Human AI Collaboration Enables More Empathic Conversations in Mental Health Support

Summary

This presentation explores the intersection of AI and empathy, particularly in the context of mental health support. It discusses the potential of AI to enhance access and quality of mental health care, focusing on peer support platforms like TalkLife. The speaker introduces a project that uses AI to provide real-time feedback to improve the empathy of peer supporters' responses. Emphasizing ethical considerations, the talk concludes with a randomized control trial demonstrating the effectiveness of AI-assisted empathy in peer support.

Takeaways

• 🤖 The talk focuses on the intersection of AI and empathy, particularly in the context of mental health support.
• 🔍 The speaker discusses the potential of AI to enhance access and quality of mental health support, while also addressing the associated risks and ethical considerations.
• ⚠️ A content warning is given for anonymized examples involving mental illness, self-harm, and suicidal ideation, used to illustrate real-world challenges.
• 🤝 The project is a collaboration involving the e-Science Institute, the Garvey Institute at UW, AI2, and a peer support platform called TalkLife.
• 🧠 The importance of empathy in mental health support is highlighted, with AI being explored as a tool to improve empathy in peer support interactions.
• 📊 Empathy is measured using adapted psychological scales, focusing on emotional reactions, interpretations, and explorations within text-based communication.
• 💬 AI techniques are used to analyze and enhance the expression of empathy in responses to individuals seeking support on platforms like TalkLife.
• 💡 The system 'PARTNER' is introduced, which uses reinforcement learning to rewrite responses to increase their empathic content.
• 📈 A randomized control trial demonstrates that AI-assisted feedback can significantly improve the level of empathy expressed by peer supporters.
• 🛡️ Ethical considerations are paramount, with a focus on co-design, informed consent, safety reviews, and ensuring the AI system does not disrupt the human connection in peer support.

Q & A

What is the main focus of the research discussed in the transcript?

The main focus of the research is the intersection of AI and empathy, specifically exploring how AI can help improve access and quality of mental health support through peer support platforms.

Why is empathy important in the context of mental health support?

Empathy is crucial in mental health support because it is associated with symptom improvement and the formation of positive, successful relationships. It allows for better understanding and feeling of the emotions and experiences of others, which is key in providing effective support.

What is the role of peer support platforms in mental health?

Peer support platforms serve as a place where individuals can find others who share similar experiences and challenges. They can complement traditional forms of care like therapy and provide a space for people to access support, especially when they might not have access to professional mental health services.

How does the AI system provide feedback to enhance empathy in peer supporters?

The AI system provides feedback by analyzing responses and suggesting edits that can increase the level of empathy expressed. It uses reinforcement learning to transform lower empathy posts into responses with higher empathy levels, focusing on emotional reactions, interpretations, and explorations.

What are the potential risks associated with AI technology in mental health, as mentioned in the transcript?

Potential risks include the disruption of meaningful human connections, the possibility of invalidating or harming individuals in crisis, and the ethical considerations of introducing AI into sensitive areas like mental health support.

How was the AI system in the study designed to minimize risks?

The AI system was designed to minimize risks by co-designing with stakeholders, focusing on a narrow scope of empathy, ensuring human supervision, and providing minimal and context-specific feedback only when needed. It also allowed peer supporters to flag and control the feedback process.

What was the outcome of the randomized control trial mentioned in the transcript?

The randomized control trial showed that the human-AI collaboration approach was effective in expressing more empathy. Participants who had access to AI feedback expressed substantially more empathy in their responses compared to the control group.

What ethical framework was referenced in the transcript for guiding the development of AI in mental health?

The ethical framework referenced is by Camille Nebeker and others at recode health, which helped inform the concrete actions taken to ensure the safety and welfare of participants in the study.

How does the transcript suggest the future of AI in mental health support?

The transcript suggests that AI-based feedback could help improve access and quality of mental health care, not only for peer supporters but also for professionals. It highlights the potential for AI to assist in moment-to-moment interventions and to enhance training programs to address the shortage in the mental health workforce.

What are some of the challenges in developing AI for mental health support as discussed in the transcript?

Challenges include understanding and navigating the ethical considerations, ensuring the safety and welfare of participants, and developing AI systems that can effectively and appropriately provide feedback to enhance empathy without disrupting human connections.

Outlines

00:00

🤖 Introduction to AI and Empathy in Mental Health Research

The speaker begins by expressing gratitude and excitement for the opportunity to discuss research at the intersection of AI and empathy, particularly in the context of mental health. They provide a content warning for sensitive topics such as mental illness, self-harm, and suicidal ideation, which will be discussed using anonymized examples. The research is a collaborative effort involving institutes and industry partners, aiming to improve mental health support through AI. The speaker emphasizes the importance of human-AI collaboration over replacing human roles and shares insights into the potential of AI to enhance access and quality of mental health support, focusing on peer support platforms like Talk Life.

05:01

🧠 Measuring and Enhancing Empathy in Peer Support

The speaker delves into the hypothesis that peer supporters could express higher levels of empathy with the help of automated, actionable feedback. They discuss the importance of measuring empathy using established psychological scales adapted for text-based communication. The empathy measurement framework includes emotional reactions, interpretations, and explorations, with examples provided to illustrate different levels of empathetic responses. The speaker then explores the downstream implications of empathy, such as increased positive engagement and relationship formation on peer support platforms. They highlight the need for empathy training and feedback, given that even with challenging posts, the average empathy score is low, indicating a need for improvement.

10:02

💬 Empathic Rewriting: AI's Role in Enhancing Empathy

The speaker introduces 'empathic rewriting', a task where AI takes a low-empathy response and transforms it into a higher-empathy one. They provide an example of how AI can edit a response to be more empathetic by adding emotional reactions and maintaining the original message's context. The system, named 'Partner', uses reinforcement learning to perform this task, being trained on increasing empathy while ensuring responses remain fluent and context-specific. The speaker emphasizes that this system outperforms previous natural language processing methods in empathic expression and sets the stage for demonstrating how AI can collaborate with humans to enhance empathy in peer support.

15:03

🔧 Human-AI Collaboration for Empathy Expression

The speaker presents a system that allows human peer supporters to collaborate with AI to express empathy more effectively. They contrast a control group, where supporters write responses without feedback, with a treatment group that receives AI-generated feedback. The AI provides suggestions for improving responses, which supporters can choose to accept or ignore. The speaker discusses the results of a randomized control trial involving peer supporters from Talk Life, which showed that the treatment group expressed substantially more empathy. The feedback system was well-received, with participants feeling more confident in their ability to support others empathetically. The speaker also touches on the ethical considerations and safety measures taken in developing and deploying the AI system.

20:03

🛡 Ethical Considerations and Future Directions in AI and Mental Health

The speaker concludes by emphasizing the importance of ethical considerations in AI technology for mental health. They discuss the challenges of ensuring safety and welfare, highlighting the use of an ethical framework to guide the development of the AI system. The speaker mentions the co-design process with stakeholders, the focused scope on empathy, and the 'AI in the loop' approach that respects the human connection on peer support platforms. They also discuss the use of a sandbox environment for safety testing and the importance of informed consent. Looking ahead, the speaker is optimistic about AI's potential to improve mental health care, not just in empathy but in various other aspects such as cognitive reframing and psycho-education. They invite further questions and thank the audience, signaling the end of the presentation.

Keywords

💡AI

AI, or Artificial Intelligence, refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the video, AI is central to the discussion on how it can be leveraged to improve mental health support. The speaker discusses AI's potential to enhance empathy in peer support platforms, demonstrating how AI can be programmed to recognize and generate empathetic responses.

💡Empathy

Empathy is the ability to understand and share the feelings of another. It is a critical aspect of human interaction, especially in providing support. The video emphasizes the importance of empathy in mental health support, suggesting that AI can be trained to recognize and generate responses that express higher levels of empathy, thus improving the quality of peer support on platforms like Talk life.

💡Mental Health Support

Mental health support refers to the assistance provided to individuals experiencing mental health issues. The video discusses the shortage of mental health professionals and how AI could help bridge this gap by improving the quality of peer support, which is a form of mental health support that can be accessed by millions through platforms like Talk life.

💡Peer Support

Peer support involves individuals receiving help from others who have similar experiences or challenges. In the context of the video, peer support is provided through online platforms where users can connect and share their mental health struggles. The speaker highlights the potential of AI to enhance the empathy and effectiveness of peer supporters on these platforms.

💡Anonymized Examples

Anonymized examples are real instances or data that have been modified to ensure the identities of the individuals involved are protected. The video mentions the use of anonymized examples of mental illness, self-harm, and suicidal ideation to illustrate the challenges and potential solutions in the domain of mental health support.

💡E-Science Institute

The E-Science Institute is mentioned as an organization that has contributed to the project discussed in the video. It is an example of the collaboration between different institutes and industries that have come together to leverage AI for mental health support, indicating the interdisciplinary nature of the work.

💡Clinical Psychologist

A clinical psychologist is a mental health professional who diagnoses and treats mental, emotional, and behavioral disorders. In the video, a clinical psychologist is part of the team working on the AI project, emphasizing the importance of incorporating professional psychological knowledge into the development of AI systems for mental health.

💡Reinforcement Learning

Reinforcement learning is a type of machine learning where an agent learns to make decisions by taking actions in an environment to maximize some notion of cumulative reward. The video describes the use of reinforcement learning in developing an AI system called 'partner' that can rewrite responses to express higher levels of empathy.

💡Randomized Control Trial (RCT)

A randomized control trial is a scientific experiment that involves random assignment of subjects into treatment groups to test the efficacy of a new intervention. In the video, an RCT was conducted with peer supporters from the Talk life platform to evaluate whether AI feedback could effectively improve the expression of empathy in their responses.

💡Ethics in AI

Ethics in AI pertains to the moral principles that guide the development and use of AI technologies. The video discusses the ethical considerations in using AI for mental health, such as ensuring safety, obtaining informed consent, and conducting reviews to minimize risks, highlighting the importance of ethical frameworks in AI research.

Highlights

The talk discusses the intersection of AI and empathy, focusing on mental health support.

Content warning is given for anonymized examples of mental illness, self-harm, and suicidal ideation.

The project was enabled by collaborations with the e-science Institute, Garvey Institute at UW, and other local institutes.

AI's potential to improve access and quality of mental health support is explored.

The importance of empathy in mental health support and its association with symptom improvement is highlighted.

The lack of mental health professionals and the role of peer support platforms like Talk life are discussed.

Empathy is defined as the ability to understand or feel the emotions and experiences of others.

A framework for measuring empathy in text-based communication is presented.

Empathy is found to be associated with positive engagement and relationship formation on peer support platforms.

The need for empathy training and feedback is identified, given the low average empathy scores in peer responses.

AI techniques are used to generate empathetic text, called empathic rewriting.

A system called PARTNER is introduced, which uses reinforcement learning to improve empathy in text responses.

A randomized control trial shows that AI feedback can effectively increase empathy in peer support responses.

The talk emphasizes the importance of human-AI collaboration over replacing human support.

Ethical considerations and safety measures in developing AI for mental health are discussed.

The potential of AI-based feedback to improve mental health care access and quality is summarized.

The talk concludes with a look ahead at future applications of AI in mental health, including training programs and addressing workforce shortages.

Transcripts

[Music]

foreign

this is very impressive doing this life

it's a great honor to speak uh kind of

in between Sean and Ali and nin uh have

many people that have invested so much

in the infrastructure and the

environment that we have here to do

great research for many people in this

room and I'm excited to share some of

that research

and specifically that will be about at

the intersection of AI and empathy and

I'll give a brief content warning before

we start I'll share a few anonymized

examples around mental illness self-harm

and suicidal ideation only do that to

illustrate the very real domain

challenges as well as the solutions and

of course these are properly anonymized

and nobody asked asked me to do this but

this will be very easy to tie in I think

precisely because of the environment

that we have here this is a project that

was unique that started at The e-science

Institute that was uniquely enabled by

the Garvey Institute at UW other local

institutes like ai2 and local industry

and of course everything I'm going to

show is teamwork including my PhD

students Ashi Sharma in a Lin as well as

UW

Stanford clinical psychologist Steve

Atkinson Adam minor and Folks at a peer

support platform talk live

and over the past few weeks you might

have come across one or multiple

headlines in the media there's a lot of

talk about the latest and greatest of AI

chat GPT this chat GPD that

um and uh many of these articles have

focused on the opportunities but also

the risks for AI technology in mental

health and uh in this talk I want to

demonstrate first some of that potential

specifically how AI could help improve

access and quality of mental health

support but I also want to focus on the

potential risks and about the ethics

that just came up earlier in the

question and what can be done to

mitigate those risks and I'll emphasize

they will focus here on human eye

collaboration not to replace anybody in

fact in the exact research I'm about to

show you we found that the human eye

collaboration turns out to be superior

to human only or also AI only approaches

so why do we desperately need to improve

access to mental health support

myself maybe many people in this room

are painfully aware that the need for

mental health care outweighs the access

to it one in five adults have a mental

illness most do not receive treatment

and often that is because of a pervasive

lack of mental health professionals

and this reality really emphasizes the

role of peer support and in this talk

this will be about how AI can help

improve the quality of this peer support

and make it more effective

so these peer support platforms are used

by millions and millions of people today

they're a place where you can find

people that might share your experience

that might share your suffering that

might understand what you're going

through and I'll talk about a particular

one today called Talk life you see here

a screenshot on the slide it looks a lot

like other social media platforms that

it might have seen and used but it's

specifically for mental health support

and the mental health challenges that

you see people Express there really run

the whole gamut of what you might

imagine

in principle these platforms could be

amazingly useful they could complement

traditional forms of care such as

therapy

unfortunately for many people it's

really all that they have access to

whether that's for insurance reasons

social isolation stigma in these

platforms could make interventions

available to people that otherwise

wouldn't have access to them

now the success of these platforms

relies on quality support and that

involves a range of skills but a central

enabling factor is empathy empathy has

been shown to be associated with symptom

Improvement and with kind of the

formation of positive successful

relationships that we call Alliance or

rapport

and by empathy I mean the ability to

understand or feel the emotions and

experiences of others

now let's say somebody that we call the

support Seeker they share and

unfortunately so my whole family hates

me

and then the peer supporter might

respond try talking to your friends

now that's by far not the worst message

we've any of us have seen on the

interwebs it just doesn't communicate

empathy actively it's pretty common for

peers to maybe a little quickly jump to

telling others what what actions they

should take instead you could express

more empathy for instance by saying I

could imagine that that makes you feel

really isolated so interpreting what

that person is going through and

underlying all the work I'm about to

show you is the central hypothesis that

peer supporters might Express higher

levels of empathy if they had access to

automated actionable just-in-time

feedback

and over the next few minutes I'll First

share how we can measure empathy that's

expressed uh then we'll learn about how

we can use AI techniques to see uh when

AI is expressed in the real world what

uh Downstream implications that has

we'll talk about how AI can

quote-unquote generate empathy in the

sense of generating text that expresses

empathy and then we'll put all of that

together and see how humans and I can

collaborate on on empathy expression

now the first important part about

measuring empathy is not to let computer

scientists Define it so in this work we

simply just adapted existing scales

about empathy that in Psychology have

been studied for decades and we had to

adapt them to the kind of exclusively

text-based and asynchronous

communication setting where I can't

interrupt you I can't vary my pitch and

we did that in Combined multiple scales

to capture emotional and cognitive

aspects I'll explain how we kind of with

the framework that we ended up with

um

there's three mechanisms of how empathy

can be communicated there's emotional

reactions kind of communicating the

emotions experience after reading a post

there's interpretations communicating

the understanding of the inferred

feeling some experiences of another

person and lastly there's Explorations

do you do anything to help improve your

understanding by exploring the feelings

and experiences of another person

we decided for each of these so we would

differentiate between not communicating

that at all doing that to a weak

relatively generic degree or a stronger

more specific degree and I'll share a

few examples

again somebody might share and

unfortunately so my whole family hates

me I don't see any point in living

somebody might respond let me know if

you want to talk not the worst response

doesn't actively Express empathy I

understand how you feel would be a weak

interpretation everything will be fine

weak emotional reaction what happened is

a generic exploration and you can be

more specific for instance by saying if

that happened to me I would feel really

isolated I really hope things would

improve or I wonder if this makes you

feel isolated question mark

and then return these into computational

prediction tasks two of them one given a

kind of post and a response what's the

level of empathy being expressed so

these are kind of three classification

tasks

and then a second task of what we call

rationale extraction this is about

recognizing which are the exact words

that are communicating empathy here here

are the regions highlighted in in red

and blue

I won't go into a lot of details where

we developed state-of-the-art methods

for performing these tasks in in

recognizing how much empathy is being

expressed in text

and then we work curious

um what the downstream effects of

empathy are now we tried really hard to

capture the Clinical Psychology

construct of empathy it's not a given

that on a social media kind of peer

support platform that people on this

platform would resonate with that exact

construct

but it turns out that they do and

whenever empathy is being expressed

that's associated with lots of positive

outcomes so here in the x-axis we have

how much empathy is being expressed and

we find that when more empathy is being

expressed we see people liking those

responses more we see them replying more

often so we see kind of positive

engagement coming from that we also saw

that whenever empathy was expressed

there was a uh kind of the social media

equivalent of relationship formation if

you respond to me with empathy I would

be 80 more likely to follow you right

after kind of starting a potential uh

kind of longer uh relationship on this

platform

but we also saw a lot of Need for

empathy training and feedback so if you

take our kind of three mechanisms you

add them up you have a scale from zero

to six and even among kind of posts

where people share their challenges

um uh responses scored on average only a

one out of six which isn't great turns

out it also doesn't get better over time

people don't just kind of learn to

express empathy better by by being on

this platform for longer times if

anything's the opposite and I've learned

from our psychology colleague colleague

said this is really true for therapists

as well without deliberate practice and

specific feedback also professionals can

diminish in their skills over time

so that set us on a path on

um kind of what we might do to give

people feedback and help train people in

expressing empathy

and we'll first looked at that just as a

task for machines and we called that

empathic rewriting where we want to have

a machine take a lower empathy post and

transform it to a similar response with

higher levels of empathy and here's an

example of what that could look like

let's say somebody shares I can't deal

with this part of my bipolar I really

need help somebody might respond don't

worry try to relax is there anyone you

can talk to

again not the worst response by far it

has this don't worry piece in here which

I've learned really is kind of a red

flag clearly the person is already

warring so this easily can come across

invalidating so let's say what if a

machine could help kind of adjust and

edit this post by replacing don't worry

with being manic is no fun it's actually

really scary adding an emotion reaction

I'm sorry to hear that this is troubling

you and then keeping as much as possible

of the original post as well

so that's what empathic rewriting is and

we built a system that's based on

reinforcement learning a system called

partner to do that exact task I'll share

in a high level how this works so this

system would look at a conversation

between a support seeker and a peer

supporter that informs the state of this

system the system then decides which

rewriting action wants to take so that

could be replacing a sentence adding a

new sentence

and then the system gets rewards and

that's how that system is trained and

specifically we rewarded this system for

increased empathy of course that's what

we want to do here but also keeping

fluent English having a coherent uh kind

of self-consistent response and also

being context specific to avoid generic

responses that we know people hate like

um kind of I'm so sorry to hear that

that you can use in in absolutely any

situation

um I'll uh kind of won't share more

details about the system in the favor of

showing you a little demo in a second

um but kind of this is a system that can

perform these tasks uh better than the

state of the the prior state of the art

in natural language processing

so now let's put all of that together

and see how uh kind of human peer

supporters and kind of AI systems can

collaborate on expressing empathy the

most effectively

and I'll I'll show you kind of the

system that we developed and I'll start

with kind of what the control group or

the status quo looks like so here kind

of there's a post you can respond to it

you're kind of greeted with an empty

text box even though it's kind of a

tough task

um and you can write out your response

so I'll write out my response here again

this is the control group that doesn't

have access to feedback let's say for

illustrational purposes I use our red

flag of don't worry and I'll write I'm

there for you

now here's the kind of treatment group

that will contrast with that and in that

condition peer supporters did have

access to feedback they didn't need to

use it we didn't incentivize them anyway

but they could access it and I'll kind

of show an example here I'll first write

out the exact same response

you might note the red flagging button

here at the bottom I'll come back to

that in a second

I request feedback takes about a second

and then we can display some potential

feedback displayed similarly how kind of

um kind of spell checking or or grammar

feedback in the AI system catches maybe

the Stonewall is a great idea you could

consider replacing it with it must be a

real struggle it also recognizes there's

no exploration of the other person's

feelings and experiences happening here

yet so it also suggests adding have you

tried talking to your boss

um

so I'll blow this up kind of now what

could a peer supporter do they could

kind of click on replace and insert they

could maybe they don't like the feedback

or they like it so much they want more

so you can reload the feedback as well

you can ignore the feedback you can edit

before and after and so on

and then we performed a randomized

control trial to figure out whether this

is actually any helpful and we worked

with actual peer supporters from this

talk life platform uh 300 of them in a

fully remote RCT we randomly divided

into these two conditions kind of the

status quo you write that on your own

and you don't get any feedback

or the treatment group that had access

to feedback

of course the primary outcome is the

level of empathy being expressed and one

thing that's really important here is we

actually set up ourselves up in a really

conservative way we didn't want to

measure oh it's kind of the use of the

that AI system kind of any useful but

what we did is we trained everybody just

before the study on kind of traditional

means of empathy so we would show people

a a screen with definitions of empathy

the mechanisms of empathy different

examples but kind of a static way how

you would do that today so if on the

next slide we see any differences

between treatment and control group what

that would mean is that this kind of

just-in-time very context specific

feedback is actually helpful Beyond this

generic training even if we literally

trained you I know three minutes ago

um of course I wouldn't set it up if the

result wasn't positive but whether you

um

um ask human peer supporters on the left

or machines on the right hand side both

agree that in the treatment condition

substantially more empathy is being

expressed about 20 more it was about

double that for people that actually

find it challenging to support others

empathically

the participants uh kind of 77 percent

wanted that system deployed and it's

still far from perfect UH 60 percent

found the feedback actionable and

helpful and one thing that was really

meaningful to us was that 69 said I feel

more confident now after this feedback

and supporting others and kind of

empowering peer supporters with these

systems is of course

um kind of what we're after here and it

suggests that there could be kind of

training effects of of using these

systems in the future as well

now I mentioned we didn't incentivize

people at all to use this feedback kind

of that that's likely how systems like

that would be used in in the future so

they have to be kind of good enough for

people to actually want to use them

um and people use them in a variety of

different ways some people chose never

to use them uh some people used it a lot

and here on the x-axis there's kind of

different clusters of how people use the

system from not using it ever even

though they could on the left to using a

kind of basically all the time on the

right one thing that's really

interesting is that the more people use

the system the more empathy they

expressed as well and these are quite

massive uh effect sizes

now of course there's selection effects

here people on the right might be

particularly motivated and interested in

doing that people on the left might have

been unfortunate and just not gotten the

best feedback uh that we could generate

but it was very interesting to see kind

of the Striking dose response type

relationship

um nevertheless

so over the past few minutes I hope that

I was able to convince you at least of

some of the potential and the potential

benefit on opportunity for AI technology

in mental health but of course AI

technology and mental health creates

lots of risk and we had a question about

that earlier as well and of course in

whatever we do here the safety and

Welfare participants has to come first

any potential needs need to be

considered and navigated

the challenge is that in kind of a very

new and emerging field they are it can

be challenging to understand like how

exactly do you do that it's a lot easier

for me to say here that I care about all

these things it's a lot more challenging

to figure out how exactly this needs to

happen and I'll give a shout out here to

a great ethical framework by Camille

nebeker and others at recode health and

that helped us kind of inform the

concrete actions that we took here that

I I want to share next

so first I believe it was critical to

the success here to co-design the system

with the various stakeholders to

understand the risks to minimize the

risks to create any benefit so we worked

over many years with the designers of

this platform with the peer supporters

and the users of this platforms and with

clinicians uh personally I'm a computer

scientist

you might have noticed that we're

intervening on the peer supporter we're

trying to help the helpers we're not

intervening on the person acutely in

crisis that's a deliberate choice for

how to do that with less risk

you might also have noticed that there's

a very focused scope here on empathy and

we understand empathy and have studied

empathy for decades this isn't a general

chat bot where we don't know what will

happen

um but it's very specifically designed

for empathy feedback

we also kind of from the beginning of

this project understood that there's

kind of this really meaningful human

human connection happening on this

platform and there's absolutely a risk

that if you kind of introduce an AI

system in the middle that you would kind

of disrupt that meaningful connection

and that's of course not what we want to

do here

in my world of computer science and

machine learning sometimes we call these

things human in the loop I think in the

course of this project we preferred

calling this AI in the loop on a back

seat with human supervision and felt

more comfortable with that perspective

and kind of part of what the system is

doing here that it was designed to give

minimal feedback and only if needed to

only pop up if there's kind of an

opportunity for empathy here and not to

do anything if that's for instance

already expressed

we also made a deliberate choice of not

to kind of answer fewer research

questions with a lot more work and

specifically what we did is that we

designed a kind of realistic looking

sandbox environment instead of actually

showing the responses yet to somebody on

the stock life platform we felt the

right way to kind of progress in this

line of work was to study the safety

implications first and for instance we

did that through that flagging button

that I showed you earlier

um and kind of with the kind of flagging

we're also using the same idea of kind

of human air collaboration where there's

lots of algorithms behind the scenes

that filter

potentially inappropriate or unsafe

content but it also allows the peer

supporter full control it allows them to

flag anything that might not be caught

by these algorithms

and in in doing that also improve these

systems over time

now participants had access to a crisis

hotline could quit at any point we did

safety reviews of these systems at every

stage over multiple years some of you

might find it interesting that in the

kind of exact AI objective functions you

can encode some of these ethical

considerations for instance to uh kind

of make the minimal changes needed in

giving the feedback rather than kind of

letting you write out a response and

then I cross it all out which seems like

an amazingly uh effective way

invalidating people and making them less

confident that they're good at

supporting others and of course there

was also informed consent and IRB

approval I think a lot of that is kind

of what we needed to learn over the last

couple years and it's pretty clear

similar to Ali's response earlier that a

lot more work is needed in in this area

on the kind of ethical principles in

digital health and also the algorithmic

approaches to ensure safety

I'll summarize and briefly look ahead

and then I very much look forward to the

Q a

I told you about empathy that empathy is

crucial but a little bit more rare than

we would like in the work I showed here

introduces new computational tasks and

data sets and tools that can be used for

facilitating empathic conversations

based on state-of-the-art natural

language processing techniques

we saw an RCT that shows that this human

eye collaboration approach can be

effective

in this and other tools we've developed

are actually already in use by several

mental health organizations

serving millions of people

so looking ahead I believe that AI based

feedback could really help improve

access and quality of mental health care

whether that's for peers or

professionals

I think it could help improve within the

moment interventions like I showed you

but also help improve training programs

to help us address the massive shortage

in the mental health Workforce here

um and of course we would focus on many

more things and kind of things Beyond

empathy like maybe cognitive reframing

complex Reflections building effect and

patient provider relationships

psycho-education challenging Health

misinformation and so on

if anybody's interested here at this URL

you can find all the papers models data

sets associated with this work uh big

shout outs to the amazing team that made

this possible again and I look forward

to your questions thank you

[Music]

thank you