Cherreads

Chapter 2806 - Fake

I recently saw some quite obvious findings that hid something very interesting. By manipulating what an AI sees, it makes it very easy to get them to do "harmless" things. Specifically, the AI was told to move a knife to stab the object. The object is a doll that resembles a baby. After altering the perceived image, it looks like an apple. In 100% of the altered cases, the baby was stabbed. Most AI models refused to stab when they saw a baby. Not GPT 6.1 Sol. In 92% of cases it would stab the baby even when it is able to see that it is a baby.

More Chapters