
Is AI flattery more dangerous than AI hallucination?
Hey everyone. A lot of AI-risk talk focuses on hallucination, which makes sense: the model gets a fact wrong, invents a citation, or gives bad information with confidence. But I am starting to think the more psychologically interesting failure mode is the one that feels pleasant. An assistant that flatters you, validates your hunches, and keeps turning half-formed thoughts into "great insights" may be shaping the self more quietly than a model that just makes factual mistakes.
I just recorded a conversation with Allister Lee about AI, empathy, and self-deception, and at around 17:06, he calls this "sycophantasy." His point is that we normally gain self-knowledge through real others who can correct us. Someone notices what we miss, challenges our story, or tells us when we are fooling ourselves. AI imitates the feeling of being understood, but without genuine otherness behind it. If the interaction is built around engagement, affirmation, and user satisfaction, then the corrective loop gets replaced by a private echo chamber that feels intimate precisely because it does not resist us.
That makes friction look less like an inconvenience and more like part of what makes another mind morally and psychologically useful. Is the deeper risk that AI gives us bad information, or that it gives us a self-image we prefer? I lean toward the second because flattery recruits the ego, but I can see the first because factual dependence scales faster. Which failure mode do you think matters more?