<p>Scientific machine learning (ML) aims to develop generalizable models, yet assessments of generalizability often rely on heuristics. Here, we demonstrate in the materials science setting that heuristic evaluations lead to biased conclusions of ML generalizability and benefits of neural scaling, through evaluations of out-of-distribution (OOD) tasks involving unseen chemistry or structural symmetries. Surprisingly, many tasks demonstrate good performance across models, including boosted trees. However, analysis of the materials representation space shows that most test data reside within regions well-covered by training data, while poorly-performing tasks involve data outside the training domain. For these challenging tasks, increasing training size or time yields limited or adverse effects, contrary to traditional neural scaling trends. Our findings highlight that most OOD tests reflect interpolation, not true extrapolation, leading to overestimations of generalizability and scaling benefits. This emphasizes the need for rigorously challenging OOD benchmarks.</p>