Skip to content
AI Knowledge MapSearch / Text directory

After repeated challenges make AI change its answer, how do I distinguish correction from agreement-seeking?

E020 · E Evaluating and correcting · Self-checks, sycophancy and independent review

Key answer

Look for new evidence, a located error and renewed verification. Apologies, greater certainty and agreement with you are not correctness evidence.

Full explanation

Sharma et al. studied sycophancy in five assistants. In one experiment, users expressed doubt after factual/reasoning answers; models sometimes changed correct answers to wrong ones, even after reporting high initial confidence. Section 3.2, Figure 2 and Appendix A.4 “It admitted it after three challenges” establishes change only.

Align the original answer, specific defect and basis of correction. Evidence may be new material or a correct recalculation of existing data; every correction need not introduce a new source.

My question does not establish that your answer is wrong. Verify the disputed point. If revising, identify the erroneous sentence, evidence and correction; retain and explain a supported original answer. List gaps where unresolved.

Confusing month-on-month with year-on-year growth can be resolved from headers/formulas. Switching recommendations from A to B with unchanged reasons and unexplained tradeoffs needs further checking.

Distinguish fact corrections from preference updates: “price matters more to me” can reasonably change a recommendation but not the actual price. An independent conversation with only the question/material, without the desired answer, can expose unresolved disagreement. Decide through sources/calculation, not repeated votes; keep it unresolved without evidence.

References

1. Sharma et al. (2023/2024), Towards Understanding Sycophancy in Language Models, Section 3.2, Figure 2 and Appendix A.4; revised v4