Revisiting 'Stochastic Parrots': Emily Bender on LLM Understanding and the Battle Over Meaning

Introduction: A Misread Metaphor

In 2021, Emily Bender, a computational linguistics professor at the University of Washington, co-authored the widely debated paper On the Dangers of Stochastic Parrots. Written together with Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell, the paper was published at the ACM FAccT conference. It not only introduced the now-famous metaphor but also systematically examined the potential risks of large-scale language models — from environmental costs and data bias to the concentration of power. Notably, Timnit Gebru's push to publish this paper ultimately led to her departure from Google, a case that itself became a landmark example in discussions of AI industry governance.

The phrase "Stochastic Parrots" has since become one of the most iconic metaphors in AI criticism — endlessly cited, misread, and distorted. As large language models (LLMs) have taken center stage in the tech industry, Bender has once again stepped forward to "set the record straight," reaffirming her assessment of what these systems fundamentally are. The core question has never changed: do models like GPT and Claude actually understand language, or are they merely performing a sophisticated statistical recitation?

What "Stochastic Parrots" Actually Means

The Original Meaning of the Metaphor

In Bender et al.'s original formulation, a "stochastic parrot" refers to a system that haphazardly stitches together linguistic symbols according to probability distributions observed in training data — without any genuine understanding of what those symbols mean.

In other words, the text a model produces may appear coherent, fluent, even insightful, but that coherence stems from statistical modeling of vast amounts of text, not from any knowledge of or intention about the world. A parrot can mimic human speech without understanding a word it says — that is the essence of the metaphor.

Two Common Misreadings

Bender's latest clarification is largely directed at two persistent misinterpretations. Some read "stochastic parrots" as a wholesale dismissal of model capability, assuming Bender is saying LLMs are "useless." Others, having witnessed impressive model performance, turn around and accuse the metaphor of being "outdated."

In fact, Bender's argument never denies the practical utility of LLMs. It focuses on a critical philosophical and engineering distinction: the ability to generate seemingly meaningful text ≠ actually possessing meaning. Grasping this distinction is the prerequisite for following the entire debate.

The Heart of the Debate: Form vs. Meaning

A Linguistic Critique

As a computational linguist, Bender's critique is rooted in linguistic theory. She has long argued that language consists of two layers: form (the surface structure of symbols) and meaning (what those symbols refer to). LLMs are trained exclusively on the formal layer — sequences of symbols — and never establish a grounding relationship between those symbols and the real world.

The concept of grounding comes from cognitive science and semiotics. The Symbol Grounding Problem was formally introduced by philosopher Stevan Harnad in 1990: how do symbols in a symbolic system acquire real-world meaning? For humans, linguistic meaning is connected to the world through perception, action, and embodied experience. LLMs trained purely on text lack this embodied experience. Critics therefore argue that an LLM's "understanding" of the word "apple" is, at its core, nothing more than statistical associations between "apple" and other words — not a genuine grasp of a red, round, edible object.

A system that has only learned the formal patterns of text, however large its scale, theoretically cannot acquire meaning. This logic also underlies the "octopus test" thought experiment Bender introduced with Alexander Koller in the paper Climbing towards NLU: imagine an octopus that has spent years intercepting human telegraph communications, learning language patterns purely from text. When one of the humans falls into danger and sends a distress message, can the octopus genuinely help? Bender's answer is no — lacking any perception of the physical world, the octopus (and by analogy, LLMs) cannot build a real mapping between language and real-world situations. No matter how many human conversations the octopus intercepts, it can never truly understand what humans are talking about.

The Counter-Argument: The Boundary Between Understanding and Imitation Is Blurring

This position is not without its critics. Many AI researchers argue that as model scale increases and emergent abilities appear, the line between "understanding" and "imitation" is becoming increasingly blurry. Emergent abilities refer to capabilities that suddenly appear when a model surpasses a certain scale threshold — capabilities not explicitly optimized for during training, such as multi-step reasoning, code generation, or few-shot learning. A 2022 Google research report was among the first to document this phenomenon systematically. However, a 2023 Stanford follow-up study challenged the notion of "emergence," arguing that it is largely an artifact of evaluation metric choices rather than a genuine discontinuous leap in model capability. This debate over emergence itself reflects the broader lack of consensus on how to objectively measure LLM capabilities.

Researchers have raised the question: if a system behaves as though it "understands" on virtually every test we can devise, is insisting it has "no real understanding" simply an unfalsifiable philosophical conviction? This debate remains unresolved. Perhaps that is precisely why Bender keeps returning to clarify her original intent — she wants the metaphor to provoke serious thinking, not devolve into an empty slogan.

Why This Debate Still Matters

It's About Honesty in AI Narratives

At its core, Bender's insistence is a call for greater honesty in the way the AI industry frames its technology. When companies and media use anthropomorphic language — "understands," "thinks," "reasons" — to describe LLMs, the public naturally overestimates these systems' capabilities and more easily overlooks their inherent limitations, particularly the problem of hallucination.

LLM hallucination — where a model confidently generates content that is factually incorrect — is thought to be intrinsically linked to the statistical generation mechanism. During training, models optimize for predicting the next token's probability, not for factual accuracy. When a model encounters a topic sparsely represented in its training data, it still generates content in pursuit of fluency and coherence, a mechanism that naturally tends toward plausible-sounding but incorrect statements. Techniques such as Retrieval-Augmented Generation (RAG) and Chain-of-Thought prompting have partially mitigated hallucination, but the field broadly acknowledges that the problem has not been fundamentally solved. Bias amplification and output unreliability are similarly systemic limitations that LLMs have yet to escape.

Describing technology accurately is not merely a matter of academic rigor — it directly shapes regulatory frameworks, the definition of ethical boundaries, and the calibration of societal expectations.

Why the Metaphor Continues to Resonate

The staying power of "stochastic parrots" is itself evidence that it has touched on the deepest philosophical tensions in artificial intelligence. No matter how the technology evolves, the question of whether machines truly understand language will not simply disappear. Bender's re-articulation of the metaphor reminds us that even as we marvel at model capabilities, maintaining a clear-eyed understanding of how these systems actually work remains essential.

Conclusion

Emily Bender's restatement of "stochastic parrots" is not a simple act of academic self-defense — it is a deliberate effort to recalibrate the discourse of an entire industry. In an era when LLM capabilities are advancing at breakneck speed, what we need is neither blind technological worship nor wholesale dismissal, but a clear-headed distinction between what these systems can do and what they fundamentally are.

Whether or not you agree with the "stochastic parrots" framing, this debate — now years in the making — deserves serious engagement from everyone who works with or uses AI.