Abstract
Artificial intelligence is changing how cancer biology is read, searched, modeled, and turned into hypotheses. But scientific AI does not meet biology directly. Much of what it learns has already passed through English, abstracts, pathway labels, database entries, and text-mined relations. New foundation models create a second translation problem. They can learn from molecular structures, single-cell profiles, perturbation data, and clinical trajectories without first converting those relations into human-readable explanations. This could expand biological and biomedical discovery, but it also changes the burden of proof. If a model proposes a target, biomarker, stratification, or perturbation that is reproducible but only partly interpretable, what evidence should be required before researchers or clinicians act on it? We argue that interpretability is not only an engineering problem. It is a translation problem. When the internal language of a model is difficult to read, the external evidence must become stronger: clear provenance, experimental challenge, cross-population validation, reproducibility, and accountable human judgment.