How to Discuss AI Without Anthropomorphizing It

Summary

Emily M. Bender and Nanna Inie argue that common ways of talking about AI often make software systems sound as if they think, feel, communicate, or act like people. Drawing on the categories developed in their research and summarized in an earlier op-ed, they recommend describing systems through their functions and assigning agency to the people who build or use them. Their examples include replacing “artificial intelligence” with “probabilistic automation,” “image recognition” with “image labeling,” “hallucination” with “undesirable output,” and “prompt” and “answer” with “text input” and “output.” They also question terms such as “AI agent,” “conversational agent,” “tutor,” and “co-creator,” arguing that these labels can obscure uncertainty, limitations, and the human labor that the systems do not replace. The authors advise avoiding emotional language such as saying a chatbot “struggles,” as well as pronouns and collective phrasing that place machines and people in the same social role. They recommend more precise descriptions for biological metaphors too, such as “weighted networks” instead of “neural networks” and describing how data is used to set model weights rather than saying a model “consumes” data. The proposed wording may be longer or initially awkward, but the authors present that friction as useful because it encourages people to examine what a system actually does. They retain “AI” for references such as the AI industry or AI as an ideology, while generally recommending that specific technologies be named by their products or functions.