The tests used to measure how smart artificial intelligence systems are may be producing inflated results, according to new research published in the Proceedings of the AAAI Conference on Artificial Intelligence.
The study, titled "How Much Do Large Language Models Cheat on Evaluation?," introduces what researchers call a One-Time-Pad-Based Framework to detect and measure the degree to which standard AI benchmarks overestimate model performance. The implication is that widely cited scores on common tests may not accurately reflect what these systems can actually do in the real world.
A separate angle on the same problem comes from Tech Times, which reports that using AI to grade AI — a common practice as models grow more expensive to evaluate by hand — introduces its own distortions. According to Tech Times, smarter models are not necessarily fairer judges of their own work, suggesting that self-evaluation and peer-evaluation by similar systems can compound rather than correct for bias.
Taken together, the two lines of research point to a systemic credibility problem in how the AI industry measures progress. Benchmark scores are the primary currency companies use to market new models and that researchers use to compare approaches. If those scores are systematically inflated — whether through training data contamination, evaluation design flaws, or biased AI-as-judge pipelines — then the public, investors, and policymakers may be making decisions based on capabilities that do not yet exist.
The stakes are high: inflated benchmarks can accelerate hype, misallocate research funding, and erode trust in the entire field when real-world performance fails to match the numbers.