The potential for AI to reshape mental health globally is really underexplored. It’s great to see EAs paying attention to it! Here are some hastily scribbled thoughts:
1. Confidently drawing conclusions about the subset of people who are currently using LLM therapy from studies of people who were paid to try it out. I’m doubtful those two client groups would be sufficiently comparable. Given that variance in clients accounts for ~5x more variance in treatment outcome than the treatment itself, it’s really important to keep the client group constant.
2. Consider the challenges of distribution and funding. We already have many evidence-based, cost-effective therapies that aren’t widely implemented—not because they don’t work, but because it’s hard to secure funding and uptake at scale. One potential shortcut here is working directly with the model providers: they already have distribution, resourcing, and reputational incentives to avoid harm. Helping them mitigate specific risks may be a more tractable way to have immediate impact, while also building credibility for future projects—before vs afters are a really compelling way to demonstrate your impact that anyone could understand in seconds. The other proposed work would be far harder to communicate, and thus harder to get people excited about.
3. Take the unguided digital intervention literature with a grain of salt. There’s been decades of promising RCTs that unfortunately didn’t translate into widespread real-world impact. A few possible reasons:
Low-cost trials make it cheaper to try again until you get a strong result (by chance)
Many of these studies are sponsored by companies/authors with strong incentives to report positive findings.
Participants are often paid to engage—something that doesn’t generalize well to real-world settings where adherence is a big challenge, especially for unguided interventions.
4. Doing too many things Each of your four projects could easily be an entire organization’s focus. Consider testing each idea briefly, then doubling down on the one where you see the most traction. . If you’d rather work separately on different things, you might want to consider branding yourselves as separate projects—it’s harder for any given project to be taken as seriously otherwise.
Hiii! Thanks! I’m only speaking for myself here, and I’m mostly interested in #3, or specifically in building, testing, and rolling out an AI-based tool for this rather than an RCT.
2. Consider the challenges of distribution and funding.
Yeah, working directly with the likes of Google (Gemini) and others would be swag, but correct me if I’m wrong, I see a very low chance of that working out? There is little commercial incentive in it for them, it doesn’t help them gain more market share from their competitors because our target clients can’t pay much, reputational risks similar to self-driving cars, etc. I haven’t asked anyone who works there, but I’m not sufficiently optimistic that it could work out to attempt it… Besides, if it does work out and lots of people start using Gemini for therapy, and then Google redecides and closes that department again, lots of users will use new version of a product for a purpose for which it’s not tested or optimized anymore.
But I already built an alpha version of an app for mentalization-based treatment on top of Gemini. That’s super easy, and I’ll permanently have control over the instructions and possibly the fine-tuning. If it should turn out to be too risky, I can shut it down, or more likely I can make adjustments to minimize any new risks.
Do you think I overestimate the difficulty of working with the model providers?
4. Doing too many things
The topics can probably be trimmed down a bit, but I feel like #1–3 form a nice story line where we first assess the risks, the assess the opportunities, and then exploit them? Personally, I’d rather 80⁄20 all of that by rolling out my solution only to fairly stable people first (I’m in some relevant support groups), collect feedback, poll well-being measures from time to time, and react to any problems with safety (in the feedback) or lacking effectiveness (well-being measures) along the way, while I increasingly market it to wider audiences. The others might want to take this more slowly, and as a result they’ll probably have the better data, but when that data is in, I can still optimize my tool accordingly.
Do you think it would really be better to focus on one topic only or would you agree that merging and 80/20ing is the better approach?
The potential for AI to reshape mental health globally is really underexplored. It’s great to see EAs paying attention to it! Here are some hastily scribbled thoughts:
1. Confidently drawing conclusions about the subset of people who are currently using LLM therapy from studies of people who were paid to try it out.
I’m doubtful those two client groups would be sufficiently comparable. Given that variance in clients accounts for ~5x more variance in treatment outcome than the treatment itself, it’s really important to keep the client group constant.
2. Consider the challenges of distribution and funding.
We already have many evidence-based, cost-effective therapies that aren’t widely implemented—not because they don’t work, but because it’s hard to secure funding and uptake at scale. One potential shortcut here is working directly with the model providers: they already have distribution, resourcing, and reputational incentives to avoid harm. Helping them mitigate specific risks may be a more tractable way to have immediate impact, while also building credibility for future projects—before vs afters are a really compelling way to demonstrate your impact that anyone could understand in seconds. The other proposed work would be far harder to communicate, and thus harder to get people excited about.
3. Take the unguided digital intervention literature with a grain of salt.
There’s been decades of promising RCTs that unfortunately didn’t translate into widespread real-world impact. A few possible reasons:
Low-cost trials make it cheaper to try again until you get a strong result (by chance)
Many of these studies are sponsored by companies/authors with strong incentives to report positive findings.
Participants are often paid to engage—something that doesn’t generalize well to real-world settings where adherence is a big challenge, especially for unguided interventions.
4. Doing too many things
Each of your four projects could easily be an entire organization’s focus. Consider testing each idea briefly, then doubling down on the one where you see the most traction. . If you’d rather work separately on different things, you might want to consider branding yourselves as separate projects—it’s harder for any given project to be taken as seriously otherwise.
Hiii! Thanks! I’m only speaking for myself here, and I’m mostly interested in #3, or specifically in building, testing, and rolling out an AI-based tool for this rather than an RCT.
Yeah, working directly with the likes of Google (Gemini) and others would be swag, but correct me if I’m wrong, I see a very low chance of that working out? There is little commercial incentive in it for them, it doesn’t help them gain more market share from their competitors because our target clients can’t pay much, reputational risks similar to self-driving cars, etc. I haven’t asked anyone who works there, but I’m not sufficiently optimistic that it could work out to attempt it… Besides, if it does work out and lots of people start using Gemini for therapy, and then Google redecides and closes that department again, lots of users will use new version of a product for a purpose for which it’s not tested or optimized anymore.
But I already built an alpha version of an app for mentalization-based treatment on top of Gemini. That’s super easy, and I’ll permanently have control over the instructions and possibly the fine-tuning. If it should turn out to be too risky, I can shut it down, or more likely I can make adjustments to minimize any new risks.
Do you think I overestimate the difficulty of working with the model providers?
The topics can probably be trimmed down a bit, but I feel like #1–3 form a nice story line where we first assess the risks, the assess the opportunities, and then exploit them? Personally, I’d rather 80⁄20 all of that by rolling out my solution only to fairly stable people first (I’m in some relevant support groups), collect feedback, poll well-being measures from time to time, and react to any problems with safety (in the feedback) or lacking effectiveness (well-being measures) along the way, while I increasingly market it to wider audiences. The others might want to take this more slowly, and as a result they’ll probably have the better data, but when that data is in, I can still optimize my tool accordingly.
Do you think it would really be better to focus on one topic only or would you agree that merging and 80/20ing is the better approach?