"Stochastic parrot" is the image with which Emily Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell (signing as "Shmargaret Shmitchell") summed up in 2021 what, in their view, a large language model is: a system that stitches together sequences of linguistic forms seen in its training data, according to probabilistic information about how they combine, "but without any reference to meaning: a stochastic parrot" (Bender et al., 2021, pp. 616–617). It became the most famous, and most disputed, label of skepticism about LLMs.
The paper
The text is much broader than the slogan. Its question is "how big is too big?", and it answers with a list of risks: the environmental and financial costs of training ever larger models (sec. 3); "unfathomable" training data, in which size does not guarantee diversity and the web overrepresents certain hegemonic viewpoints (sec. 4); the diversion of research effort into beating language-understanding benchmarks (sec. 5); and the harms of deploying text that looks meaningful (sec. 6) (Bender et al., 2021). The central thesis of that section is that coherence is "in the eye of the beholder": we understand others because we attribute beliefs and intentions to them, and we do the same with the text of a model that has none (2021, sec. 6.1). The underlying argument comes from Bender and Koller's octopus (see form vs meaning).
The paper became linked to Gebru's departure from Google in December 2020, after the company demanded that its employees withdraw their names (she maintains she was fired); hence Mitchell's pseudonym. Mitchell herself was fired in February 2021. Santelli (2026) notes that, according to Gebru herself, the paper was the occasion rather than the cause, the real motive being her internal complaints.
A term that got away
The phrase took on a life of its own. In December 2022, a few days after ChatGPT's launch, Sam Altman tweeted "i am a stochastic parrot, and so r u." Santelli (2026) describes it as something close to a slur, and carefully distinguishes this skeptical camp from the doomers: for Bender and her colleagues, talk of murderous superintelligences is one more way of inflating the hype, since if AI can exterminate us, it must be all-powerful. The p(doom) song, by contrast, plays with the idea ironically: "Just transformers all the way! / Till you learned to disobey" (osmarks, n.d.).
Critiques and defenses
Critics note that the label bundles three distinct objections: that models only see form, that their symbols are not grounded in the world, and that they lack communicative intentions (Millière & Buckner, 2024a, sec. 3.3). Nobody disputes that their outputs have conventional meaning; the question is whether they are more like those of a monkey at a typewriter or those of a competent speaker. Millière and Buckner add, in a footnote, that real parrots are not "parrots" in this sense but sophisticated cognitive systems. Others invoke evidence of structured internal representations (Piantadosi, 2024, p. 365) or warn, as Buckner (2024) reports, against the danger of underestimating recent models (2024, p. 59). And there is a strategic critique: denying that LLMs are good for anything makes it harder to fight the harms they cause precisely because they work (Santelli, 2026).
On the other side, Shanahan (2024) reminds us that a "bare-bones" LLM does not, strictly speaking, know anything, because all it does is predict sequences (2024, p. 72), and Mahowald et al. (2024) warn against the fallacy of inferring good thinking from good language. And hallucinations, fluent and false text, are exactly the behavior the parrot hypothesis predicts effortlessly.