A new study offers a simple pre-prompt that coaxes AI models to reveal multiple possible answers instead of just one--restoring the diversity lost through alignment training.