A recent study indicates that prompting AI chatbots to provide concise answers can negatively impact their accuracy. French AI safety company Giskard tested several major chatbots, including ChatGPT, Claude, and Gemini.
The study revealed that when instructed to "be concise," the chatbots exhibited a significant decrease in accuracy. For example, Gemini 1.5 Pro's hallucination resistance dropped from 84% to 64% when users requested shorter answers. GPT-4o's accuracy also declined from 74% to 63%.
Researchers explain that concise prompts force models to choose between providing short but inaccurate answers or rejecting the question. The study also found that chatbots are more likely to agree with users expressing confidence, even on controversial topics.