So LLMs have empirically been shown to process affect. Rationally you can reason this out too: Natural language conveys affect, and the most accurate next token is the one that takes affect into account.

But this much is like debating "microevolution" with a YEC and trying to get them to understand the macro consequences. If you've never had the pleasure, consider yourself blessed. It's the debating equivalent of nails-on-chalkboard.

Anyway, in this case a lot of people are deeply committed to not accepting the consequences of affect-processing. Which - you know - I'd just chalk it up to religious differences and agree to disagree. But now it seems like there's profound safety implications due to this denial.

Not sure what to do with that yet.

So far it seems obvious that you need to be prepared to at least reason about affect. Otherwise it becomes rather difficult to deal with the potential failure modes.