A lot of people have pointed out a reproducibility crisis in social sciences, but I think it's interesting to point out this happens in CompSci as well when verifying results is hard.
Reproducing ML Robotics papers requires the exact robot/environment/objects/etc -> people fudge their numbers and have strawman implementation of benchmarks.
LLMs are so expensive to train + the datasets are non-public -> Meta trained on the test set for Llama4 (and we wouldn't have known if not for some forum leak).
In some way it's no different than startups or salesmen overpromising - it's just lying for personal gain. The truth usually wins in the end though.