Emily M. Bender and Nanna Inie

In our op-ed for Tech Policy Press ("We Need to Talk About How We Talk About 'AI'"), we made the case against the anthropomorphizing language that makes it harder to have clear discussions of what so-called "AI" technologies actually do, and when and whether to use them. But these ways of speaking are deeply ingrained at this point, and it takes work carve new conversational and writing habits. That work involves at least three steps:

1. Noticing which word choices are anthropomorphizing
2. Finding alternatives
3. Getting in the habit of using the alternatives

In our research (summarized in the op-ed) we have been working on the first two steps, categorizing the kinds of anthropomorphizing language and using those categories to organize potential alternatives.

De-anthropomorphizing language talks about computer systems in terms of their functionality (what people build and/or use them to do), assigns agency to people using systems and not systems, and avoids aggrandizing metaphors about cognition.

We aim to find substitutes that are as self-explanatory as possible, so that you can just go ahead and use them without having to explain. (Though of course, if someone asks "Why are you calling it that?" that's also a great opening.)

Some of these rephrasings may feel a little clunky, and they can end up longer than the anthropomorphizing shorthand. This means it takes a little more dedication to use them, but also isn't necessarily a bad thing. We should stop and think about the tech we are using, or even discussing, and what it actually does.

We'll go through the categories of anthropomorphizing language we identified in Inie et al 2026, and give examples of de-anthropomorphized versions for each.

Our suggestions

Cognizer and products of cognition

This category is super frequent, because it's right in the marketing term artificial intelligence itself. This is language that locates thinking in an algorithm. Instead, we recommend describing software as performing calculations or other algorithmic operations, and locate the thinking with the people using the system. (In some cases, people clearly aren't thinking when they use them, but they are still the ones who should be.)

Examples:

artificial intelligence → probabilistic automation
hybrid intelligence → augmented human intelligence
image recognition → image labeling
speech recognition → automatic transcription
the model shows bias → the model reflects bias
model mistakes → model errors
chatbots are good at … → chatbots are good for …
hallucination → undesirable output

In general, we recommend avoiding using artificial intelligence or AI in reference to technologies. We do still talk about the AI industry, because that is the name of a thing, and talk about AI as an ideology. But when the intended referent is some specific technological system, it is always better to name that system itself. Maybe it's some specific product. Or maybe it's a system with a particular function like automatic transcription. Either way, it's worth finding names that aren't also anthropomorphizing. If you need a more general term, our recommendation of probabilistic automation above works for many (but not all) things sold as "AI".

We've also put hallucination in this category, because in its original sense it refers to perceiving things that are not there, but of course software systems (and conversation simulators in particular) don't perceive anything. Our proposed one-to-one replacement phrase is undesirable outputs, but it is also important to know that all LLM output is probablisitically produced synthetic text; there is no fundamental difference between desirable and undesirable outputs on the system side, but only for the people interpreting them.

Emotion

These are turns of phrase that suggest that software systems have emotional lives. We don't have particular rephrasings to recommend here because there is no accurate way to talk about emotional states of computers other than to reassert the obvious, that they don't have any. What's perhaps most subtle (and thus most fun for linguists) about this category is that these allusions to emotional experience can sneak in in surprising ways: If you say that ChatGPT struggles to do something, or that you had to coax it into some output, you are describing it as if it had emotional states.

Communication

In this category, we find words that place automated systems, usually synthetic text extruding machines, on an equal footing with people in communicative situations. If we ask something of Claude, we are describing Claude as a conversational partner. Instead of verbs like ask, say, inform, discuss, use verbs appropriate to computers like input and output. Another strategy is to foreground the fact of simulation.

Examples:

prompt → text input
answer → output
chatbot / conversational agent → conversation simulator

Agency

Turns of phrase that locate agency with a machine often serve to obfuscate the interests and goals of people. We suggest revising to locate agency with people or choosing less agentive verbs.

Examples:

ChatGPT assisted students → the students used ChatGPT
revealing the solution → displaying the solution
AI agent → probabilistic, unverified software manipulator

The elephant in the room of this category is the buzzword AI agent (and its variants like agentic AI systems). This is a term for software systems that connect LLMs (probabilistic synthetic text extruding machines) and/or other components up with other systems that can impact the world, i.e. systems previously designed for people to do things like schedule appointments, book flights, or make other purchases. Our suggestion for this one for now is probabilistic, unverified software manipulator, which has the advantage of giving a suitably gross acronym ("No thank you, I don't want to use your PUSMic system.") But, we are definitely open to other ideas! Send them our way and if any seem particularly apt, we will add them to this list.

Human Role Analogy

These are words that cast systems as doing the same work as people in various roles, and serve to hide all of the ways in which such automation falls short of what is needed all the while devaluing the actual work that people do and relationships that we form. Calling systems tutor or co-creator are overclaims that describe what a developer might wish they could develop—for those who want to replace people in these roles.

For this category, our recommendation is to use language that describes algorithms as tools (or products) that people use, rather than as human-like entities, and more clearly indicates system functionality while also not telegraphing a plan to replace people.

Names and Pronouns

The names and pronouns we use to refer to systems can also function in anthropomoprhizing ways. With system names, its somewhat trickier, because the system developers usually get to do the naming, and if they use a person's name for it, everyone else is stuck repeating that anthropomorphizing choice (we're looking at you, Anthropic, with Claude) or going for circumlocutions (Anthropic's conversation simulator).

Pronouns are chosen each time, and avoiding pronouns usually reserved for people (and pets), e.g. he, she, and singular they is a good first step. But subtle choices—such as grouping algorithms and people under you or them—can anthropomorphize. Separating systems from people and avoiding collective pronouns is preferable.

Examples:

who’s right? → is the machine output correct?
they produce results → the team uses it [the system] to produce results

Biological metaphors

Computer scientists working in "AI" (and its subfields) have been embedding biological metaphors in their technical terminology for a long time. These turns of phrase might have been metaphorical in their origins, but they also function to suggest more similarity than is actually there. When revising away from biological metaphors, ask how system functionality can be more precisely described to give readers a clearer sense of what is actually happening.

Examples:

neural networks → weighted networks (from Hunger 2023)
the model consumes data → data is used in setting model weights

Reflections

We encourage you to try out the above rephrasings and to create some of your own in the same spirit! It can feel awkward at first, but in our experience it is easier that reliably pronouncing or spelling the word anthropomorphization, so there is that.

It can also feel a bit socially awkward, because you are swimming against linguistic and cultural currents, but that can also be rewarding in and of itself. At a talk she gave in January, Emily was asked by a student how to contribute to the resistance against "AI" in conversations with friends, without being a stick in the mud. Emily said: Be a stick in the mud! If you think about our current situation as mired in mud that's hard to walk in, if you plant a stick, you can start to create firm ground for others to join you on.