Emily M. Bender and Nanna Inie have released a practical guide for developers and writers who want to stop attributing human traits to software. Published via their Tech Policy Press op-ed and expanded in a new blog post, the article argues that terms like "hallucination" and "AI agent" obscure the reality of probabilistic systems. The authors propose a three-step process: noticing anthropomorphizing language, finding functional alternatives, and building new habits. Their goal is not to be pedantic, but to ensure clear communication about what these tools actually do.

Replacing Cognitive Metaphors

The authors identify "cognizer" language as the most pervasive issue, noting that the term "artificial intelligence" itself locates thinking inside an algorithm. They recommend replacing "AI" with specific functional descriptions, such as "probabilistic automation" or "automatic transcription." For instance, instead of saying a model "shows bias," they suggest saying it "reflects bias," placing the source of the bias in the training data rather than the machine's judgment. Similarly, the term "hallucination" is discouraged because software does not perceive reality; the authors suggest "undesirable output" as a more accurate, albeit less catchy, alternative.

Agency and Communication

A significant portion of the guide addresses agency, arguing that current language often hides the people behind the technology. The phrase "ChatGPT assisted students" is critiqued for removing student agency; the authors prefer "the students used ChatGPT." The term "AI agent" receives special attention, with the authors proposing the deliberately clunky "probabilistic, unverified software manipulator" to highlight the lack of verification in these systems. They also advise against treating LLMs as conversational partners, suggesting "conversation simulator" instead of "chatbot" to clarify that no genuine dialogue is occurring.

Biological and Emotional Language

The guide also tackles biological metaphors like "neural networks" and "training," which the authors argue imply a similarity to human biology that doesn't exist. They suggest "weighted networks" as a more precise technical descriptor. Furthermore, the authors warn against emotional language, such as saying a model "struggles" or that a user "coaxed" it. These phrases falsely attribute emotional states to code. The authors acknowledge that these alternatives can feel socially awkward, but they argue that resisting linguistic trends is necessary to create "firm ground" for accurate technical discussion.

Key Takeaways

  • Replace "AI" with specific functional terms like "probabilistic automation" or "image labeling" when referring to technology.
  • Stop using "hallucination"; use "undesirable output" to describe incorrect LLM generations.
  • Shift agency from the software to the user: change "ChatGPT assisted students" to "the students used ChatGPT."
  • Avoid emotional verbs like "struggle" or "coax" when describing software performance.
  • Consider "conversation simulator" over "chatbot" to emphasize the lack of genuine interaction.

The Bottom Line

Language shapes perception, and if we keep calling software intelligent, we will keep expecting it to behave like a person. It’s time to embrace the clunky, precise terminology that reflects the probabilistic, unverified nature of current tech.