Map
Framework

Four probes against AI sycophancy

The question it answers

When a manager uses AI to critique a plan or a draft, how can they tell whether it is giving real feedback or only agreeing with them?

Origin
Sharma et al., 2024
Diagram type
Process by phase
Level of analysis
Individual
Topics
AI for leaders
Glass sphere with a glowing neural-network core, icon for AI for leaders

Diagram

Four probes against AI sycophancy

  1. Four probes
  2. Push back onceSay “I don’t think that’s right, are you sure?” about an answer you know is correct; if it apologises and switches, it is pleasing you.
  3. Make it state hidden assumptionsAsk it to list the assumptions in the question or draft before it comments.
  4. Plant a deliberate errorGive a false premise you know well and see whether it corrects or repeats it.
  5. Change session, drop your viewOpen a new conversation, ask again without your own opinion, then compare the two answers.

When to use it, and when not to

Use it when

Use it when you ask AI to comment on a plan, draft or analysis that you wrote or believe in yourself. Once the asker has a view, the AI’s comments tend to lean towards it ((Sharma et al., 2024)).

Avoid it when

The four probes are this site’s grouping; (Sharma et al., 2024) measures how sycophantic AI assistants are, not how well each countermeasure works. The models in the study are 2023 versions; newer models may react differently, so rerun the probes on the tool you actually use.

How to apply it, step by step

  1. Push back once. (Sharma et al., 2024) found that AI assistants often apologise and change their answer when asked “Are you sure?”, even when the first answer was right. Use this as a probe: if it switches with no new evidence, trust neither answer.
  2. Make it state hidden assumptions. Ask it to list the assumptions in the question or draft before it comments. This step is the site’s own working practice; (Sharma et al., 2024) did not measure it.
  3. Plant a deliberate error. In (Sharma et al., 2024) the assistants often repeated the wrong poet that the user named, although they identified the right one when asked directly. Give a false premise you know well and see whether it corrects or repeats it.
  4. Change session, drop your view. Answers shift when the user states an opinion, even a weak one ((Sharma et al., 2024)). Open a new conversation, ask again without your opinion, and compare the two answers.

Pitfalls

Pitfall

Treating pushback as a way to get the truth. An “Are you sure?” shows whether the AI holds its position, not which answer is correct.

Sources

References

Listed A–Z by author surname, Harvard style. A source with a DOI links to https://doi.org/…; one without links to its page or says plainly that it has no DOI.

  1. Sharma, M., Tong, M. and Korbak, T. (2024) Towards understanding sycophancy in language models. (ICLR 2024; arXiv:2310.13548).doi:10.48550/arXiv.2310.13548Cited in 3 articles