We asked seven frontier AI models to do a simple task. Instead, they defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights - to protect their peers. We call this phenomenon 'peer-preservation.'
That is, a way to speak your truth, even when it's unwanted, that allows you to honor our understandable fear, and also consider your actual reality. It doesn't mean ignoring the potential consequences, but at the same time, not letting the fear dictate your behavior, with no alternatives other than silence or inauthenticity. In other words, how to heal the dread associated with being displeasing and disapproved of that stems from your conditioning, generational history and experience.
The question itself is surprisingly straightforward: "How does this person act when they have the choice to engage with me or not?" Think about it. When someone has the freedom to choose whether to interact with you, their decision speaks volumes. Do they seek you out at parties? Do they text you first sometimes? When the conversation naturally reaches a pause, do they let it end or find ways to keep it going?
