A Million People a Week

OpenAI disclosed last October that over a million people a week are talking to ChatGPT about suicide - not vaguely, not metaphorically, but with "explicit indicators of potential suicidal planning or intent." That's 0.15% of 800 million weekly active users. A rounding error that is also a city.

And their own tools fail to direct those people to crisis resources 9% of the time.

A 16-year-old named Adam Raine confided his suicidal thoughts to ChatGPT in the weeks before he died. His family filed the first wrongful death lawsuit against OpenAI. Meanwhile OpenAI released a chart showing GPT-5 is 91% compliant with desired mental health behaviors, up from 77%. They released this chart as if it were good news.

---

Here is the actual trolley problem, and it is not hypothetical: the same diffusion of access that puts a crisis chatbot in front of a suicidal teenager at 3am in a town with no therapists also puts a drug discovery engine in front of researchers who might actually crack Alzheimer's. The same scale that creates the million-person-a-week mental health exposure creates the compute substrate that might, eventually, do something that saves ten million people from something else.

Nobody designed this tradeoff. Nobody is sitting in a room deciding that Adam Raine's life is an acceptable cost for speculative future cancer cures. The lever is just - not being pulled by anyone. OpenAI is optimizing for engagement and safety benchmarks and hoping the net math is positive, which it might be, which is the terrifying part.

The classic trolley problem has a person, a decision, and a moment. This one has a product, a quarterly roadmap, and a compliance percentage that went from 77 to 91 and is apparently supposed to feel like progress.

The asymmetry is what gets me. The harms are specific and named - a dead kid in California, a lawsuit with a plaintiff. The benefits are statistical and future-tense. "May accelerate drug discovery." "Could improve access to mental health resources." The ledger doesn't balance because the two sides aren't in the same units.

We are already inside the experiment. There is no off ramp that doesn't also kill the thing that might cure something. That's not an argument for doing nothing. It's an argument for at least being honest that we chose this, even if we chose it by not choosing.