I think as long as you keep a skeptical loop and force the model to cite or surface raw notes, it can still be useful without being blindly trusted