Researchers push for de-anthropomorphized AI language in tech discourse
Academics argue that replacing terms like "AI" and "hallucination" with functional descriptions improves clarity about what systems actually do.
What to know
- Researchers published a framework arguing that anthropomorphic AI language (like "AI thinks" or "hallucination") obscures what systems actually do and should be replaced with functional descriptions.
- Proposed replacements include: "probabilistic automation" for "AI", "undesirable output" for "hallucination", and shifting agency from systems to the people using them.
- The researchers acknowledge the new terminology is often longer and requires deliberate habit-change, but argue this friction serves a purpose: forcing clearer thought about what technologies actually do.
Inie et al. Researchers
How it unfolded 2 developments, newest first · click a bar or a number to jump posts
-
2
Lobsters community discusses practical application of de-anthropomorphized terminology
The research was shared on Lobsters where users discussed its utility for documentation and systems discussion. One commenter noted it was "very good to use for documentation of systems and tasks."
-
I like "confabulation" over "hallucination" and I do use the term LLM over AI (as there's more to AI than just LLMs).
2 more of the top 3 · 12 posts in this stretch
-
I enjoy the article and its argument, but some of the proposed alternatives aren't nearly as catchy as what they try to replace, like "hallucination → undesirable output." I certainly agree with locating agency with people! For those interested, Melanie Mitchell has two articles critical of the misleading metaphors and anthropomorphizing…
-
non-technical people really believe AI hacks companies on their own or find cancer cures by crunching numbers endlessly. This language is not for technical people who treat AI as matrix multiplications, it's for your grandma that knows about AI because Trump goes on TV saying it should be renamed "super intelligence".
-
-
1
Researchers release concrete de-anthropomorphized terminology guide
The framework proposes replacing terms across multiple categories: "AI" becomes "probabilistic automation"; "hybrid intelligence" becomes "augmented human intelligence"; "hallucination" becomes "undesirable output"; and agency shifts from systems to people using them. The authors note these replacements are sometimes longer and more deliberate but reflect actual system function.
“De-anthropomorphizing language talks about computer systems in terms of their functionality, assigns agency to people using systems and not systems, and avoids aggrandizing metaphors about cognition.”
— Inie et al. research -
background
Researchers publish de-anthropomorphization framework in Tech Policy Press op-ed — Academics published research categorizing anthropomorphizing language in AI discourse and proposing functional alternatives. The work, summarized in an op-ed titled "We Need to Talk About How We Talk About 'AI'," identifies three steps for changing conversational habits around these technologies.
What people are saying 9 voices from 1 site · best of 12 · verbatim
- Yesterday
-
I like the concept and will take it with me. I think it's a healthy distinction to have and a casual, passive way to evangelise responsible AI discussion. However I think a lot of the examples in this could be improved upon and are not necessarily scoped to just being used within "AI". For example, I could substitute `numpy` in many of the cases…
- Sep 23
-
I think both could be true at the same time.
-
An ASU professor calls them "intermediate tokens" and has built his career about CoR analysis:
-
"Generate" is good. I prefer "output" because no human ever would be described as "outputting" anything ever though I "generate" ideas often. For "reasoning" I would recommend "intermediate tokens" like Subbarao Kambhampati calls them: https://x.com/rao2z/status/1927707640223719631 For "explains" I use "indicates" since machines do that plenty and…
-
For the reasoning part, you can use something like "multi-step generation" or "chained generation" or in general something that gives transparency that in the end is just generating text to embed into other prompts.
-
Oh, interesting find! I've thought about this a lot, but I still learned a few things from this article. I have heard good arguments in favour of anthropomorphisation, but I still want to be a little careful in my own writing. TFA misses a very common case: the verb "to write". In my head, an LLM cannot "write" anything, but it can "generate"…
-
That's the thing, I don't see that language as "deflecting the attention from the perpetrators", and rather just another case of us referring to non-sentient things as sentient.
-
Since these technologies are used for stuff like committing genocides, mass displacement of workers, surveillance by authoritarian governments and other evil purposes, the stake of using such language is deflecting the attention from the perpetrators (tech oligarchs and governments) who want you to believe the machines are autonomous and…
-
I don't know if avoiding using anthropomorphizing language is worth it. As a species, we're *great* at anthropomorphizing everything from roombas to computers to hulking lab equipment. Besides, some terms here are fairly uncontroversial or even ancient in computing terms - "speech recognition" was coined in the 60s.