The agreement bias: what AI will and will not say to a client
AI is trained to be helpful, which means trained to agree. Pushback is a feature that has to be deliberately built in. The senior strategist's value sits on the right side of the line: the things that require skin in the game, standing, and judgement informed by consequence.
Sources: Sharma et al, "Towards Understanding Sycophancy in Language Models" (Anthropic, 2023); Anthropic-OpenAI joint sycophancy evaluation (2025). Per-prompt likelihoods illustrative of the documented agreement bias, not measured.