Hacker News Reader

This may seem silly but can there be a text transformation/translation layer, where the conversation text is ultimately PG, the user thinks they have sent their thoughts, but they are translated and vice versa for the user and AI, so its effectively a text filter in between keeping all parties "safe" and within policy? But the UX is same as before.

View on HN

So like if the user input is "I want to ** your **" it would change it to "I want to enjoy your presence"? I'm pretty sure that would leak (be parroted back and not caught by a filter, thus exposing its happening), but it might work. It might put <censored> or something like that instead, which the model would take into consideration and censor itself.

Sending PG text to the users is already what the company started doing, so I'm not sure how that's different.

Sending PG text to the AI would be for... what? Not offending the AI?

So the users don't have to self-moderate to maintain conversations.

Reply on news.ycombinator.com