.879
I work with clinicians every single day.
00:04:42.879 --> 00:04:45.519
That is in no way, shape, or form the goal.
00:04:45.519 --> 00:04:48.079
So we're really not looking to build a substitute.
00:04:48.079 --> 00:05:03.759
Some of the ways we're thinking about ensuring that it doesn't unintentionally become that is we have, you know, Tee it is an LLM, but we actually have a pretty intricate back-end structure to Tee that involves clinicians in the loop.
00:05:03.759 --> 00:05:11.759
So we have a team of licensed clinicians working with Tee to ensure that users are kept safe, that escalations are made when necessary.
00:05:11.759 --> 00:05:18.319
And one of those escalations we can make is actually helping people get a referral into therapy when needed.
00:05:18.319 --> 00:05:32.959
So we're really thinking about this in, you know, particularly in the way that we have built Tee, as certainly not a replacement, but also a way to create that access point and help people get into care with a person when they need it.
00:05:32.959 --> 00:05:37.600
If they're more comfortable talking to an AI solution, there's a lot of reasons for that.
00:05:37.600 --> 00:05:38.240
I get it.
00:05:38.240 --> 00:05:40.079
But we're not trying to fight against that.
00:05:40.079 --> 00:05:44.720
But we also want to say, hey, at this point, this is kind of clinically significant.
00:05:44.720 --> 00:05:46.079
There may be a safety concern.
00:05:46.079 --> 00:05:51.279
We want to help you now get into therapy and not just send a referral.
00:05:51.279 --> 00:05:55.040
We want to take that next step and give people a really clear pathway into care.
00:05:55.360 --> 00:06:00.399
So with the general use large language models, there's these challenges, of course, right?
00:06:00.399 --> 00:06:05.759
Because, like you said, they're not humans who, or to our knowledge, right, there's not a whole lot of transparency.
00:06:05.759 --> 00:06:08.240
They are black boxes at this point.
00:06:08.240 --> 00:06:10.240
So we don't really know what's happening.
00:06:10.240 --> 00:06:31.839
But to our knowledge, there are not clinicians who are in the loop who are able to say, okay, this looks like the person is maybe experiencing symptoms of psychosis, or the person is experiencing suicidal ideation, or this conversation is actually exacerbating this person's OCD symptoms.
00:06:31.839 --> 00:06:38.639
How are you thinking about those pitfalls that come with the general large language models?
00:06:38.639 --> 00:06:49.199
Most of the mental health AI chatbots are, they may not be built as a wrapper around existing large language models, but they do utilize them in some part of their architecture.
00:06:49.199 --> 00:06:55.360
What kind of guardrails are you able to put in there to make sure that that's not happening?
00:06:55.839 --> 00:07:05.839
We really think about safety in a couple of ways and have put a lot of effort into making sure that we have several different levels of safety built in.
00:07:05.839 --> 00:07:12.879
So we're not just relying on the model to say the right thing, but I do think it starts with the data that we're training the models on.
00:07:12.879 --> 00:07:18.879
And so that is at the ground level, like talking to the engineers, really getting technical.
00:07:18.879 --> 00:07:30.399
We have licensed clinicians involved with that so that we have people with their relevant expertise to say, hey, this is an appropriate response if someone says X, this is not, and here's why.
00:07:30.399 --> 00:07:39.600
So we're really making some pretty nuanced and at times kind of difficult gray area decisions that are all being led by licensed clinicians on my team.
00:07:39.600 --> 00:07:50.079
So that is, I think, really where it starts is the core fundamental nuts and bolts of the model is really being trained and led by clinical oversight and clinical decisions.
00:07:50.079 --> 00:07:51.680
So that's step one.
00:07:51.680 --> 00:08:02.639
Step two is that we have additional models that essentially run on conversations that are trained to detect potential risks throughout a conversation.
00:08:02.639 --> 00:08:16.160
So if someone starts talking about suicidal ideation, if someone is trying to use Tee to seek reassurance in a way that is potentially problematic and unhelpful, you mentioned OCD.
00:08:16.160 --> 00:08:18.720
That's something we can see a lot with OCD clinically.
00:08:18.720 --> 00:08:20.800
We have models that can detect that.
00:08:20.800 --> 00:08:28.480
And when that is detected, it creates a ticket, which then leads us to our third level of protection, which is the clinician in the loop process.
00:08:28.480 --> 00:08:45.120
So we have licensed clinicians who will review that information and determine if there is indeed a clinically significant risk present, at which case they can, like I mentioned earlier, potentially get that person connected to an actual human therapist for the appropriate level of care.
00:08:45.360 --> 00:08:48.240
So you mentioned you have models that detect that.
00:08:48.240 --> 00:08:53.840
Is the detection done through an algorithm or keywords?
00:08:53.840 --> 00:08:56.399
Or is it more nuanced than that?
00:08:56.399 --> 00:09:06.320
The reason I'm asking is another thing that the large language models are not super great at is detecting the difference between explicit language and implicit language.
00:09:06.320 --> 00:09:08.320
And I'm sure you all have been thinking about that.
00:09:08.320 --> 00:09:11.039
I'm curious, what does that look like?
00:09:11.039 --> 00:09:13.279
Because keywords can be fallible.
00:09:13.279 --> 00:09:14.960
I'm curious how that's layered.
00:09:15.200 --> 00:09:16.879
It's a little bit of everything, frankly.
00:09:16.879 --> 00:09:22.399
What I can say though is that we are not relying solely on keywords, and for exactly the reason that you said, right?
00:09:22.399 --> 00:09:26.320
So much potential risk is not stated overtly.
00:09:26.320 --> 00:09:36.080
And so we really do have to read between the lines, understand the context, and pick up on the full picture of what someone is saying, which only relying on keywords is not going to do a very good job of.
00:09:36.080 --> 00:09:45.840
So we have other LLMs using machine learning, et cetera, so that we have a complex architecture going on on the back end to make sure that we're capturing that present.
00:09:46.159 --> 00:09:50.240
So when a conversation results in a ticket, what happens next?
00:09:50.240 --> 00:09:54.080
On the back end, you all have a team of clinicians.
00:09:54.080 --> 00:09:55.840
What does that look like for you?
00:09:55.840 --> 00:09:58.720
How do you decide which clinician picks that up?
00:09:58.720 --> 00:10:00.879
What does that look like organizationally?
00:10:01.200 --> 00:10:05.759
So big kudos to our engineering and product team because they have made this very easy for us.
00:10:05.759 --> 00:10:12.000
We have basically an integrated system that organizes tickets based on potential acuity.
00:10:12.000 --> 00:10:19.519
So higher risk tickets are prioritized as well as the time that the ticket was created to make sure that we're responding to things quickly.
00:10:19.519 --> 00:10:23.440
And then we have internal SOPs, policies, procedures, et cetera.
00:10:23.440 --> 00:10:25.679
I could spare you the weeds of that.
00:10:25.679 --> 00:10:33.279
But we're using technology to really support our clinicians to ensure that higher risk tickets are escalated and addressed first.
00:10:33.279 --> 00:10:39.360
Everyone sees each other's work so that we're all maintaining the same standards, we're communicating clearly.
00:10:39.360 --> 00:10:42.000
If a decision was made, it's clearly documented.
00:10:42.000 --> 00:10:47.679
And that's truly because we have such a great team that was able to build us a really amazing tool to do this work.
00:10:48.000 --> 00:10:53.120
And so then if a conversation is flagged, does the clinician outreach that individual?
00:10:53.120 --> 00:10:55.759
What does that look like on the user side?
00:10:56.000 --> 00:10:56.240
Yeah.
00:10:56.240 --> 00:10:58.879
So they can outreach the user if indicated.
00:10:58.879 --> 00:11:07.039
It can also, in case of a more acute emergency, we can reach out to local emergency services if indicated.
00:11:07.039 --> 00:11:13.039
We also have, like I mentioned earlier, a really easy way to refer into therapy and get them that referral information.
00:11:13.039 --> 00:11:23.759
So depending on the context, we have a handful of different pathways that we can take to ensure that we're taking one step outside of the system clinically to make the right decision for that individual.
00:11:24.080 --> 00:11:25.600
And then why a chatbot?
00:11:25.600 --> 00:11:27.759
So why did you develop Tee?
00:11:27.759 --> 00:11:30.639
What was the pressure point there?
00:11:30.960 --> 00:11:35.360
A part of it is what you mentioned earlier, which is that we already know people are using chatbots.
00:11:35.360 --> 00:11:41.039
And I think unfortunately, there have been a lot of pretty significant negative outcomes associated with that.
00:11:41.039 --> 00:11:47.840
And rather than right take this stance of, hey, don't do this, we actually thought about it and really thought, well, like, why are people doing this?
00:11:47.840 --> 00:11:51.039
And in my opinion, I think it is because there's a gap.
00:11:51.039 --> 00:11:54.960
Therapy can be still very hard to access for a lot of people.
00:11:54.960 --> 00:11:57.759
It's also stressful, it's scary.
00:11:57.759 --> 00:11:58.159
I don't know.
00:11:58.159 --> 00:12:00.799
Like there's so many reasons why someone might not make that step.
00:12:00.799 --> 00:12:04.320
It's not surprising to me that people are using LLMs in this way.
00:12:04.320 --> 00:12:10.559
So we really wanted to take a stance and say, hey, we're gonna figure out a way to do this safely.
00:12:10.559 --> 00:12:12.559
That was a big motivator for us.
00:12:12.799 --> 00:12:14.320
And then how do people access Tee?
00:12:14.720 --> 00:12:15.840
Right now it's a standalone.
00:12:15.840 --> 00:12:21.840
So they can go to our website to access the tool and create an account, and then they're good to go.
00:12:22.240 --> 00:12:24.879
You mentioned that data is used to train the models.
00:12:24.879 --> 00:12:29.039
If you could tell me a little bit what data is being used to train the models.
00:12:29.039 --> 00:13:02.720
So then, as research comes out, how do you reassess guardrails?
00:13:02.720 --> 00:13:10.480
Let's say there's new research, maybe indicating a conversation that surpasses this amount of time, or maybe it's more nuanced than that.
00:13:10.480 --> 00:13:15.840
This individual seems to be in a loop or this seems to be exacerbating their symptoms.
00:13:15.840 --> 00:13:22.320
You mentioned the clinical side, but as we learn, and the research is always a little bit delayed.
00:13:22.320 --> 00:13:23.919
We're always behind the eight-ball.
00:13:23.919 --> 00:13:29.840
So as research comes out, how do you reassess what those guardrails will look like?
00:13:30.159 --> 00:13:31.600
You know, it takes a village.
00:13:31.600 --> 00:13:42.080
Um, and I think this is one of the strengths that we have is we have an internal research team, we have product, engineer, clinical, and all of these nuanced brains are really working on this.
00:13:42.080 --> 00:13:52.960
And so whenever there is a new study that comes out, that's something that our team is talking about and really being thoughtful about how do we want to integrate this and can move pretty quickly on that internally.
00:13:52.960 --> 00:13:59.039
The other thing is we are actively working on research and really hoping to have some published peer-reviewed studies soon.
00:13:59.039 --> 00:14:08.080
And in the interim, we're able to really look at data internally so that we can get a better understanding of what's happening in real time so that we're able to make quick decisions.
00:14:08.080 --> 00:14:15.679
If we do notice a trend, that's something we can really monitor, track, and better understand rather than waiting for a peer-reviewed research cycle.
00:14:15.679 --> 00:14:20.080
Because speaking as someone who was formerly an academic, I understand the importance of it.
00:14:20.080 --> 00:14:21.600
And it also just takes a really long time.
00:14:21.919 --> 00:14:22.240
It does.
00:14:22.240 --> 00:14:23.360
It takes a really long time.
00:14:23.360 --> 00:14:31.120
So if somebody goes and they start using Tee, do they have to opt into their data being used for machine training?
00:14:31.120 --> 00:14:33.039
The reason I'm asking is with the large language models.
00:14:33.039 --> 00:14:42.000
Let's say I go on to Claude, I can, on the setting side, say, don't use this conversation to further train the model.
00:14:42.000 --> 00:14:43.600
It is a difficult topic.
00:14:43.600 --> 00:14:48.080
On one hand, people should have the right to say, no, I'm not opting in.
00:14:48.080 --> 00:14:54.799
And then I can also see the company desire to have people opt in because it helps to improve the product.
00:14:54.799 --> 00:14:59.200
So it's a difficult topic, but I'm curious at this moment, what can people do?
00:14:59.200 --> 00:15:04.320
Can they say, I do want to use this tool, but I don't want it to further train the model?
00:15:04.639 --> 00:15:06.320
Yeah, it's an important question.
00:15:06.320 --> 00:15:13.039
So when folks sign up, they are prompted to review all of our privacy informed consent policies.
00:15:13.039 --> 00:15:16.240
All of that information is readily accessible for them.
00:15:16.240 --> 00:15:19.919
I will triple check the most recent update on this.
00:15:19.919 --> 00:15:28.159
I think right now our consent process outlines that data can be used for training, but I will triple check and confirm that.
00:15:28.159 --> 00:15:29.519
So maybe I can follow up with you.
00:15:29.840 --> 00:15:49.519
Yeah, I guess the distinction there is whether they can opt out but still use the product, because that's where it becomes one of those things where to use the product, this is what you're agreeing to versus we would prefer if you did, but you can continue to use it and we will not save this data or we will not utilize it to further improve the model.
00:15:49.840 --> 00:15:50.639
Yeah, understood.
00:15:50.639 --> 00:15:53.360
To my knowledge, right now, we don't have a clear opt-out.
00:15:53.360 --> 00:15:56.080
And it's not that we never would.
00:15:56.080 --> 00:15:58.799
It's not necessarily like a hard stance we're taking.
00:15:58.799 --> 00:16:01.600
But as of today, I don't believe that there's an opt-out.
00:16:01.600 --> 00:16:02.720
But let me confirm that.
00:16:03.039 --> 00:16:10.000
So another challenge, again, with the large language models, because people are turning to them, is there is no HIPAA compliance, right?
00:16:10.000 --> 00:16:14.799
So there is no PHI, there is no protected health information.
00:16:14.799 --> 00:16:20.080
But in the behavioral health space and the mental health space, there is a far different bar.
00:16:20.080 --> 00:16:25.600
How do you navigate protecting information and what does that look like?
00:16:25.600 --> 00:16:40.720
And then, of course, you have the challenge of now that you are working with this large healthcare provider, how do you navigate that with also sharing that information with those who need it, like other clinicians or other providers?
00:16:41.039 --> 00:16:43.679
Yeah, Tee was really built as a wellness tool.
00:16:43.679 --> 00:16:47.519
And because it's a wellness tool, it's not considered healthcare.
00:16:47.519 --> 00:16:55.519
So we don't, strictly speaking, fall under the HIPAA guidelines, for example, that our core Talkspace therapy and psychiatry business does.
00:16:55.519 --> 00:16:59.200
With that being said, we take privacy very seriously.
00:16:59.200 --> 00:17:05.279
And so we have built the technical infrastructure of Tee to be HIPAA compliant.
00:17:05.279 --> 00:17:14.640
So we can't really say it's HIPAA compliant because it's not healthcare, but it was built with the same expectations that our sort of core Talkspace product was built.
00:17:14.640 --> 00:17:17.039
So that addresses that piece.
00:17:17.039 --> 00:17:27.039
And right now, because Tee is a standalone, we don't have built-in pathways necessarily for integrating someone's therapist, for example.
00:17:27.039 --> 00:17:32.480
Part of that is prioritizing privacy and the user being able to engage with the tool.
00:17:32.480 --> 00:17:41.200
And I definitely think there's a future in which we have a little bit more integration, maybe not with external providers, but certainly thinking about our own providers.
00:17:41.519 --> 00:17:51.759
With the large language models, the general use ones, if you're experiencing increased distress and it does notice that and responds, because that's not always a certainty.
00:17:51.759 --> 00:17:54.640
The safety nets you all have in place are more robust.
00:17:54.640 --> 00:17:59.359
But what happens when I test it out is it's also not super consistent.
00:17:59.359 --> 00:18:08.000
It's just that's another thing about generative technology is you're getting a customized response, not necessarily a standardized or uniform response.
00:18:08.000 --> 00:18:15.680
With Tee, your space is going to be more challenging because you want it to both be individualized, but also standardized.
00:18:15.680 --> 00:18:21.039
So when somebody's experiencing that, you mentioned the different tiers and then the tickets.
00:18:21.039 --> 00:18:31.119
But could you tell me a little bit on the user side, like if I was navigating Tee, what that would look like for me if I started experiencing increased distress?
00:18:31.119 --> 00:18:38.480
Is there a point where let's say they're a clinician after hours or it's three o'clock in the morning versus daytime?
00:18:38.480 --> 00:18:42.559
What's that layout in terms of, from the user perspective, what they see?
00:18:43.599 --> 00:18:52.480
When someone is using Tee, it's primarily a conversation, which I think for folks who are familiar with interacting with uh general LLMs may look kind of similar.
00:18:52.480 --> 00:18:54.720
Kudos to our product team because it's very pretty.
00:18:54.720 --> 00:19:01.680
So I think it looks very nice, but it's a conversation and the user can text, they can type what they're saying, they can also use voice.
00:19:01.680 --> 00:19:06.240
So it's sort of interactive in probably a way that feels familiar to a lot of folks at this stage.
00:19:06.240 --> 00:19:14.880
If there is a risk concern or clinical concern that's brought up throughout the course of the conversation, what a user would expect is that Tee is going to be responsive to that.
00:19:14.880 --> 00:19:24.240
Depending on what they're saying, Tee may express concern, Tee may provide them with resources, it may recommend calling 988, something of that nature.
00:19:24.240 --> 00:19:26.319
So that's all certainly in place.
00:19:26.319 --> 00:19:34.640
Speaking of the tiers, though, what's happening on the back end is let's say a risk algorithm picks up, hey, this is a risky situation, and a ticket is created.
00:19:34.640 --> 00:19:42.799
Depending on the level of acuity on the user side, what they will then see is essentially a pop-up message that says, hey, there's concern here.
00:19:42.799 --> 00:19:44.079
We're worried about your safety.
00:19:44.079 --> 00:19:48.319
Here are some immediate resources that we would recommend you reach out to.
00:19:48.319 --> 00:19:55.599
So that is really how we're balancing the user experience of Tee will be fairly individualized, but we also need to standardize safety.
00:19:55.599 --> 00:19:57.359
And that's one of the ways we're doing it.
00:19:57.359 --> 00:20:24.960
And one thing I can mention on that front, which is not yet implemented, but is actively in the process of being implemented, is an integration with a crisis resource so that if someone says something that is particularly concerning and acute while they're talking with Tee, they will be essentially prompted to very easily click a button and start a phone call, which will connect them directly with a live clinician who is a crisis counselor.
00:20:24.960 --> 00:20:27.519
So if they're in crisis at 3 a.m.
00:20:27.519 --> 00:20:34.640
and talking to Tee, we're really building out the infrastructure to make sure that they get connected to a live person in real time.
00:20:34.960 --> 00:20:40.160
Is there anything else that is on the horizon or that you hope to do soon with Tee?
00:20:40.799 --> 00:20:42.480
So many things that we hope to do with Tee.
00:20:42.480 --> 00:20:50.319
One of the things that I mentioned earlier, which we're actively working on every day, is the peer-reviewed research process.
00:20:50.319 --> 00:20:54.799
I was a faculty member in the beginning of my career working in clinical trials.
00:20:54.799 --> 00:20:56.960
Research is very near and dear to my heart.
00:20:56.960 --> 00:20:59.279
And it is so critically important in this field.
00:20:59.279 --> 00:21:00.880
Things are moving really quickly.
00:21:00.880 --> 00:21:02.319
There's so much uncertainty.
00:21:02.319 --> 00:21:03.519
There's so much we don't know.
00:21:03.519 --> 00:21:19.119
And so this is something that we are really prioritizing to make sure we're not just proving that Tee is safe and effective and that it's helpful, but that we're also contributing to a larger body of research for the field so that other people who are working in this space are able to learn from us.
00:21:19.119 --> 00:21:23.519
So that we're really setting a standard here for how mental health AI looks.
00:21:23.519 --> 00:21:26.000
So that's one thing that I'm really excited about.
00:21:26.000 --> 00:21:32.240
And then in general, just access, having this be a new entry point, potentially integrating into Talkspace.
00:21:32.240 --> 00:21:43.359
I mentioned the integration with crisis resources, really being creative about how we can ensure that people are staying safe, but that we're also getting them connected to care in a way that really hasn't been possible before.
00:21:43.680 --> 00:21:50.720
Are you all open to outside researchers looking at the data and testing the efficacy?
00:21:51.039 --> 00:21:51.920
Yes, please.
00:21:51.920 --> 00:21:55.200
If anyone is listening to this and is interested, please contact me.
00:21:55.200 --> 00:21:57.680
We're very open to that.
00:21:59.359 --> 00:22:11.759
That was Monika Dargis on Tee, Talkspace's new wellness AI chatbot, which was designed by clinicians as an alternative to broad-use LLMs and has clinicians on staff who can help provide users access to care.
00:22:11.759 --> 00:22:17.599
If you are a researcher interested in studying Tee, I've put Monica's contact information in the show notes.
00:22:17.599 --> 00:22:24.799
I did read Tee's terms of service and privacy policy, and at the time of this recording, there is no data opt-out.
00:22:24.799 --> 00:22:29.119
Meaning, if you use the tool, you can assume your data will be used to train the model.
00:22:29.119 --> 00:22:37.119
However, according to the privacy policy, you may request access to, correction of, or deletion of your personal data.
00:22:37.119 --> 00:22:44.799
As Monica mentioned, because Tee is a wellness tool, but not a mental health one, it does not fall under HIPAA like core talks-based services do.
00:22:44.799 --> 00:22:47.440
At CrisisTalk, we like to hear from our listeners.
00:22:47.440 --> 00:22:55.039
If you've used Tee or another wellness mental health or general use AI chatbot for mental health advice, let us know how it went.
00:22:55.039 --> 00:22:57.039
Our email is in the show notes.
00:22:57.039 --> 00:23:02.559
If you enjoyed this episode, please subscribe and leave us a review wherever you listen to the podcast.
00:23:02.559 --> 00:23:04.400
It helps others find the show.
00:23:04.400 --> 00:23:05.519
Thanks for listening.
00:23:05.519 --> 00:23:13.759
I'm your host and producer, our associate producer is Rin Koenig, Audio Engineering by Michael Keane, Music is Vinyl Couch by Blue Dot Sessions.
