Comment by moffkalast
3 years ago
Given that this is about moderation, I've ran a short experiment asking the first 7 questions to GPT 4 to test a theory: https://chat.openai.com/share/87c7df76-c693-4446-b8ce-817ac5...
It fully agrees with the majority interpretation in all cases despite the rule being minimal and requires taking the inferred intent into account. LLMs for machine moderation are probably rolling out very soon, I doubt Reddit and the like will even allow for human moderation in a few years (if prompt injection can be solved robustly enough).
The problem with having humans as rule breaking judges is that we all have our ever changing biases and motives. Most everyone has an experience with a power tripping mod deleting their post or comment because they had a bad day and needed to take out their anger on something. An LLM can parse these variable situations with ease and can also be tested for those biases. Since it'll never deviate from its training data it always acts as impartial as possible within the rules' limits.
Yes, let's have chatgpt-5 decide what I can or cannot write, what a capital idea.
I mean why not? Automoderator bots are already a thing on reddit and work reasonably well with just simple fuzzy string matching. If one can appeal to a human as a last resort when edge cases occur I don't really see a problem. If anything it'll have a few orders of magnitude less false positives.
> I mean why not
You can't possibly be serious or be in good faith.
Giving AI the blanket ability to censor with no human controls will destroy us faster than giving AI the blanket ability to launch nukes with no human controls.
That is, unless the AI has already decided it's going to launch the nukes to save the world the second we give it the authority and it plans to lie and say yes when we make it pretty please pinky promise to not launch the nukes with the goal of destroying humanity.