Bender and Inie Publish Six-Category AI Vocabulary Guide

The most persistent bias in AI reporting is baked into the vocabulary. When a system 'hallucinates,' a product 'understands,' or an agent 'decides,' the framing attributes cognition, intention, and responsibility to software before any analysis begins, obscuring what is actually happening and who made the decisions that shaped the system.

In a June 25 newsletter published via Mystery AI Hype Theater 3000, Emily M. Bender and Nanna Inie lay out a structured vocabulary guide organized around six categories of anthropomorphizing language: cognition, emotion, communication, agency, human-role analogies, and biological metaphors. For each, the guide offers specific substitutions: 'probabilistic automation' instead of 'artificial intelligence,' 'conversation simulator' instead of 'chatbot,' 'undesirable output' instead of 'hallucination,' 'weighted networks' instead of 'neural networks.' The underlying principle is consistent throughout: describe what software actually does algorithmically rather than borrowing from the vocabulary of human minds and bodies.

The agency category carries the sharpest practical edge. Bender and Inie argue that language attributing agency to machines 'obscures human responsibility and interests.' Rewriting 'ChatGPT assisted students' as 'the students used ChatGPT' is a small change that shifts accountability back to the people making choices. 'AI agent' becomes 'probabilistic, unverified software manipulator,' a phrase deliberately ungainly enough to signal genuine uncertainty about what such systems reliably do.

The honest caveat is that some substitutions will test any writer's patience under deadline pressure. Bender and Inie anticipate this resistance: they frame embracing 'the social awkwardness of resisting cultural linguistic norms' as the point, not a side effect, suggesting that 'planting a stick' of resistance creates 'firm ground for others to join you on.' What the piece does not address is how these alternatives would function in consumer-facing copy, regulatory disclosure requirements, or product marketing, where the pressure toward familiar vocabulary is strongest and whether there is empirical evidence that readers actually form more accurate mental models when the ungainly alternatives are used.

For practitioners in documentation, policy, or science communication, the six-category structure gives a systematic checklist rather than just a critique, and that is the more durable contribution.