Bender and Inie push back on anthropomorphic 'AI' language

'AI' is not your friend. That is the opening line of a Tech Policy Press essay by Emily M. Bender and Nanna Inie, and it sets the tone: the words we routinely use for these systems, they argue, are doing a lot of quiet work on their behalf.

Bender, a linguist at the University of Washington, and Inie, at the IT-University of Copenhagen, sort the trouble into a few buckets. There is cognitive and emotional language, verbs like 'think', 'recognize', 'understand', and nouns like 'reasoning', 'chain of thought', 'skills'. There is agent framing, sentences that put the automated system 'in the driver's seat' and 'obscure the actions and accountability of people building and using the systems'. And there is the language of communication, casting a prompt exchange as something two parties are doing together. Their claim is that all three risk masking limitations of probabilistic automation systems that make them different from human cognition.

The historical hook they lean on is Drew McDermott's 1976 critique of 'wishful mnemonics', the argument that naming the main loop of a program 'UNDERSTAND' does not prove the program understands anything. They also cite research suggesting that the label 'artificial intelligence' carries more perceived competence than 'machine learning' or 'decision support systems', and a 2025 study finding that people with lower AI literacy are more likely to perceive AI as magical and to feel awe. On the harm side they point to reports of 'AI psychosis' and to a finding that frequent chatbot use in exchanges that mimic romantic conversations is associated with higher depression and lower life satisfaction.

The honest caveat is that this is an argumentative essay, not a systematic review. The piece names studies and cases but does not hand you their methodologies, sample sizes, or the specific products behind the harms it cites, and a reader who wants to interrogate any single claim will have to go to the underlying sources.

Their proposed fix is small but sharp: talk about what a system is 'good for' rather than what it 'is good at'. That is a reframing anyone writing a product page, a policy memo, or an internal deck can adopt this week, and it is the part of the argument that travels beyond linguistics.