Questions about the Social Impacts of AI

First, I found the backlash against OpenAI’s decision to sunset GPT-4o with the release of GPT-5 surprising. How big was this backlash? Is a vocal but small subset of users distorting my perception here? Or, are a substantial share of people developing emotional attachments to LLMs? While the reaction was enough to warrant reinstatement of 4o, I suspect that the reversal was not terribly expensive for OpenAI.

Second, I want to separate out these possibly consequential but quotidian changes to our lives from personally catastrophic outcomes. For the latter, I am thinking primarily of “AI psychosis” and AI-related suicide. This recent piece on a suicidal teen’s concerning interactions with ChatGPT is heartbreaking. I am grateful that I didn’t have access to ChatGPT during my depressive episodes in high school. I am worried what I may have asked of it. I think I would retrospectively be grateful if it refused those queries. This technology could be quite powerful, and we should ask much of it. Even if you think LLMs are not causing these cases of psychosis or suicide, I think we should ask that LLMs help us reduce the incidence of these tragedies.

LLMs strike me as especially amenable to moderation (through refusing responses) and monitoring for dangerous behavior. I will admit that I am uncertain about how to handle reporting when a user might present a risk of harming themselves or others. I’d like to develop these thoughts in a later post. So, I’ll focus this post on the first topic: how might AI reshape our relationships to others and ourselves.

Why might LLMs have large social impacts?

My interactions with LLMs feel very different than my interactions with humans. First, these interactions are highly personalized. I lead the conversation, and the conversation is centered around my needs, and my needs only. Interpersonal relationships are far more reciprocal. Reciprocal relationships are fulfilling, but they are not always easy.

Second, LLMs are always available. An instantaneous response at any time of day is hard to beat. I do take joy in making or receiving the spontaneous phone call or text message. Nevertheless, convenience often wins out.

Third, LLMs are largely sycophantic. This feature does not seem necessary, but I suspect users will in aggregate prefer a fairly sycophantic model. I believe that GPT-5 is less sycophantic, but it still seems sycophantic to me. I am also not sure if the model is less sycophantic in a deep or shallow way. I often say or think things that are stupid. I appreciate and respect my friends and colleagues who call me out on this, and I learn when they do so. However, I again think that sycophancy is the easier path. Adulation feels good!

I’d group these three features under a broader category of friction. Human relationships are full of friction, which makes them challenging. Overcoming those challenges deepens the sense of connection we have to others. LLMs give us a frictionless simulacrum of human interaction. I won’t deny the allure. When we fall on hard times, who among us doesn’t long for this kind of unconditional, always-on support? Given this lack of friction, I think it’s possible more and more of us will find LLMs deeply integrated into our personal lives.

I’ll now formalize some of my biggest questions about this topic. If you are or know someone working on these topics, please let me know. If you are interested in working on these topics, please get in touch. If you’ve seen good work on any of these questions, please send it my way! I’d like to write and think about these questions more, and I’m exploring how we might study them at FRI (my employer).

• What proportion of people have a friendship-like attachment to an LLM?

• How reciprocal are these friendship-like attachments?

• How is the nature of friendship changing for young people? Do young people have more or fewer friends? Are these friendships deep or shallow?

• If AI-produced content can be highly personalized, do young people have fewer shared cultural artifacts? In high school, my friends and I watched the same YouTube videos.

• What proportion of people have a romantic attachment to an LLM?

• How does this proportion vary by age?

• Is new relationship formation among young people falling?

• Is the duration of relationships among young people changing?

• How prevalent is “AI psychosis?”

• Are cases of AI psychosis displacing episodes of psychosis that would have other triggers? Or, is there a meaningful increase in psychosis due to LLMs?

• I have the same two questions for AI-enabled or -encouraged suicide.

• Additionally, could LLMs reduce the prevalence of psychosis, suicide, or other similar mental health issues?

I’m curious to learn more. Please get in touch if you are too.

The thumbnail photo is from Susan Wilkinson on Unsplash.