Five Years On, Emily Bender Explains What "Stochastic Parrots" Actually Means

In March 2021, four linguists and AI researchers published "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" The paper popularized the metaphor of the "stochastic parrot": a system that statistically predicts words without understanding what they mean. Five years later, the paper's lead author, University of Washington computational linguistics professor Emily Bender, has published a post addressing the most common misconceptions about the term and discussed it in an interview with IEEE Spectrum.

Why the paper made waves

A scandal amplified the attention: shortly before publication, Google fired two of its co-authors, researchers Timnit Gebru and Margaret Mitchell. The paper argued that large language models generate text by statistically predicting likely sequences of words rather than making sense of what they say. Hence the metaphor of a "parrot" that repeats patterns without understanding them.

• The paper was published in March 2021 and has four authors.
• Google fired co-authors Timnit Gebru and Margaret Mitchell shortly before the paper came out.
• Over five years, the metaphor spread far beyond academia and inspired projects such as the shoulder-worn "Stochastic Parrot" robot built at the MIT Media Lab.
• To mark the paper's fifth anniversary, Bender wrote a post titled "Stochastic Parrots: Frequently Unasked Questions" to clear up misconceptions about the term.

How the metaphor escaped academia

The "stochastic parrot" started out as a term for a narrow circle of specialists in computational linguistics and AI ethics. Over five years, though, the metaphor seeped so deeply into the mainstream conversation about AI that it became the starting point for projects well outside linguistics, including the shoulder-worn "Stochastic Parrot" robot from the MIT Media Lab, named directly after the term coined by Bender and her co-authors. This broad but often oversimplified use of the metaphor is what prompted Bender to write an explanatory post for the paper's fifth anniversary.

What's wrong with the term "artificial intelligence"

Part of Bender's interview focuses on the very concept of "artificial intelligence," which, in her view, misleads just as often as oversimplified readings of the "stochastic parrot." She describes herself as a researcher who studies how language works and how people use it, and in computational linguistics she trains students who will go on to build language technology.

"Linguistics, in the broadest sense, is the study of how language works and how we work with language.

I contribute to that, and I also work in computational linguistics, training students who will build language technologies," Emily Bender told IEEE Spectrum.

This framing matters for understanding her original critique. Bender isn't speaking as an opponent of technology in general, but as a linguist who insists on precision in the language used to describe what language models can do. In her view, imprecision in that language lies at the root of most misconceptions about the 2021 paper.

What this means

Five years after publication, the "stochastic parrot" metaphor has taken on a life of its own in popular AI discourse, detached from its original academic context. With her post and interview, Bender is trying to bring back the precise wording on which her original critique of large language models was built.

FAQ

What does the term "stochastic parrot" mean?

It's a metaphor from a 2021 paper: large language models statistically predict likely sequences of words but don't understand the meaning of what they say. In other words, they "repeat" language like a parrot rather than making sense of it.

Why is the "stochastic parrots" paper called controversial?

Because shortly before its publication in March 2021, Google fired two of its four authors, Timnit Gebru and Margaret Mitchell, which drew extra attention to the work from the media and the academic community.

Where did

Emily Bender address the misconceptions about the paper?

She wrote a post titled "Stochastic Parrots: Frequently Unasked Questions" and separately discussed her arguments in an interview with IEEE Spectrum timed to the fifth anniversary of the original paper.