Hackaday · Maken

Avoiding Anthropomorphism When Discussing LLMs

Emily M. Bender and Nanna Inie propose language adjustments to prevent people from anthropomorphizing Large Language Models (LLMs). Their analysis highlights how common language used to describe LLMs implies human-like qualities that the software does not possess.

The authors observe that terms like "recognize" in "speech recognition" and "hallucination" suggest a level of consciousness absent in LLMs. They suggest alternatives such as "automatic transcription" for the former and "undesirable output" for the latter, though they acknowledge the latter isn't perfect.

Bender and Inie also recommend using "input" and "output" instead of "prompt" and "response" to maintain a more technical and less human-centric description of LLM interactions.

Given the increasing prevalence of LLMs, the authors believe evolving the language used to discuss them is important. They suggest their proposed changes could help foster a more accurate understanding of these technologies.

AI-samenvatting op basis van de bron.