The Stanford Prison Experiment only had 24 participants and implementation problems that should have concerned anyone with a pulse. But it’s been taught for decades.
A lot of psych research uses small samples. It’s a problem, but funding is limited and so it’s a start. Other researchers can take this and build upon it.
Anecdotally, watching people meltdown over the end of ChatGPT 4o indicates this is a bigger problem that 0.1%. And business wise, it would be odd if OpenAI kept an entire model available to serve that small a population.
The outcry when 4o was discontinued was such that open AI kept it on paying subscriptions. There are at least enough people attached to certain AI voices that it warrants a tech startup spending the resources to keep an old model around. That’s probably not an insignificant population.
How do we know if these examples aren’t just the 0.1% of the population that is, for all intend and purposes, “out there”?
So much of “news” is just finding these corner cases that evoke emotion, but ultimately have no impact.