How it works

Researchers tested open-ended continuation prompts on ChatGPT, measuring how far output diverged from user intent across varied task types.