Hot topics

Is “AI” the right word?

What ChatGPT-style systems actually do at the mechanical level, the researchers who argue “stochastic parrot” describes it better, and why “AI” isn't going anywhere regardless.

Last updated September 22, 2026

Terms

LLM: Large language model — the category of system (GPT-4, Claude, Gemini, and similar) this page is about. "AI" is used elsewhere on this page to mean the broader popular term; "LLM" is the more specific, technical name for this one kind of AI system.

A 68-year-old umbrella term

"Artificial intelligence" is older than almost everything it now gets applied to. John McCarthy coined it for a 1956 Dartmouth research proposal, specifically to distinguish his group's work from the era's dominant framework, cybernetics. From the start it was a broad label: over the following decades it came to cover chess engines, spam filters, route-planning software, protein-folding models, and, since 2022, the chatbots most people now mean when they say "AI." Those are extremely different pieces of technology, built different ways, doing different things — grouped under one word mostly because each one, when it arrived, seemed to do something previously thought to require a mind.

What a large language model actually does

Strip away the branding and a system like ChatGPT is, mechanically, a next-word predictor. It's a neural network trained on a very large amount of text to do one thing: given the words so far, estimate the probability of what word comes next, then sample from that estimate. Do that one word at a time, repeatedly, and you get paragraphs. Nothing in that training objective requires the system to track whether what it's saying is true, to hold a model of the conversation's goal, or to know what any of the words refer to in the world — it only has to get good at predicting text that resembles its training data.

"A stochastic parrot"

That's the mechanism a group of four researchers — Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell — were describing in a widely cited 2021 paper, "On the Dangers of Stochastic Parrots." Their exact words:

"an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot."

Bender has since been explicit that this isn't an insult, just a description — and a narrow one, aimed specifically at large language models producing text, not at every technology people call "AI." Her broader complaint about the umbrella term itself: "The phrase 'artificial intelligence' both groups together disparate technologies and oversells what each one of them can do."

Not everyone agrees this is the whole story

This is a live, unresolved argument among serious researchers, not a settled question with an obvious winner — worth presenting both sides of rather than picking one.

Evidence for "just" pattern-matching

OpenAI's own 2023 technical report claimed GPT-4 scored in the 90th percentile on the Uniform Bar Exam — widely cited as evidence of real capability. A 2024 re-analysis found that comparison pool was skewed toward repeat test-takers who'd already failed once, and that GPT-4's actual standing among first-time takers was far lower — a caution about taking benchmark claims at face value, which cuts both ways in this debate.

Evidence for something more

Researchers training small transformer models purely to predict legal moves in the board game Othello found the models spontaneously developed an internal representation of the actual board state — something never explicitly taught. Separate interpretability work at Anthropic on production-scale models has found internal circuits that appear to plan several words ahead and operate independently of the language the model ultimately answers in. Whether that adds up to "understanding" in any meaningful sense is exactly what's contested.

Why "AI" isn't going away

Even researchers who think the label oversells the technology mostly agree it isn't disappearing, for reasons that have little to do with accuracy. It's short. It's the name the academic field has carried since 1956, with departments, funding lines, and decades of publications built on it. And it benefits from what's sometimes called the AI effect: once a capability is achieved, the public tends to stop calling it "AI" and starts calling it "just software" — optical character recognition and GPS routing were both once considered artificial intelligence research. The goalposts move; the label persists for whatever's newest. "Large language model" or "next-token predictor" are more precise, but "AI" is what fits in a headline and what everyone already understands well enough to keep using.

Nonpartisan, plainly

Whether large language models "understand" anything in a meaningful sense is a genuine open question among researchers who study this for a living — we're not in a position to settle it, and we don't take a side. What's not in dispute: what the training objective actually is, what a specific group of researchers has argued that objective implies, and what other researchers have found that complicates the simplest version of that argument. These are the sourced facts, plainly. See also our pages on why some AI models don't follow shutdown commands and on recent AI extinction-risk warnings, both specific, tested angles on the same underlying question.

Talking points

These are the questions we think you should ask those who are running for office and will represent you. We don't give our opinion on the answer, but we DO think you should be talking about them.

Should AI companies be required to disclose, in plain language, how a model actually works when marketing it as "AI"?
Should the term "AI" carry any legal or regulatory definition, given how loosely it's used today?

Bender, Gebru, McMillan-Major & Shmitchell (2021): "On the Dangers of Stochastic Parrots"dl.acm.org
The paper that coined the term, in the ACM Conference on Fairness, Accountability, and Transparency.
IEEE Spectrum: What Emily Bender Really Meant by "Stochastic Parrots"spectrum.ieee.org
Bender, five years later, on what the term does and doesn't claim — and her critique of "AI" as a label.
Li et al. (2023): "Emergent World Representations" (Othello-GPT)arxiv.org
The interpretability study behind the counterargument that some internal representation can emerge from next-token training.
Martínez (2024): "Re-Evaluating GPT-4's Bar Exam Performance"law-ai.org
The follow-up study behind the bar-exam caveat below.
Dartmouth: "Artificial Intelligence (AI) Coined at Dartmouth"home.dartmouth.edu
The 1956 proposal where the term was first written down, and why McCarthy chose it.