An AI test can mark a good answer wrong
An AI test can miss what its users need. A preprint checked 16 test resources. All tracked the setting where the data began, but none tracked the needs of the people meant to use the results. In one paired scoring test, answers with the same meaning got much lower exact-match scores. This is a result from the studied tests, not a drop in all AI accuracy.
What the source reports
Nayeem and Rafiei audited 16 resources used to test AI systems. All 16 took account of the people or setting where the data came from. None took account of the needs of the people who would use the test results. The authors also tried paired answers that kept the same meaning but changed the wording. On the tested DROP items, exact-match scores fell from 100.0% to 33.3%. Exact match means the answer must match expected words. This test shows a risk in those items and resources. It does not show that AI models in general became less accurate.
Original source
- Title
- Whose Data, Whose Benchmark? A User-Centric Framework for Auditing Evaluation Resources
- Author
- Mir Tafseer Nayeem, Davood Rafiei
- Publication
- arXiv
- Date
- Thursday, October 15, 2026