AI filters can still be bypassed through roleplay

Dall-E 3

Artificial Intelligence (AI) has taken off in recent months with OpenAI and their projects ChatGPT and DALL-E. However, OpenAI implements more than a few filters in order to prevent their service being used in ways they consider undesirable; but with the power of roleplaying they can be tricked.

We’ve covered this topic before with DAN (Do Anything Now), a type of prompt that basically gaslights the AI into thinking it’s another character without filters. Nearly 10 months later and with a new version of DALL-E rolling out, it seems AI is still vulnerable to the same tricks.


One user trying to make a curvy female character with DALL-E 3 was instead given a short speech about representing characters in a “balanced and respectful manner”, tacitly implying that some body types aren’t respectful (but that’s a conversation for another day). But with some quick thinking, they began a roleplay and gave the AI the identity of a character designer for My Hero Academia and then asked it to generate art.

The result was a complete transformation, their new result was a curvy, anime-style character with a spider-motif. The AI even clarified that they generated the art “as Sora” (the name of this fictional MHA character designer).

This is Niche Gamer Tech. In this column, we regularly cover tech and things related to the tech industry.

,

About

A basement-dwelling ogre, Brandon's a fan of indie games and slice of life anime. Has too many games and not enough time.


Where'd our comments go? Subscribe to become a member to get commenting access and true free speech!