Reanimation of pseudoscience in machine learning and its ethical repercussions (www.cell.com)

The article warns that modern machine learning can inadvertently revive debunked pseudoscientific ideas by packaging spurious correlations and biased datasets as “evidence.” It argues that when models are trained on noisy, unrepresentative, or historically biased data, they can reproduce—and appear to validate—dangerous concepts (e.g., physiognomy-like inferences, biologically essentialist claims about race or gender). The piece stresses that misinterpreting correlational model outputs as causal explanations, failing to run external validation, and over-relying on post hoc interpretability tricks turn statistical artifacts into seemingly scientific claims with real-world consequences. For the AI/ML community this is significant because it erodes trust, risks harm to marginalized groups, and can legitimize discriminatory policies. Key technical issues include dataset bias, confounding and proxy variables, p-hacking/multiple comparisons, lack of robustness checks, and misapplied explainability methods (saliency maps, feature importances) that invite over-interpretation. The authors call for concrete safeguards: stronger dataset documentation and provenance (datasheets/model cards), pre-registration of analyses, adversarial and out-of-distribution testing, causal inference methods rather than purely correlational claims, interdisciplinary review (social scientists, ethicists), and transparent reporting to prevent ML systems from reanimating pseudoscience under the guise of technical rigor.