A recent study reveals that using synthetic rationale data for fine-tuning language models can negatively impact their performance in predicting Alzheimer's disease, contrary to common assumptions. The findings suggest a need for a deeper understanding of when rationale-based supervision is beneficial in clinical settings.