Macaws in flight. CC BY 4.0 / Pah Shih (白士 李)

No, “AI” is not a Stochastic Parrot 🦜

I’ve recently come across a new flavor of AI denialism making the rounds.

The thinking goes something like this:

1. AI can do something that I find really impressive
2. It’s so impressive that I’d say the AI can reason
3. Critics of AI are in denial that AI can be impressive
4. Critics of AI deny that AI might be able to reason
5. Critics of AI often invoke the “stochastic parrot” framing
6. Therefore, the “stochastic parrot” framing is wrong
7. (Bonus round:) Therefore, AI critics are wrong in general
8. (Bonus round #2:) Furthermore, parrots are intelligent and this framing is insulting to parrots

Let me first jump to the claim that’s most painful for me, speaking as a technologist and as an author on the Stochastic Parrots 🦜 paper: No, “Artificial Intelligence” is not a stochastic parrot. Large language models (LLMs) are. And large language models can be extremely useful.

Yet increasingly, “AI” systems of all stripes — from agents on Moltbook to commercial products like ChatGPT or Claude — are referred to as “stochastic parrots”, particularly as critics of the stochastic parrots framing argue their case. This is a result of a category error, treating one type of technology, large language models, as synonymous with AI. But there is a vast array of technology called “AI” that is not reducible to LLMs, and many current AI systems that utilize LLMs also leverage a variety of other technologies, including hand-written-rules, deterministic (non-stochastic) programs, various algorithms, and non-language models. This means that “AI”, broadly, is not equivalent to a large language model; it is not “just a stochastic parrot”.

Stochastic Parroting 🦜 is Technically Amazing

But setting aside this confusion: Stochasticity combined with parroting is remarkably powerful, a point that seems to be lost on many who push against the stochastic parrots framing. Different forms of stochastic parroting are fundamental to LLM training and generation, making it possible to effectively process text and to produce fluent, human-like linguistic output. It’s the product of decades of work, and a massively impressive computational feat.

To briefly explain the technology behind the metaphor, LLM training relies on stochasticity — in the form of stochastic gradient descent — to build statistical representations of text-based language. “Parroting” is central to how LLMs learn: Given a bunch of text, the model follows each piece in sequence, and is rewarded for correctly predicting the next piece [1]. The result is a model where each piece of text is associated with numeric information about the sequences it tends to occur in.

This means that LLMs are not designed for verbatim playback of text [2]. Instead, when prompted to generate, they draw on their learned representations to parrot smaller spans of text based on whether they’re a probable continuation of what came before — that’s yet another form of stochasticity. The end result is text that looks human-written because that’s exactly what LLMs have been exposed to.

One of the conceptual keys here is that LLM training works by consuming text-based descriptions of peoples’ thoughts, intents, and experiences; they do not participate in these experiences themselves, but can produce descriptions of them as if they were there, and say the kinds of things we say about them. An LLM can tell you about how much it loves camping if you’d like, even though it has never gone camping. It uses our deep thoughts about our existence and reflects them back to us: “I think,” it might say, and then follow through with the corresponding implicature, “ ‘I’ am”.

Parrots at the Singapore Bird Paradise. CC BY 4.0 / Margaret Mitchell / m-mitchell.com

Why They’re So Compelling

Together, stochasticity and parroting of virtually everything people have written on the internet (and beyond) provide fodder for incredibly compelling AI chat systems [3]. They are deceptively human-like because the LLMs that they use have incorporated troves of human interaction and expression. They give the sense of “understanding” like people, and “thinking” like people. In other words, it can seem that LLMs are not stochastic parrots because they are stochastic parrots.

The increasing tendency to lose track of the fact that “stochastic parrots” was penned to refer to LLMs specifically, and not to all of “AI”, makes a ton of sense: LLMs deal with language. Language is an entry point to interact with many AI systems, and language holds a privileged status in our brains. It has played a central role in interpersonal understanding and human connection, aiding in the development of societies that provide for our survival [4].

It’s worth reflecting on how currently within AI research, there are two powerful generative technologies whose use has skyrocketed — language generation and image generation (audio and video are quickly catching up) — yet it is the singular generative paradigm of language that we hold as indicative of reasoning and intelligence. This follows a tradition of associating intelligence with language (see work from Ludwig Wittgenstein or Jerry Fodor; more recently, see work from Temple Grandin and Amanda Baggs). The historic shift from describing people as “deaf and dumb” to “deaf-mute” to simply “deaf” was the product of a long prejudicial history of conflating spoken language with intelligence [5]; today, AAC (augmentative and alternative communication) users frequently report being treated as intellectually disabled because they don’t speak verbally. Just as the absence of spoken language was wrongly taken as evidence of absent intelligence in the past, the presence of fluent language output is now being wrongly taken as evidence of present intelligence. It’s a similar error, inverted.

Two vibrant Sun Parakeets. Image by Melanie van Zanten from Pixabay.

When Metaphors Stop Being Metaphors

But modern chatbots do something to get from a user’s language-based input to a linguistically well-formed output (response). The processes within that arc are increasingly referred to as “reasoning”, a word that carries connotations of human cognition, even though computationally-grounded terminology such as “goal-based processing” might be more appropriate. Naming conventions often draw from human experience: Biology research gives us “hitchhiking” genes, graphical modeling work gives us “belief” propagation. The etymology of “invasive” plants and parasite “hosts” documents how the earlier meanings of these terms were expanded as we applied them outside of human social life. Etc.

Such metaphors can be genuinely useful: They light up new ways of understanding phenomena, and provide high-level intuitions about scientific mechanisms that are otherwise unnamed. They help us appeal to things that people are deeply familiar with, in order to “give a sense” of something they’re not familiar with. Yet as these terms are absorbed into technical jargon, the work of rigorously interrogating such metaphors — testing their applicability, elaborating on them, discharging them, adjusting them, etc. — is rarely done [6].

In the case of artificial intelligence, this is becoming a bit of a problem. The 1950s metaphors of “intelligence” and machine “learning” have proven especially productive [7], catalyzing mentalistic terminology such as natural language “understanding”, neural network “attention” mechanisms, and more recently, “chain-of-thought”. This has led to the moment we now find ourselves in, where the distinction between literal and metaphorical language is collapsing in AI discourse. Metaphors are beginning to double back on themselves: What are AI systems doing before they produce an answer? Reasoning. How do we know? Because we call what they do reasoning.

This circularity has muddied peoples’ ability to understand what different AI systems are doing, demonstrably misleading people into conceiving of AI processing as roughly equivalent to human reasoning. It manifests the “jingle” part of the jingle jangle fallacy and pushes us to see the systems that we already anthropomorphize as being all the more human.

Why Call Them Stochastic Parrots?

The stochastic parrot metaphor is useful to pinpoint these anthropomorphic traps and disrupt them. As we discuss in the paper, when we interact with the output of a system and experience something human-like, we move to assumptions about the underlying mechanisms and nature of the system producing it, where the relevant concepts in our human experience — thought, understanding — were all developed to describe other kinds of things. Applying them to a fundamentally different kind of system is both an entirely natural human impulse, and a serious obstacle to understanding what is actually going on. These terms do not illuminate AI systems so much as they project onto them, and what gets lost in that projection is clarity on what these systems are.

Does this mean that an AI chatbot based on an LLM can’t be useful? No. Even “talking out” an idea can be helpful, although a generative system’s responses can be problematic in a few different ways (it can persuade people who are vulnerable, it may be incorrect, it reinforces hegemonic viewpoints, it relies on unconsented and uncompensated data — you should definitely check out our Stochastic Parrots 🦜 paper, we get into a lot of that). And when augmented with non-generative technology (including non-generative uses of LLMs), their utility expands.

The Warning

The idea that the stochastic parrot concept is wrong because LLMs can be useful, or because LLMs can do cool things, misunderstands what the term refers to. It provides a warning about losing track of the relationship between input data and output response, leading to a particular kind of inferential error — one with a long and damaging history — in which the outputs of a system are mistaken for evidence of its inner nature. LLMs produce fluent language. Fluent language is what humans associate with intelligence. Therefore, the argument goes, LLMs are intelligent. This is an error in logic that the framing was designed to name.

If you feel that something being a stochastic parrot is derogatory, it’s worth reflecting on why that is. Being a stochastic parrot is very cool.

Green Amazon parrot with emoji sunglasses. CC BY 4.0 / Philip Nalangan / Wikimedia

The Solution

To alleviate misunderstandings and the problems of talking-past-each other in discussions on AI’s internal processing, it’s useful to be precise about which systems we’re discussing when we say “AI”, and what they specifically do. It’s helpful to reflect on our metaphors rather than letting them calcify into assumptions. Resist the pressure (and there is enormous financial and political pressure) to treat fluent outputs as indicative of human-like understanding, and try to see AI systems for what they are.

This is what the Stochastic Parrots paper called for in 2021. It’s all the more relevant today.

Stickers made by Ruth Starkman’s class at Stanford.

Acknowledgements

Thanks to Emily Bender, Alex Hanna, and Timnit Gebru for comments and suggestions.

End Notes

[1] The term “reward” is a simplification of a more complex process. There is also something more directly called a “reward” in the related process of reinforcement learning.

[2] Although LLMs can be used for verbatim parroting:

• Jing Huang, Diyi Yang, and Christopher Potts (2024). Demystifying Verbatim Memorization in Large Language Models. https://aclanthology.org/2024.emnlp-main.598/
• Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo, Percy Liang (2026). Extracting books from production language models. https://arxiv.org/abs/2601.02671. (Provides further links to peer-reviewed articles.)

[3] Although LLMs can provide the content of conversations, they are not very good at “chatting” out-of-the-box (right after they’ve been trained) — in keeping with the concept of a stochastic parrot. A lot of additional work is required for interactive dialogue.

[4] As-stated, this statement is not controversial (as far as I know), but it is contestable. Further details are provided in literature on the evolution of language, among them:

• Robin Dunbar (1996). Grooming, Gossip, and the Evolution of Language.
• Michael Tomasello (2019). Becoming Human.

[5] See, e.g.,:

• FAQ from the (U.S.) National Association of the Deaf: (https://www.nad.org/resources/american-sign-language/community-and-culture-frequently-asked-questions/
• Jack Gannon (1980). Deaf Heritage: A Narrative History of Deaf America. https://gupress.gallaudet.edu/Books/D/Deaf-Heritage

[6] A point very well-made in:

• Susanne Knudsen (2005). Communicating novel and conventional scientific metaphors: a study of the development of the metaphor of genetic code. Public Understanding of Science, 14(4), pp.373–392. ff10.1177/0963662505056613ff. ffhal-00571070f. https://hal.science/hal-00571070v1/document

Examples of work engaging with this includes:

• Inie, N., Zukerman, P., & Bender, E. M. (2026). De-anthropomorphizing “AI”: From wishful mnemonics to accurate nomenclature. First Monday, 31(2). https://doi.org/10.5210/fm.v31i2.14366.
• Stark, L. & Hoffmann, A. (2019). Data Is the New What? Popular Metaphors & Professional Ethics in Emerging Data Culture, Journal of Cultural Analytics 4(1). https://doi.org/10.22148/16.036.

[7] In the neutral/non-positive sense. Ruha Benajmin has a piece on how “racism” is productive: https://today.emerson.edu/2019/10/18/ruha-benjamin-how-race-and-technology-shape-each-other/