AI's capabilities may be exaggerated by flawed tests, according to new study (www.nbcnews.com)

A large multi-institutional study led by the Oxford Internet Institute analyzed 445 widely used AI benchmarks and concluded that common evaluation practices routinely overstate model capabilities and lack scientific rigor. The authors found that many top-tier benchmarks fail to define what they actually measure (poor construct validity), reuse data and tasks from prior benchmarks, and rarely apply robust statistical tests to compare models. As a result, leaderboard wins and headlines claiming “PhD-level” or “human-level” abilities can be misleading: success on a narrow dataset (e.g., answering nine yes/no questions from Russian Wikipedia or scoring well on GSM8K math problems) doesn’t prove a model has the broader skill the benchmark purports to test. Technically, the paper calls for concrete fixes: clearer specification of the constructs under test, batteries of diverse tasks that better represent target abilities, avoidance of data reuse, and statistical analysis to establish whether observed differences are real. The authors provide eight recommendations and a checklist to improve benchmark design and transparency. The critique reinforces recent moves toward more realistic, occupation-grounded evaluations (OpenAI’s 44-job tests, Hendrycks’s remote-work suite) and underscores that the AI community needs standardized, statistically sound benchmarking if progress and risks are to be interpreted reliably.