Der er mange artikler om dette fx

Unmasking Clever Hans predictors and assessing what machines really learn | Nature Communications
Citat herfra:

A model implementing a 'Clever Hans'-type decision strategy will likely fail to provide correct classification and thereby usefulness once it is deployed in the real world, where spurious or artifactual correlations may not be present.

The issue of bias: the framing powers of ML (
Citater herfra:

A second problem is that the more data we have, the more spurious correlations or patterns can be found in the data, because unless data is truly random, many different mathematical relationships can be detected between the data.

testing for spurious correlations is not always obvious.

Agerbos Side