Stop Talking About AI Like It’s Human

Given the volume of chatter around AI, it’s worthwhile to examine the vocabulary we use to describe it. A lot of words used to describe AI are personifying. Maybe you remember from English class that personification means associating inanimate objects with human traits. In literature, if we want to emphasize the sorrow a character is feeling, we might write something like “Michael was devastated by his bad luck losing all his money. He didn’t even notice he walked into the forest, right into the rainstorm, drenching him. The sky wept torrents of rain, and the trees reached out to pull him into the darkness.”

In this example, the author personifies aspects of the setting to mirror or emphasize the character’s feelings. The sky cannot weep like a person cries; it can rain heavily. Trees don’t reach out to touch people; they are moved by the wind. Personifications are vivid, useful shortcuts and we use them all the time to efficiently describe things. How would it sound if we instead described the scenery in the following way: “Humidity in the air accumulated into a precipitate and fell from the sky, drenching Michael”.

As useful as personification can be, I want to use this article an exercise to attempt to define concepts in AI without using personifications. AI technology and LLMs in particular often appear to have human characteristics, even though this is an illusion (an extremely useful illusion, but an illusion nonetheless). So that’s why I think we have to be especially cautious when using personifications and anthropomorphic language to describe it.

Agent - This is a technical term existing for decades, but it can be contorted by a marketing team to make the technology sound spooky and autonomous. It can evoke the CIA or Mr. Anderson from the Matrix. But an agent is a computer program that interacts with an LLM–a system outside of the immediate program–in a circular feedback loop. A typical computer program constantly interacts with systems outside itself. Microsoft Word interacts with the software that controls your keyboard, reaches into the operating system’s memory to copy and paste, and hands your document off to Adobe when you request to print to PDF. Apps we use every day write to databases and make network calls to far off servers without us even thinking about. (Note again, the personified language I can’t even escape here: “interacts”, “reaches”, “hands off”, “writes”, “calls”, but nobody assumes that Microsoft Word has a brain!)

An agentic program adds one more exterior system that a program can interact with, an LLM. (See the article breaking down what an LLM is). The LLM itself can't do anything; all it ever produces is text. The programmer writes functions the program knows how to carry out, “save this file,” “send this email” and wires them up so that when the LLM's text says "send an email," the function runs. But the program is the thing that actually sends it. After a function executes, the program records the result as text, sends the result as text back into the LLM, and the LLM outputs another bit of text. Continuing the loop, that output text either signals the program to run another execution or end the loop (see the interactive simulation from an “agent’s perspective”). To beat a dead horse, when an agent “sends an email” or “saves a file” an “agent” didn’t decide or even do that on its own: a human wrote the function, connected it, and allowed it to run. As an extreme example, if a programmer accidentally leaves a switch in the program that can wipe the company’s entire mainframe, and a stray sentence from the LLM can flip the switch, then sure, something's going to go wrong eventually. But that's not the LLM's fault. (There is a lot to talk about here related to agents and agentic code patterns, so stay tuned for more on this!)

Behavior - It’s worth reiterating what we discussed above, an LLM never “behaves” in one way or another, in fact, it doesn’t do anything besides output text that a program uses to execute functions we’re all familiar with. Suggested alternative: Computer program actions taken based on text output from an LLM.

Bias - This is a serious technical term meaning lopsided patterns in the data used to train the LLM get carried through to its output. LLMs cannot be "prejudiced" in the human sense, somehow taking on some sort of emotional opinion or motive. Suggested replacement: a statistical pattern causing an LLM output to skew to favor certain kinds of responses.

Hallucination - Now here’s a scary word with overtly negative connotations! The last time I had a hallucination was when I had a high fever, and it was a disturbing experience. Even more confusingly, hallucination connotes perception, which is hardly analogous to what it's trying to convey in the context of AI. A “hallucination” refers to when the LLM produces output that looks and sounds great, is internally coherent, presented as correct, but is actually false. Suggested replacement: false output from an LLM that unfortunately appears logical. Confidence without substance.

Learning, training: An LLM contains billions of adjustable values that can be changed to help produce the desired output patterns. “Training” means feeding an entry from a dataset, computing an output, and measuring the mathematical “distance” of the output from the target. Alternative: mathematical optimization.

Neural/neuron - The term neuron in terms of machine learning is a metaphor for a neuron-inspired mathematics model. These mathematical models are obviously very powerful but in reality barely hold a candle to the actual biological structure. Suggested replacement: neuron-inspired models.

Thinking - LLMs don’t think. If we reduce the LLM to a black box where text gets input and different text gets output (again, see my previous article on how LLMs work), then what happens inside the box is not “thinking”: it’s billions of mathematical operations happening at an extremely high speed. If we paused and examined one of these operations, someone with an understanding of high school algebra could follow the logic. But the value comes from the insanely large volume of the operations. Suggested replacement: Processing.

Taking - Here’s another big one that the news loves to use rile us up. Is AI going to “take” your job? This is a great example where AI is exploited to shift blame: The innocent CEO didn’t decide to fire half the company, AI just took the jobs instead. Suggested alternative: Companies choosing to automate jobs. Or even better, let’s just let people own their decisions and actions (good luck on that one!).

Are these terms bad? No, and it’s worth repeating that many of them are established terms from technical fields, coined by fine people who wanted to make these concepts easy to understand. Personifications often are a much more concise and efficient way to convey function. The clockface “tells” us the time, the train "belches” out steam. LLMs are different because they appear to talk back to us: the clock “telling” the time remains an obvious figure of speech, but meanwhile an LLM produces fluent sentences. So when we say something like “the LLM thinks,” we are more prone to read the metaphor as a fact. The marketing department exploits this to convince us AI is a magic genie that will solve all of our problems. Maybe even more concerningly, a human-sounding entity is a perfect scapegoat when something goes wrong: It wasn’t the engineer’s fault that everyone’s bank routing number was exposed, it was simply the evil AI breaking out of its shackles and running amok.