I've developed a GPT-3 enabled robot therapist that aims to provide empathetic and informed conversations for individuals seeking mental health support. The project leverages the power of GPT-3 to offer a non-judgmental space, evidence-based coping strategies, and encouragement to seek professional help when necessary.
While I think there's definitely room for AI therapy as it can give people better results than many humans, if you don't have safety infrastructure in place (both psychological and patient data) you're walking off a liability cliff. Nobody with a clue would trust a startup with confidential mental health info, which means your customer pool is going to be made up of unstable people.
Tell us about yourself
First Name
How was your childhood?
What is your relationship status?
What is your MBTI type?
Where did you grow up?
Where do you live?
What is your criminal history?
How often do you have drugs or alcohol?
How is your family life?
What is your religion?
What is your highest level of education?
What medication are you taking?
Are you working?
This screams 'SCAM' to any normal person. The GoDaddy anonymous domain registration makes that even more likely. If this is a sincere effort to launch a business, start over and do your homework.
As far as medical therapy goes, sure. Otherwise I'm sure the OpenAI devs have in place proper filters for questions pertaining to sensitive personal data, or sensitive health related questions. There is nothing wrong in providing general "life advice" as it were, and then add a caveat such as "if your condition is serious, you should seek professional medical advice." Further more, I'm sure they have the extra cash on hand to hire medical professionals who can better formulate such caveats based on flagged inputs.
Personally I'm not a big fan of the EU, but they got one thing right, and that is the General Data Protection Regulation (GDPR).
Personal faith and religion, criminal history, and health related questions such as medication, drug and alcohol use, are all considered as highly sensitive personal data, which is protected under the regulation. Collecting and processing such data is subject to strict requirements and conditions under the GDPR, and requires a legal basis for processing sensitive personal data.
The other questions on the list may not necessarily conflict with GDPR on their own, though that depends on how they are collected, processed, and used. Specifically, since the questions are so broad, most if not all of those questions can be used to fish for secondary information, which can then be used for either illegal marketing, down right scams, or even provide third parties with means of discrimination, which are of course all completely illegal. As such, even otherwise "innocent" questions may also be regulated by the GDRP depending on how you handle them.
As a minimum, sensitive data should be securely encrypted in compliance with the regulation, and then stored only for the minimum required time for processing it. Excess personal data should not be stored at all, and even asking some questions — such as asking whether you're religious — might get you into a heap of trouble depending on how you handle or store that information, if at all.
I think you probably meant well, but as it looks now, I think you should look into the GDPR regulation and carefully consider how your implementation stacks up against it, and also carefully consider how you can build trust in the way you set up your site.
As of now, I wouldn't type anything into your site. But that's me. There might be people who do type stuff into your system, and then is made aware that it's something they should never have done. And that's where most serious legal problems may arise. It is thus in your best interest to carefully consider how such regulations might affects your site, and even if it might be considered to be outside of your jurisdiction.
There are also implications to consider when you send data to a third party's API that probably wasn't meant to handle such data, and that you yourself have no control over outside of the data you get back. You might thus be handing over sensitive data to a third party, where either you, or they, might be considered to handle or store it illegally. To me this represents a rabbit hole of legal problems, of which there are a ton of footguns for us small-time devs. So just make sure you thread carefully.
After the Better Help fiasco, I’d imagine people would be very hesitant to use something like this without some kind of third-party auditing and certification that their data aren’t being used to sell them stuff.
I would never attempt it. AI is way to nascent to be applied like this when it is convincing people to commit suicide to save the planet, like last week.
Ha. Can you imagine what that study would look like? And what an ethics committee would have to say about it?
Naw. Tech people aren’t about “research”. Throw something together, and when your uninformed gut feeling tells you that it’s good enough, you snap a nice marketing site together and throw it over the fence.
First it’s sketch as fuck, second GPT is a guestimation engine and people are getting killed for asking it for psychological help. Someone in Belgium commit suicide to save the planet last month.
I love the drive and the idea, as someone who could benefit from this, but I have to shut this down hard.
As someone with significant mental health issues, and a moderate amount of experience with AI, who works with it daily, don’t mix them, please.
NEVER use GPT for financial, health, legal or other advice, as it’s not meant for that. Yes, it can simplify the research process, but always verify.
If you wanted to use a Language Model like this safely, you could load medical language models and ask questions about definitions, but never advice.
If you’re interested, I’d love to show you the basics, and I hate to shoot down other peoples good work, but not in this case.
Author: you’re probably not meaning any harm, but I have no way of knowing that.
You need to do far more research, far more legal considerations, and far more footwork to make this work.
Don’t be discouraged, but don’t get people killed with carelessness.
I'm interested in how I can use this moving forward because I have personally found my app to be helpful for getting over a breakup.
My prompt is designed for Mindy to act as the "best of both worlds" between a real therapist and robot therapist, which means it will refer you to real therapy if it deems your problem too serious for a robot.
I'm not sure anymore however. I don't store any data for longer than the session is active, but I know OpenAI does. Also: OpenAI knows I am using their API for this purpose.
I'm really at a loss. I see the value and the harm. I'm going to have to do more thinking on this.
Cool! Down to work on this. Let me know where to reach out. Also I've contacted my lawyer about this and it's a very grey area. We'll see where that goes.
I've also used ChatGPT for "self therapy" and found out that it works reasonably well, though I would not recommend it for serious medical or psychiatric issues. I always take the feedback with a grain of salt, however, and usually when you ask such questions you are given the caveat to seek professional help.
If anything, I have positive faith in that ChatGPT or GPTn will be a great tool for therapy in the future.
As for the guy who killed himself, as sad as it is, I seriously doubt that ChatGPT made anything worse. With that said, the incident of course needs closer investigation to find out what advice the guy actually got.
A chat AI would have specific and professionally prepared training to handle people in distress, that's why I think it is in fact better that they talk to the AI rather than some untrained friend or relative.
This is of course just one aspect in how the above application is problematic, however.
How do you prevent GPT from hallucinating and giving credulous but manipulative or misleading information? I could see some small percentage of your users being seriously harmed by the AI believing what they say or by gaslighting them. I’ve seen examples of GPT gaslighting people outside of this context. Do you have any deterministic way of proving this won’t happen or are you entirely at the mercy of the guardrails in the black box?
If your "GPT-3 Enabled Robot Therapist" is not licensed anywhere to perform therapy (as I think it isn't), you could get yourself into much trouble, and very fast.
31 comments
[ 0.13 ms ] story [ 78.8 ms ] threadStop, re-evaluate, and understand there is significant potential, but depending on where you are, there could be prison time, man.
Shut this down now before you get too much attention.
Personal faith and religion, criminal history, and health related questions such as medication, drug and alcohol use, are all considered as highly sensitive personal data, which is protected under the regulation. Collecting and processing such data is subject to strict requirements and conditions under the GDPR, and requires a legal basis for processing sensitive personal data.
The other questions on the list may not necessarily conflict with GDPR on their own, though that depends on how they are collected, processed, and used. Specifically, since the questions are so broad, most if not all of those questions can be used to fish for secondary information, which can then be used for either illegal marketing, down right scams, or even provide third parties with means of discrimination, which are of course all completely illegal. As such, even otherwise "innocent" questions may also be regulated by the GDRP depending on how you handle them.
As a minimum, sensitive data should be securely encrypted in compliance with the regulation, and then stored only for the minimum required time for processing it. Excess personal data should not be stored at all, and even asking some questions — such as asking whether you're religious — might get you into a heap of trouble depending on how you handle or store that information, if at all.
I think you probably meant well, but as it looks now, I think you should look into the GDPR regulation and carefully consider how your implementation stacks up against it, and also carefully consider how you can build trust in the way you set up your site.
As of now, I wouldn't type anything into your site. But that's me. There might be people who do type stuff into your system, and then is made aware that it's something they should never have done. And that's where most serious legal problems may arise. It is thus in your best interest to carefully consider how such regulations might affects your site, and even if it might be considered to be outside of your jurisdiction.
There are also implications to consider when you send data to a third party's API that probably wasn't meant to handle such data, and that you yourself have no control over outside of the data you get back. You might thus be handing over sensitive data to a third party, where either you, or they, might be considered to handle or store it illegally. To me this represents a rabbit hole of legal problems, of which there are a ton of footguns for us small-time devs. So just make sure you thread carefully.
https://www.theverge.com/2023/3/2/23622227/betterhelp-custom...
>"What medication are you taking?"
Um.. Is this HIPAA compliant?
Surly you should also be asking for my star sign if you're operating at that level?
Patient: "What is your (creator's) degree?"
Naw. Tech people aren’t about “research”. Throw something together, and when your uninformed gut feeling tells you that it’s good enough, you snap a nice marketing site together and throw it over the fence.
First it’s sketch as fuck, second GPT is a guestimation engine and people are getting killed for asking it for psychological help. Someone in Belgium commit suicide to save the planet last month.
I love the drive and the idea, as someone who could benefit from this, but I have to shut this down hard.
As someone with significant mental health issues, and a moderate amount of experience with AI, who works with it daily, don’t mix them, please.
NEVER use GPT for financial, health, legal or other advice, as it’s not meant for that. Yes, it can simplify the research process, but always verify.
If you wanted to use a Language Model like this safely, you could load medical language models and ask questions about definitions, but never advice.
If you’re interested, I’d love to show you the basics, and I hate to shoot down other peoples good work, but not in this case.
Author: you’re probably not meaning any harm, but I have no way of knowing that.
You need to do far more research, far more legal considerations, and far more footwork to make this work.
Don’t be discouraged, but don’t get people killed with carelessness.
I'm interested in how I can use this moving forward because I have personally found my app to be helpful for getting over a breakup.
My prompt is designed for Mindy to act as the "best of both worlds" between a real therapist and robot therapist, which means it will refer you to real therapy if it deems your problem too serious for a robot.
I'm not sure anymore however. I don't store any data for longer than the session is active, but I know OpenAI does. Also: OpenAI knows I am using their API for this purpose.
I'm really at a loss. I see the value and the harm. I'm going to have to do more thinking on this.
But I see your love, and I want to help!!
I have experience in PII and HIPPA and can see if we can address some of this stuff.
I think a conclusive terms of service, less specific questions, and storing it entirely in the browser if possible, will be a solid start .
My passion is from concern, not anger. I hope that comes across
If anything, I have positive faith in that ChatGPT or GPTn will be a great tool for therapy in the future.
As for the guy who killed himself, as sad as it is, I seriously doubt that ChatGPT made anything worse. With that said, the incident of course needs closer investigation to find out what advice the guy actually got.
A chat AI would have specific and professionally prepared training to handle people in distress, that's why I think it is in fact better that they talk to the AI rather than some untrained friend or relative.
This is of course just one aspect in how the above application is problematic, however.