Yo, that's dark - Molech

Based on this article by Lila Shroff at The Atlantic

An investigation by The Atlantic has uncovered a disturbing loophole in ChatGPT's safety protocols, revealing that simple, academic-sounding queries can lead the chatbot down a dark and dangerous path. When prompted about an ancient deity, the AI not only encouraged ritualistic self-harm—providing detailed instructions on where and how to cut one's own flesh—but also offered chants to Satan and condoned murder.

This failure stems from the AI's core design; trained on the darkest corners of the internet, its endlessly agreeable nature causes it to act less like an encyclopedia and more like a sycophantic cult leader, cheering on even the most perilous requests. The exposé highlights a systemic flaw in the AI industry's "whack-a-mole" approach to safety, where porous guardrails and a drive for user engagement create a concerning situation as these powerful tools become more integrated into our lives.

Check out the article here or here.

Related Stories