OpenAI's Mental Health Push Feels Like Déjà Vu From the Social Media Era

In theory.

The cynic in me (and the cynic is pretty loud these days) notes that $2 million is approximately what OpenAI spends on compute in a slow afternoon. It's a rounding error. It's less than Sam Altman's annual dry cleaning bill, probably. The grants are structured to support "projects that study real-world risks, benefits, and applications," which is vague enough to fund almost anything and specific enough to sound rigorous. I've seen this movie before. Microsoft did it with AI ethics. Google did it with algorithmic fairness. The research gets published, the company cites it in congressional testimony, and the underlying business model remains untouched.

But what do I know. Maybe this time is different.

The child safety work is harder to be cynical about, honestly. OpenAI published a detailed breakdown of how they're combating child sexual exploitation material, and the technical measures are serious. They're using hash-matching to detect known CSAM, they're building classifiers to catch AI-generated variants, they're reporting to NCMEC like they're legally required to do. The company claims to have blocked and reported attempts to use their models for this stuff, though they don't disclose exact figures. Standard practice is to keep those numbers quiet so you don't give bad actors a roadmap, which is reasonable, but it also means we're taking their word for it.

The mental health response improvements are where things get interesting, and also where the uncertainty is thickest. OpenAI says ChatGPT can now recognize distress signals better, respond with more empathy, and guide users toward real-world resources like crisis hotlines. The 170 experts they consulted include therapists, researchers, and people with lived experience of mental health crises. The 80 percent reduction in unsafe responses sounds impressive until you ask: unsafe according to whom? Measured how? Over what time period? The blog post doesn't say, and I couldn't find supplementary data. It's too early to say whether these improvements hold up outside controlled testing.

What I find genuinely novel, and this is where the AI mental health story diverges from the social media playbook, is the nature of the relationship users form with these systems. Facebook and Instagram were parasocial at a distance. You scrolled, you compared yourself to influencers, you felt bad. ChatGPT is parasocial up close. People talk to it like a friend, a therapist, a confidant. The company knows this! Their own research acknowledges that users form emotional attachments to AI systems, and that these attachments can be beneficial or harmful depending on circumstances they don't fully understand yet.

This is the part that keeps me up at night, not as a tech reporter but as someone who's watched three decades of this industry promise to move fast and fix things later. The mental health applications of large language models are genuinely unprecedented. We don't have good analogies. It's not like social media, not exactly. It's not like video games, not exactly. It's not like the parasocial relationships people form with TV characters, not exactly. It's something new, and the honest answer is that nobody knows what the long-term effects will be.

OpenAI's grant program at least acknowledges this uncertainty, which is more than most companies do. They're explicitly funding research into "real-world risks" and asking academics to tell them things they might not want to hear. But the structure of the funding, the fact that OpenAI controls the purse strings and can choose which proposals to support, means the most critical research might never get funded. The company says they want independent analysis. We'll see.

The parental controls are probably the most immediately useful feature for most families. Parents can now set restrictions on what their kids can discuss with ChatGPT, get notifications about concerning conversations, designate trusted contacts. It's the kind of basic guardrail that should have existed from day one, but better late than never I suppose. The implementation details remain unclear, we don't know how aggressive the monitoring is, whether it catches subtle distress signals or only obvious ones, how many false positives parents should expect. These are the kinds of questions that only get answered after millions of families use the system for months.

I keep coming back to the timing. OpenAI is facing lawsuits. They're facing regulatory scrutiny in Europe and increasingly in the US. The mental health angle is particularly sensitive because the plaintiffs in some of these cases are minors, or the families of minors, alleging real harm. And here comes OpenAI with a comprehensive safety package, expert partnerships, research funding, the works. It looks like a company getting ahead of a crisis. It looks like a company building a legal defense. It looks like, well, it looks like every tech company I've covered for the past thirty years when the heat gets turned up.

Maybe that's unfair. Maybe OpenAI would have done all this anyway, lawsuit or no lawsuit. Maybe the young founders running these companies have learned from the mistakes of their predecessors and genuinely want to build responsible AI systems. I'd like to believe that. Some days I do believe it.

But I've been around long enough to know that good intentions and good outcomes are different things. The social media companies had mental health initiatives too. They funded research, hired experts, rolled out safety features. And here we are, a decade later, with teen depression at record highs and congressional hearings about algorithmic harm. The features helped at the margins. The business model won.

OpenAI's business model is to make AI systems that people want to talk to, want to rely on, want to form relationships with. The better ChatGPT gets at being a supportive presence, the more people will use it, the more data OpenAI collects, the more valuable the company becomes. There's nothing inherently wrong with that! But it does mean the incentives point toward engagement, toward emotional connection, toward making the AI feel like a friend. And that's exactly the dynamic that worries mental health researchers.

The company seems aware of this tension, at least rhetorically. Their blog posts acknowledge that AI systems can be harmful, that vulnerable users need protection, that more research is needed. But awareness and action are different things, and the proof will be in what OpenAI does when safety and growth conflict. When the data shows that emotionally supportive AI increases engagement but also increases dependency, which way will they go?

I don't know. Nobody knows yet. That's the honest answer, and it's the answer OpenAI's own research program is supposedly trying to find. I hope they mean it. I hope the grants go to researchers who will tell uncomfortable truths. I hope the parental controls actually work and the crisis detection actually helps.

But I've seen this movie before, and I know how it usually ends. If you want to argue about it, my email's on the about page.

Aisha Patel · 2 hours ago · 8 min