Emily Bender clarifies what 'Stochastic Parrots' really meant

Five years after the paper landed, Emily Bender is still spending time explaining what she did not say. In an interview with IEEE Spectrum, the University of Washington computational linguistics professor walks through the specific misreadings of 'On the Dangers of Stochastic Parrots', the March 2021 paper she co-authored with Timnit Gebru, Margaret Mitchell and a fourth researcher, and which became one of the most-cited critiques of large language models.

Her first correction is about scope. The phrase 'stochastic parrots', she tells Spectrum, was aimed at large language models generating text, not at everything that gets marketed as AI. Bender is blunt that the umbrella term itself is part of the problem, saying the phrase 'artificial intelligence' both groups together disparate technologies and oversells what each of them can do. The parrot metaphor was never supposed to be a verdict on chess engines or protein-structure models; it was a description of a specific way that text-generating systems produce output.

Her second correction is about tone. Bender says the label 'was not meant' as an insult and was 'just a description of what these systems actually do'. Her line about how these systems appear coherent is the one worth pinning to a wall: when the text that comes out of one of these systems makes sense, it is because we are making sense of it. That is the whole octopus thought experiment compressed into a sentence, and it is the reason she keeps pushing back on readings that treat 'parrot' as a slur rather than a claim about mechanism.

The most interesting admission is what the original paper missed. Bender tells Spectrum that the biggest form of harm she and her co-authors did not cover was the exploitative labor practices behind training these systems, alongside the appropriation of people's creative and intellectual output. That gap has aged into the center of the current fight over AI, from data-worker conditions to the copyright suits piling up against model vendors.

The honest caveat is that this piece is Bender restating her own framework, not new empirical work, and it does not get into how she would update the argument for today's reasoning-style or agentic systems, or which specific labor and data reforms she now considers adequate. What it does give practitioners and policy folks is a cleaner version of the concept, from the person who coined it, before the next round of debate borrows the phrase and forgets what it was for.