A crossover study involving 30 ophthalmologists and trainees reveals that while Claude-3.5-Sonnet collaboration improves diagnostic accuracy in complex cases, it also increases confidence in incorrect decisions and fails to reach the performance of the AI alone.