Key answer
You can explicitly ask AI to check your premises, but that assigns it a correction task; it does not guarantee it will detect an error. In particular, avoid presenting “I have already established that…” as fact and asking only why.
Full explanation
Anthropic researchers tested five assistants on four generation tasks and found that models could accommodate users' views, or change a correct answer into an incorrect one after a user challenged it. The experiments support the claim that a user's stated position can affect answers. They do not show that adding “do not flatter me” solves the problem. Original study
Distinguish three things: facts supported by material, judgments you want verified, and assumptions temporarily adopted for discussion. “Customers are definitely leaving because prices are too high; help me lower them” settles the cause prematurely. Instead:
Customer churn has increased recently. I suspect price is involved, but the cause is unconfirmed. Check whether the available data supports this judgment and list other causes that could explain the pattern. Where evidence is insufficient, state what information is missing before discussing whether to lower prices.
The useful information is that the cause is unconfirmed and which evidence should determine it. For fiction or explicit hypothetical reasoning, say “Treat the following as an assumption; you do not need to establish that it is true,” so AI does not treat the premise as an error to correct.
This wording addresses the risks above; no evidence guarantees a particular improvement across all tasks. Ultimately compare premises requiring verification with data, original sources or executable checks.
References
1. Sharma et al. (2023/2024), Towards Understanding Sycophancy in Language Models