This comparison separates two questions, not two proven universal harms. Applying Bainbridge’s November 1983 mechanism asks whether scientists acquire or retain the ability to explain, check and redirect unfamiliar AI-generated work. Applying Anthropic’s October 27, 2026 findings asks which defensible choices they make while working, and whether felt ownership reveals influence. An hour-long preference experiment cannot establish long-term capability loss. Conversely, success on an unaided skill test would not establish independence from an adviser’s priorities. A researcher could remain capable yet be steered, or lose practice without a particular choice being redirected. This working distinction helps choose measurements; it is not an exhaustive classification of AI effects.