All news

Research Shows AI Performance Changes with the Emotional Tone of Prompts

Think of a question you could ask calmly or with fear. Manuel Gozzi and Francesca Fallucchi changed the tone of requests while keeping the requests themselves the same. They tested five language models on eight tasks.

The models scored up to 4.5 percentage points higher with positive or neutral wording than with fearful wording. In another study, Ziqi Yin and colleagues tested politeness in English, Chinese and Japanese. Rude wording often hurt performance. Extra politeness sometimes hurt it too.

AI is awakening consciousness, and how we speak to it is worth studying. The next step is to look inside the model while it responds. Then we can see what changes as the tone of the conversation changes.

Read the source article

All news