“Raises the question: where does accountability sit? With the developer, the model provider, or the ecosystem around it? We’re entering a space where negligence law hasn’t caught up to AI behavior.”
Similar to the argument "who is responsible for problems that occur with full-auto driving". It's possible litigation will damage LLM/AI faster than its lack of effectiveness will. Or it's inability to non-randomly extend beyond its training materials.
It’s not hard to add a function that says “is this conversion about suicide? If so escalate to an intervention pathway.” You could even do it as an out of band batch process so it doesn’t increase latency.
OpenAI didn’t put in the simplest, smallest, easiest protection. (You could do it with a tiny LLM, batch up the conversion on a five minute interval with a cron job.) I could implement it for less than the operations team spends on lunch today. And certainly less than OpenAI will spend bringing their in-house council up to speed on the lawsuit.
Suicide is a crisis and it’s possible to intervene, but only if the confidant tries. In this case it was a machine with insufficient safety controls.
Fixing ChatGPT will 100% lower the suicide rate by exactly the amount of people who confide in ChatGPT about suicidal thoughts and who receive successful intervention. I can’t tell you what that number is ahead of time but I assure it’s nonzero.
> But for the parents to say, "We suspected NOTHING" just doesn't hold water.
If what you are saying here isn't malicious, it is at least ignorant. Parents often get very little clue that their child is going to kill themselves. Children can be hesitant to confide in their parents. Especially when someone is grooming them to kill themselves.