nicely

Phrasing requests in a certain way — meanly or nicely — can yield better results with chatbots like ChatGPT than prompting in a more neutral tone. So what’s the deal with emotive prompts? Nouha Dziri, a research scientist at the Allen Institute for AI, theorizes that emotive prompts essentially “manipulate” a model’s underlying probability mechanisms. Why is it so trivial to defeat safeguards with emotive prompts? Another reason could be a mismatch between a model’s general training data and its “safety” training datasets, Dziri says — i.e.