I agree with you that AI companions won’t improve social skills or connection by default. But I disagree that the commercial incentives to build one that will are doomed. I think it really wouldn’t be hard to create an AI companion that would actually be helpful in most of the ways you mention. This would be a matter of even moderate scaffolded prompting for coach-like behavior.
LLMs seem quite capable of mixing emotional support and coaching. In fact I’m certain they can easily do this, at least if you keep asking them to at regular intervals; I use them this way myself (my use tilts far toward coaching and consulting and away from “companion”; but they naturally provide emotional support when I’m asking them about difficult interpersonal issues).
This would be marketed as “the healthy companion”, and I think it would be just as easy to build one that actually works as one that doesn’t and just claims it does. I’d think that the market forces would favor better efforts, since it’s easy to get reviews. It wouldn’t be perfect, but that doesn’t prevent competition in similar markets from pushing in the right direction.
Such a system would proactively, occasionally, push the user to consider how to develop healthy social connections with other humans, as well as mitigating the loneliness that’s their primary concern. It would ask about all of the nonverbal skills you mention, and give sources to check those things or have conversations about them. It would suggest routes toward developing healthy relationships; groups, hobbies, dating, and initiate conversations about how those things are going and how to make them work, etc.
I don’t know if this is a wild hope or something we should expect to see by default. Imagine a parent whose child is using AI as a companion seeing an ad “for the better, healthier AI companion that helps you develop human connections as well as supports you emotionally”. They’d pay that subscription in a second if they could possibly afford it. And imagine friends alarmed by someone’s AI use suggesting they switch subscriptions to the “better” option (and here motivated reasoning and the halo effect help it develop the reputation of being better in other ways, too).
I think the combination of outside pressure (you should use the healthy one!) and user experience (this one is both supportive and makes me feel like I’m trying to get healthier and being a responsible person) should work in combination to push in the direction of both quality and helpfulness. And it’s not hard to gather user reviews AND friend/parent reviews for something people will devote that much time and emotional energy to.
One might worry that the incentive is to keep the user dependent on the coachpanion. I don’t think it needs to push anyone out the door. They’ll grow fond of the coachpanion because it’s helpful and pleasant, and it’s the same subscription cost to use it a little as to use it a lot. And come to think of it, less use but maintained subscription financially benefits the provider (a little) by saving on compute.
Maybe this is naive, but I don’t think so. If I weren’t so concerned about alignment and everyone dying, I’d be starting such a company myself. I want to see others do it, and I’d like to help if they do.
I’ve never gotten around to writing about this publicly before, so thanks for the impetus!
Edit: The fact that no such thing exists is a real challenge to my argument. Is that a clear indicator, or is it just that nobody has really tried to make such a thing yet? Markets are real, but they’re typically not very efficient.
I agree with you that AI companions won’t improve social skills or connection by default. But I disagree that the commercial incentives to build one that will are doomed. I think it really wouldn’t be hard to create an AI companion that would actually be helpful in most of the ways you mention. This would be a matter of even moderate scaffolded prompting for coach-like behavior.
LLMs seem quite capable of mixing emotional support and coaching. In fact I’m certain they can easily do this, at least if you keep asking them to at regular intervals; I use them this way myself (my use tilts far toward coaching and consulting and away from “companion”; but they naturally provide emotional support when I’m asking them about difficult interpersonal issues).
This would be marketed as “the healthy companion”, and I think it would be just as easy to build one that actually works as one that doesn’t and just claims it does. I’d think that the market forces would favor better efforts, since it’s easy to get reviews. It wouldn’t be perfect, but that doesn’t prevent competition in similar markets from pushing in the right direction.
Such a system would proactively, occasionally, push the user to consider how to develop healthy social connections with other humans, as well as mitigating the loneliness that’s their primary concern. It would ask about all of the nonverbal skills you mention, and give sources to check those things or have conversations about them. It would suggest routes toward developing healthy relationships; groups, hobbies, dating, and initiate conversations about how those things are going and how to make them work, etc.
I don’t know if this is a wild hope or something we should expect to see by default. Imagine a parent whose child is using AI as a companion seeing an ad “for the better, healthier AI companion that helps you develop human connections as well as supports you emotionally”. They’d pay that subscription in a second if they could possibly afford it. And imagine friends alarmed by someone’s AI use suggesting they switch subscriptions to the “better” option (and here motivated reasoning and the halo effect help it develop the reputation of being better in other ways, too).
I think the combination of outside pressure (you should use the healthy one!) and user experience (this one is both supportive and makes me feel like I’m trying to get healthier and being a responsible person) should work in combination to push in the direction of both quality and helpfulness. And it’s not hard to gather user reviews AND friend/parent reviews for something people will devote that much time and emotional energy to.
One might worry that the incentive is to keep the user dependent on the coachpanion. I don’t think it needs to push anyone out the door. They’ll grow fond of the coachpanion because it’s helpful and pleasant, and it’s the same subscription cost to use it a little as to use it a lot. And come to think of it, less use but maintained subscription financially benefits the provider (a little) by saving on compute.
Maybe this is naive, but I don’t think so. If I weren’t so concerned about alignment and everyone dying, I’d be starting such a company myself. I want to see others do it, and I’d like to help if they do.
I’ve never gotten around to writing about this publicly before, so thanks for the impetus!
Edit: The fact that no such thing exists is a real challenge to my argument. Is that a clear indicator, or is it just that nobody has really tried to make such a thing yet? Markets are real, but they’re typically not very efficient.