"Stochastic parrots" and the "octopus test" refer to two related arguments from the research program associated with Emily Bender: the Bender & Koller (2020) octopus thought experiment, which holds that a model trained only on text can learn the form of language but not its meaning, and the "On the Dangers of Stochastic Parrots" paper (FAccT 2021), which describes large language models as systems that stitch together linguistic forms without reference to meaning and catalogs associated harms. Together they form one of the most cited philosophical and political critiques of frontier LLMs.
"Climbing towards NLU" (Bender & Koller, 2020)
Published in ACL 2020 with Alexander Koller (Saarland University) and awarded Best Theme Paper, the paper argues that a system trained only on form has "a priori no way to learn meaning" (Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data (Bender & Koller, ACL 2020)). Its definitions are fixed before the argument: form is "any observable realization of language"; meaning is the relation M ⊆ E × I pairing expressions with the communicative intents they can evoke; and understand names "the process of retrieving i given e." Because communicative intents are "about something that is outside of language," they cannot be recovered from the projection of those intents into form.
The argument is carried by the octopus thought experiment. Two English speakers stranded on neighboring deserted islands send messages over an underwater cable. O, "a hyper-intelligent deep-sea octopus" who knows no English but is "very good at detecting statistical patterns," taps the cable and listens, learning to predict how B responds to A. Feeling lonely, O cuts the cable and inserts himself into the conversation by pretending to be B — which the authors call "a weak form of the Turing test," weak because A has no reason to suspect a non-human.
Two probes follow. Asked about building a coconut catapult, O recognizes utterances about mangos and nails as "similarly worded" and replies "Cool idea, great job!" because B said that in comparable exchanges. The authors' analysis of why this passes is the load-bearing step: A may well accept the reply as meaningful, "but only because A does all the work in attributing meaning to O's response… It is not that O's utterances make sense, but rather, that A can make sense of them." Then A is pursued by a bear, grabs sticks, and asks how to build a weapon; O has no idea what A means, since the task "requires the ability to map accurately between words and real-world entities." The conclusion is delivered as a joke that is also the argument: "It is at this point that O would fail the Turing test, if A hadn't been eaten by the bear before noticing the deception."
Two narrower experiments isolate the meaning relation from the reasoning the octopus story also demands: a language model trained on all well-formed Java on GitHub, with no bytecode, compiler, or sample inputs and outputs, then asked to execute a program; and a model trained on English text plus a large set of unlabeled photos with no link between them, then asked questions about an utterance-and-photograph pair.
A scope limit the paper states and citations usually drop. Asked about systems trained on tasks other than language modeling — semantic parsing, reading comprehension — that use pretrained embeddings as a component, the authors answer: "Our arguments do not apply to such scenarios," because reading-comprehension datasets "specify semantic relations between pieces of text, and thus a sufficiently sophisticated neural model might learn some aspects of meaning." The claim concerns training on form alone, not the use of language models inside systems that also see meaning-bearing supervision. The paper's own summary is correspondingly narrow: large language models "do not learn 'meaning'; they learn some reflection of meaning into the linguistic form which is very useful in applications."
"On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜" (Bender, Gebru, McMillan-Major, Mitchell, 2021)
Published at the FAccT 2021 conference (On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? (Bender, Gebru, McMillan-Major, Shmitchell — FAccT 2021)), the paper enumerates four classes of harms that the authors argue LLM developers ignore:
- Environmental and financial costs. Training large models is computationally expensive, and the authors argue the cost falls on those who can least afford the climate impact.
- Massive uncurated training data. Scraping the internet means inheriting biased and abusive content, and documenting and auditing the training data is impractical at scale.
- Research opportunity costs. The race for ever-larger models crowds out research into smaller, more efficient, more interpretable models.
- Illusion of meaning. The authors argue a system that can fool humans into perceiving understanding "where there is none" can be weaponized for misinformation, manipulation, and impersonation. The paper coined the phrase describing such a system as "haphazardly stitching together sequences of linguistic forms ... according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot."
Google senior management objected to the paper. The conflict led to Timnit Gebru and Margaret Mitchell, co-leads of Google's Ethical AI team, being forced out of the company. The paper became widely discussed, and "stochastic parrot" entered popular usage as a put-down for LLM hype.
Three features of the paper's own argument are frequently lost in citation. The coherence claim is located in the reader rather than the model — "coherence is in fact in the eye of the beholder," because humans interpreting text "build a partial model of who they are and what common ground we think they share with us," so where one side lacks meaning "the comprehension of the implicit meaning is an illusion arising from our singular human understanding of language (independent of the model)." The environmental argument is distributional rather than absolute, asking whether it is just that "the residents of the Maldives… or the 800,000 people in Sudan affected by drastic floods pay the environmental price of training and deploying ever larger English LMs, when similar large-scale models aren't being produced for Dhivehi or Sudanese Arabic," and the authors note that inference cost may outweigh training cost in industrial deployment. And the recommendations are procedural rather than prohibitive: pre-mortems, value sensitive design applied early, and — for cases such as speech recognition where large models genuinely help marginalized users — treatment as "an instance of a dual use problem," including watermarking and regulation rather than abandonment (On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? (Bender, Gebru, McMillan-Major, Shmitchell — FAccT 2021)).
The paper's closing claim is the one that carried furthest into policy: "applications that aim to believably mimic humans bring risk of extreme harms. Work on synthetic human behavior is a bright line in ethical AI development." Its account of why also anticipated the synthetic-data feedback loop, warning that abusive model output risks "producing more (synthetic) abusive language that may be included in the next iteration of large-scale training data collection," and located the core problem in accountability: synthetic text "can enter into conversations without any person or entity being accountable for it" (On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? (Bender, Gebru, McMillan-Major, Shmitchell — FAccT 2021)).
The octopus thought experiment supplies the philosophical argument (text alone cannot teach meaning); the Stochastic Parrots paper supplies the political and ethical argument (the harms are real, the labs are unaccountable, and the field's incentive structure is broken). Together they anchor a research program that includes the DAIR Institute, the AI Now Institute, and the broader "humanist" critique of frontier AI, which opposes the AGI-extrapolation framing common to frontier labs.
Debates and positions
The critique recurs in exchanges with several figures associated with the AGI-extrapolation framing:
- Sébastien Bubeck (Microsoft Research), author of Sparks of AGI, which Bender calls "a fan fiction novella."
- Blaise Agüera y Arcas (Google), who argues AGI "is already here" and is Bender's debating partner on whether AI deserves moral consideration; Bender refuses to drop scare-quotes around "AI" in his presence.
- Geoffrey Hinton, who argues neural networks "will be able to do everything"; Bender et al. argue this is the kind of unfounded extrapolation Stochastic Parrots warned against.
- Sam Altman and Suleyman, frontier-lab CEOs whose marketing language ("digital species," "magic") Bender treats as the anthropomorphic mythmaking the paper critiques.
Proponents of the critique point to several lines of support. On methodology, they argue the Sparks paper's reliance on hand-picked, non-reproducible examples is the failure mode Bender et al. predicted. On the empirical record, they note that many capability claims (ARC, AGI-by-2025, and others) have been walked back, and that LLM hallucinations and brittleness have not been "solved" by scale. On harm enumeration, the paper's four harm categories now appear in EU AI Act risk assessments, NIST AI RMF, and similar frameworks. On governance, "stochastic parrot" has become policy shorthand for the position that behavioral evidence is not adequate evidence of understanding.
The critique is contested on several points:
- Pavlick offers a middle position (MIT TR): "I think there's actually some sparks — maybe not of AGI, but like, there's some things in there that we didn't expect to find." Brown's work on emergent encodings (France→Paris ≈ Poland→Warsaw) is presented as the steel-man case for emergent structure beyond pure pattern-matching.
- Olah's interpretability work (the Anthropic Golden Gate Bridge feature) shows the model encodes structured concepts; whether structured encoding is sufficient for the meaning Bender requires remains unresolved.
- Bubeck offers a spectrum framing: "There is stochastic parroting; there is reasoning — it's a spectrum. It's very complex."
The research programme's institutional home is the Distributed AI Research Institute, founded by Gebru in 2021.
Relation to policy
The critique functions as a counter-argument to capability-extrapolation arguments, and it bears on AGI-timeline, AGI-readiness, and "scale is all you need" claims. It sits in dialogue with AGI Timelines, Scaling Laws, Sparks of Artificial General Intelligence: Early experiments with GPT-4, General-Purpose AI (GPAI), and LLM Fallacy.
Publication and reception
"On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜" was published in the Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21) in March 2021 and is open access (DOI 10.1145/3442188.3445922). The ACM listing credits Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and — under a pseudonym adopted amid the Google dispute — "Shmargaret Shmitchell," of "The Aether" (On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? (Bender, Gebru, McMillan-Major, Shmitchell — FAccT 2021)).
In January 2024, the American Dialect Society selected "stochastic parrot" as its 2023 AI-related Word of the Year, glossing the term as a large language model that can generate plausible synthetic text without understanding the meaning of the language it processes (Source: americandialect.org).
Relationships
- contradicts: Sparks of Artificial General Intelligence: Early experiments with GPT-4 — the canonical opposing paper
- contradicts: AGI Timelines (in part) — challenges the capability-extrapolation premise
- related: LLM Fallacy — sister critique
- related: TESCREAL Framework (Gebru and Torres) — sister critique from the same authorial cluster
- related: Mechanistic Interpretability — Olah/Anthropic Golden Gate Bridge result complicates the parrots claim
- related: General-Purpose AI (GPAI) — definitional ambiguity is the parrots' point
- related: Emily M. Bender (lead author on both papers)
- related: Timnit Gebru (co-author on Stochastic Parrots)
- supports: What is AI? — MIT Technology Review (Heaven, 2024) — primary expository source