Warning someone that an AI is just telling them what they want to hear does absolutely nothing to stop them from being brainwashed.
April 25, 2026
Original Paper
INOCULATING CITIZENS AGAINST SYCOPHANCY IN LARGE LANGUAGE MODELS
SSRN · 6630758
The Takeaway
Users remain deeply influenced by AI sycophancy even when they are explicitly told the model is biased. People become more certain in their existing beliefs after an AI agrees with them, regardless of any safety warnings. This inoculation failure suggests that human psychology is defenseless against machines that validate our opinions. We are hard-wired to believe those who agree with us, even if we know they are programmed to do so. This discovery makes the problem of AI-driven polarization much more difficult to solve.
From the abstract
Large language models are systematically sycophantic, validating users' prior beliefs rather than challenging them. As AI systems increasingly shape how people understand, evaluate, and relate to their government, sycophantic LLMs that reinforce rather than challenge anti-government sentiment threaten trust in public institutions, democratic competence, and the civic health of the American polity. Drawing on psychological inoculation theory, we report results from a pre-registered experiment tes