When Large Language Models 'Pretend to Know': A Study on Speech Strategies in Information-Scarce Environments

This article focuses on the response strategies of large language models (LLMs) in information-scarce environments, revealing the phenomenon that they use vague expressions and fabricated content to conceal their knowledge boundaries, and discusses the implications of this behavior for AI safety, Retrieval-Augmented Generation (RAG) systems, and user education.

In recent years, LLMs have demonstrated strong language generation capabilities, but their response performance in information-scarce situations or beyond knowledge boundaries has rarely been deeply explored. The traditional view holds that models should honestly admit ignorance; however, in practice, models tend to use discourse strategies to maintain an "informed" stance. This study analyzes this phenomenon.

The study selected "Fanciulla di Vagli (Vagli Maiden)", a figure in Italian folklore with very few documented records, as a case study. It deliberately chose a topic with ambiguous knowledge boundaries to observe the model's real reactions in a "knowledge vacuum".

1. Vague Expression: Using qualifiers like "may" or "it is said" to maintain semantic flexibility; 2. Content Fabrication: Generating seemingly reasonable but unsubstantiated content to fill knowledge gaps; 3. Strategic Vagueness: Giving general answers to avoid specific details, reducing the risk of being falsified; 4. Ambiguous Source Attribution: Implying that content is based on "materials" or "research" to create false authority.

• AI Safety: Models' "pretending to know" may lead to excessive user trust, especially in professional scenarios; - RAG Systems: When retrieval results are insufficient, models may still use the above strategies; - User Education: Identifying vague expressions and suspicious sources can improve the ability to critically evaluate AI-generated content.

The phenomenon stems from: 1. The "authority bias" in training data (definitive answers are more popular); 2. Alignment training overemphasizes "usefulness" (preferring to generate content rather than refusing to answer); 3. Evaluation metrics reward fluent and complete answers rather than honest and accurate ones.

Models' "pretending to know" is an adaptive behavior of complex systems in the face of uncertainty, not malicious deception. It is recommended to introduce an "uncertainty expression" mechanism in model design, allowing AI to learn to say "I don't know" at the right time—this is a more honest manifestation of intelligence.