xu2024-context-memory-87pct-drop

IN premise — summaries/2026/08/24/xu-2024-knowledge-conflicts-survey-sR-references-chunk-2.md

Created 2026-08-25T02:59:03+00:00

Misinformation in context causes up to 87% performance drop in ChatGPT on the NQ-1500 and CovidNews datasets (Pan et al. 2023, as reported in Xu et al. 2024 Table 2).

Summary

Feeding false or misleading information into a language model's context can destroy most of its accuracy, with one reported experiment showing ChatGPT's correct-answer rate falling by as much as 87 percent. This means any system that injects external content into a model's prompt, whether from search results, cached documents, or user-supplied text, must treat context contamination as a critical failure mode rather than a minor noise issue.