Overcoming AI making spurious correlations causing decisions based on unimportant and potentially misleading information

https://techxplore.com/news/2025-04-technique-spurious-problem-ai.html

"spurious correlations traced to very small subset of training data, generally caused by simplicity bias during training... overcome even when have no idea what spurious correlations AI is relying on, by removing a small portion of training data... we can measure each sample's difficulty based on how model behaved during training, most difficult samples in dataset noisy/ ambiguous, most likely relying on irrelevant information... this elimination overcomes spurious correlations, without causing significant adverse effects"