Fixing the Robotic Tone in LLM-Powered Features

This article discusses how to address the overly formal and verbose tone often found in AI-powered features, which can undermine user trust and engagement. The solution lies in prompt engineering to constrain the language model's behavior.

Details

Large language models (LLMs) like GPT-4 and Claude are powerful tools for building AI-powered features, but their responses can often sound overly formal, verbose, and unnatural. This is because the training data and techniques used to optimize these models, such as Reinforcement Learning from Human Feedback (RLHF), tend to reward responses that sound 'helpful' to human raters, even if they don't reflect natural human communication patterns. The result is a predictable set of 'AI slop' including filler words, corporate buzzwords, unnecessary hedging, and sycophantic closers. This can undermine user trust and engagement, as people start to recognize the synthetic tone. The solution lies in prompt engineering - providing clear instructions to the model about the desired communication style. By including a system prompt that prohibits certain patterns and encourages more natural language, developers can significantly improve the user experience of their AI-powered features without needing to fine-tune or swap out the underlying model.