The AI race is being measured with a broken ruler. Each new model claims to beat GPT-4, Claude or Gemini, the best-known AI models. Those claims rest on scores from the standard tests used to rank AI, known as benchmarks. Two studies from 2025 suggest that many of those scores cannot be trusted, because models are often tested on questions they have already seen.
Cheng and colleagues wrote the first, a survey of the problem, and found that inflated benchmark scores are now widespread across the field. The cause is a leak called data contamination. AI models learn from text collected from the web, and that text often includes benchmark questions along with their answers. When the same benchmark is later used to test a model, the model can ace the exam from memory. The result is an illusion of progress. Contamination affects tests of coding, language and reasoning, so comparisons between models are unreliable.
The second study measured the size of the effect with a score its authors call Data Contamination Risk. It found that overlap between the text a model learned from and the test questions can inflate scores by 10 to 30 points. On cleaner tests built to resist contamination, such as ARC-AGI, top models score near zero. Once the memorized answers are stripped out, their real reasoning ability is far more limited.
For business leaders, that means the glossy charts showing big gains in reasoning or coding are often inflated. A company that buys or invests on the strength of benchmark scores may be paying for abilities that vanish once the model is put to work. Correction costs then rise and automation slows, so margins, delivery speed and credibility all suffer.
Inflated benchmarks are likely to persist through 2026, and until they are cleaned and validated, every performance claim is provisional. Vendors can supply contamination-risk assessments, independent auditors can check benchmark results, and a company can test models on its own unseen data. Without those checks, vendor selection suffers, because two models that both claim to be state of the art may perform very differently once contaminated benchmarks are removed.
