ITG NowHome › AI glossary › Benchmark contamination
What is Benchmark contamination?
Benchmark contamination is when an AI has already seen a test's questions or answers, so its score looks better than its real skill.
- Category Benchmarks
- Also called data contamination
- Updated 20 Oct 2026
Benchmark contamination, explained simply
AI models learn from huge piles of text from the internet. Test questions and answers can end up in that pile.
If a model has seen the answers, it may just recall them. The high score then says little about real ability.
This matters because it can make a model look smarter than it is. Be careful with claims based on one old test.
Makers fight it with fresh questions, hidden test sets and checks. It is also called data contamination or leakage.
On ITG Now, the AI models page links each benchmark to its source, so you can see how it is run and how often it is updated.
Related terms
BenchmarkTraining dataLiveBenchLeaderboardTrainingAI model
Questions people ask about Benchmark contamination
How does benchmark contamination happen?
Test material gets published online. Later it is swept into the data used to train a model.
How can you spot contamination?
It is hard to see. Fresh tests, hidden questions and sudden jumps in score on old tests are clues.
Is contamination always cheating?
Not always. It often happens by accident because the internet is so large.