Ask a modern AI system to explain quantum mechanics, debug a Python program, analyze an image, summarize a scientific paper, interpret a clinical vignette or operate a computer interface, and it is tempting to conclude that something resembling understanding must be happening. That impression is precisely what the debate about stochastic parrots and large language models calls into question.
There is another description, though, and it is much less flattering. Perhaps the system is simply a stochastic parrot: an extraordinarily sophisticated machine that has learned how human language tends to look, and can therefore produce convincing sentences without understanding what any of them mean.
The expression has become one of the most famous criticisms of large language models. It is also one of the most oversimplified. Its cleverness is not the issue. What is worth asking today is whether it still adequately describes what modern AI systems actually do. To answer that, we need to go back to where the idea came from.
The birth of the stochastic parrot
The term became famous through a 2021 paper, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, written by Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell.
It appeared at a time when increasingly large language models such as GPT-2 and GPT-3 were showing remarkable abilities. Its authors asked whether the race toward ever larger models was overlooking important scientific and social costs. Their concerns included the environmental and financial cost of training huge models and the biases and problematic content contained in enormous internet-derived datasets. They also pointed to the concentration of AI development in organizations with access to immense computational resources. And they warned about the risks that arise when fluent machine-generated language is mistaken for genuine understanding.
The “stochastic parrot” was a memorable metaphor for that last problem. A parrot can reproduce human sounds without necessarily understanding what they mean. “Stochastic” adds the probabilistic component: a language model learns statistical relationships between pieces of language and generates new sequences according to those learned relationships.
Put simply, during training the model solves, over and over, a problem that looks something like this:
Given everything that has appeared so far, what is likely to come next?
If the sentence begins:
“The capital of France is…”
then Paris is an extremely probable continuation.
Now scale the process up to billions or trillions of words, very large neural networks and immense amounts of computation. Something remarkable happens. The machine becomes extremely good at producing language that sounds as though it were written by someone who knows what they are talking about.
Sounding knowledgeable and being knowledgeable, however, are not necessarily the same thing. That distinction lies at the heart of the stochastic-parrot argument.
Form is not necessarily meaning
The intellectual roots of the argument go back further than 2021. In 2020, Emily Bender and Alexander Koller published Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data, built around a distinction between linguistic form and meaning.
A language model trained only on text receives linguistic forms, that is, sequences of symbols produced by humans. Human language, on the other hand, normally refers to things outside language: objects, intentions, sensations, actions and situations in the world. Bender and Koller therefore argued that a system trained exclusively on form cannot, from form alone, acquire meaning in the full human sense of the term.
To make the point, they proposed a now-famous thought experiment involving an extraordinarily intelligent octopus.
Two people are stranded on separate islands and communicate through an underwater cable. An octopus intercepts the cable and spends years observing their messages. Eventually it becomes so good at recognizing linguistic patterns that it can generate replies indistinguishable from those of the humans. Then one of them sends a message describing an emergency involving a bear.
The octopus has never seen a bear. It has never experienced danger, and it does not know what running, injury or fear actually feel like. All it knows is how messages containing words such as bear, danger and run relate statistically to other messages. The question is simple:
Does the octopus understand the conversation, or has it merely learned its structure?
For Bender and Koller, the distinction matters enormously. Human language is grounded in a world, and words are connected to perception, action, social interaction, goals and bodily experience. A text-only language model, by contrast, encounters descriptions of those things. For us, the word fire is bound up with light, heat, danger, memories and physical consequences. For a purely textual model, fire begins as something closer to a position in a vast network of relationships between symbols.
Why medicine makes the problem impossible to ignore
Medicine shows with unusual clarity why this distinction matters. Suppose an AI writes:
“This pattern is highly suggestive of acute aortic dissection and urgent CT angiography is indicated.”
The sentence may be fluent, clinically plausible and, stylistically, indistinguishable from something written by a physician. Yet none of this, by itself, tells us whether the conclusion is correct for the patient in front of us.
A model can produce the linguistic form of clinical expertise without possessing the biological experience, the bedside perception or the responsibility of a clinician. It has not seen the patient’s skin colour change or felt a pulse. It has not noticed that the patient looks different from the previous hour, nor experienced the consequences of a delayed diagnosis.
Medical language is also particularly dangerous when plausibility is mistaken for truth. A fabricated drug interaction, an incorrect staging rule or an invented reference can still sound entirely professional.
Recent clinical research confirms that the concern is not merely theoretical. Studies evaluating LLMs in realistic medical decision-making have found failures in diagnosis, guideline adherence and laboratory interpretation, as well as sensitivity to the way clinical information is presented. Other studies have documented hallucinated or propagated errors during the extraction of structured information from medical records.
In medicine, then, the stochastic-parrot problem becomes a practical question:
When the cost of an eloquent mistake is high, how much epistemic authority should fluency be allowed to carry?
To my mind, this remains one of the strongest arguments for keeping the original warning alive.
Linguistic plausibility is not truth
A traditional language model is fundamentally optimized to predict language well. It is not optimized to maximize
“the probability that this statement is true.”
That helps explain hallucinations. A fabricated scientific reference can have exactly the linguistic shape of a real one. A nonexistent historical event can sound perfectly plausible, and a medical explanation can be elegant and coherent while still containing a factual error.
Fluency, in other words, sets an epistemic trap. Humans naturally associate articulate language with intelligence and expertise. A system capable of generating extraordinarily fluent language can exploit that psychological association without intending to. The result is something historically unusual:
linguistic authority without guaranteed epistemic authority.
In medicine the distinction is sharper still. A clinician may be tempted to trust an answer because it resembles the style of a guideline, a consultation or a specialist report. Resembling an expert answer, however, is no evidence that the underlying reasoning, evidence or references are valid.
But something unexpected happened
Language models were already impressive when the stochastic-parrot argument was formulated. What followed was considerably more surprising.
As models grew larger and training methods improved, they did not simply get better at completing sentences. They became capable of translating between languages, writing and debugging software, solving previously unseen problems, manipulating structured information, analyzing images, interpreting charts and diagrams, using external tools and carrying out ever longer multi-step tasks. More recent systems add stronger reasoning procedures, multimodal perception and agent-like interaction with software and other external tools.
This created a problem for the simplest reading of the stochastic-parrot metaphor. A parrot imitates. Modern language models often generalize, producing solutions that were almost certainly never present verbatim in their training data.
So the question changed. Instead of asking
“Are language models memorizing text?”
more and more researchers began to ask:
“What internal representations must a model develop in order to predict text this well?”
It is a far more interesting question.
Prediction may require a model of the world
Take a simple example. Suppose a model repeatedly encounters statements about Rome, Milan, Naples and Turin. To predict that language successfully, it may gradually become useful for the model to represent relationships such as:
- cities have locations;
- Rome is south of Milan;
- Italy has a geographical structure;
- distances between locations are not arbitrary.
Nobody explicitly instructed the model to build a map. An internal approximation of geography may simply help it predict text.
Research has found evidence that language models develop internal representations related to spatial and temporal relationships. Gurnee and Tegmark, for example, reported linear representations of space and time in Llama-2 models, including neurons whose activity correlated with spatial or temporal coordinates.
Work on Othello-playing transformer models offers a particularly intuitive example. A model trained simply to predict legal sequences of moves can develop an internal representation related to the state of the board. Further work showed that such representations can play a causal role in the model’s subsequent behaviour.
Why does this matter? Because it challenges the simplest version of the parrot analogy. The model may be trained through statistical prediction, but successful statistical prediction can force it to discover latent structure. Put differently, predicting the next token does not imply thinking only about the next token. The training objective and the internal solution learned by the network are two different things.
“It only predicts the next word” is therefore misleading
One of the most common descriptions of LLMs is that “they are just next-word predictors”. Technically, there is an important truth in this: prediction is indeed fundamental to how language models are trained. But the word just carries far more philosophical weight than the rest of the sentence.
Imagine training a machine to predict what happens next in thousands of chess games. To become really good at prediction, the easiest solution may eventually involve learning something about pieces, positions, legal moves, threats and strategy. In the same way, predicting human language may reward internal representations of objects, people, intentions, causality, space, time and mathematical structures. Language contains an enormous amount of compressed information about the world, because it was produced by creatures living in that world.
Which points to an important counterargument to the stochastic-parrot thesis: statistical learning and representation of reality are not mutually exclusive. A system can learn through statistics and still discover structure. Human learning itself has large statistical components (children extract patterns from speech, and biological brains continuously generate predictions about sensory input). Calling something “statistical” does not, by itself, make it superficial.
But does representation equal understanding?
This is where it gets difficult.
Finding internal representations of geography, time or board states does not demonstrate that an AI understands those things in the same way a human does. Several concepts are frequently collapsed into one another: competence, knowledge, representation, reasoning, understanding, intelligence and consciousness. They are not equivalent.
A system may manipulate a representation successfully without possessing subjective experience. It may solve a problem without knowing that it is solving one, or behave as though it understands fear without ever experiencing fear. Nothing in the discovery of increasingly sophisticated internal representations demonstrates consciousness, nor does it solve the philosophical problem of meaning.
So the stochastic-parrot argument survives, in a subtler form. The original binary question (“Does the model understand or merely predict?”) may itself be too crude. The more interesting possibility is that understanding is not binary. Machines may develop forms of functional representation and reasoning that deserve to be distinguished both from simple imitation and from human understanding.
Medicine again: representation is not clinical understanding
In medicine the distinction is particularly important.
A model may learn a rich representation connecting aortic stenosis, valve area, gradient, left ventricular hypertrophy and symptoms, and that representation can be clinically useful. Clinical understanding, though, involves more than correctly associating concepts. A physician integrates measurements with uncertainty, examination, trajectory, patient preferences, comorbidities and consequences. Any cardiac surgeon will recognize this. The same echocardiographic result can carry very different meaning in an asymptomatic athlete, a frail elderly patient and a patient in cardiogenic shock.
So even if an LLM develops something that deserves to be called a medical world model, the relevant question is not simply
“Does it know cardiology?”
but rather
“What aspects of clinical reality are represented, which are missing, and under what conditions does the representation fail?”
That is a more useful question, both scientifically and clinically.
The grounding problem is changing
The original debate also assumed something that is becoming less and less true: that language models interact only with language. Modern frontier systems increasingly process multiple kinds of information, combining text with images, audio, video, structured data and outputs from external tools.
This changes the grounding problem. A text-only model encountering the word cat sees the linguistic contexts in which cats appear. A multimodal model may encounter the word together with many visual representations of cats, and can learn associations between language and visual properties.
In medicine the difference is even more tangible. A multimodal medical system can potentially connect the phrase pulmonary edema not only with textual descriptions, but with chest radiographs, ultrasound images, physiological signals and structured laboratory data. Recent surveys of multimodal medical LLMs describe applications spanning medical reporting, diagnosis and treatment support across multiple data modalities. That is already different from purely linguistic grounding.
It is still not equivalent to human experience. A clinician does not merely process an image of pulmonary edema. Breathlessness, respiratory effort, auscultation, blood pressure, urine output, prior imaging, response to therapy and the changing appearance of the patient over time all enter the clinician’s judgement. Human grounding is embodied, causal and interactive.
Multimodal AI therefore weakens the claim that modern models operate only on linguistic form. It does not, however, eliminate the larger question of how symbols and representations acquire meaning.
Reasoning changes the picture, but does not settle it
Another major development has been the rise of models optimized to devote more computation to difficult problems. Instead of producing an answer immediately, these systems can perform extended internal computation, explore alternatives, use tools and verify intermediate results.
The appeal for medicine is obvious. Diagnostic reasoning is rarely a single association between symptoms and a disease. It involves differential diagnosis, updating probabilities, deciding what information is missing and choosing the next useful test. Recent work specifically examining reasoning-driven LLMs in medicine reflects this shift from pure text generation toward systems designed for multi-step problem solving.
None of this should be mystified. There is no logical reason why a machine performing sophisticated computation must possess subjective awareness. Nor, on the other hand, should its behaviour be dismissed as simple regurgitation.
A useful distinction here is that the training mechanism is not the same thing as the learned capability. Artificial neural networks learn through mathematical optimization, human brains through biological mechanisms. In neither case does describing the learning mechanism automatically provide a complete description of the resulting abilities.
From language model to tool-using system
A further development makes the parrot metaphor look ever more incomplete. Modern AI systems are no longer necessarily confined to generating text: they can be connected to tools. Depending on its environment, an AI system can now:
- formulate a hypothesis;
- search for information;
- open a file;
- execute code;
- inspect the result;
- interact with software;
- discover that its original assumption was wrong;
- modify its strategy;
- try again.
What matters conceptually is not any single commercial product. It is the architecture. The classical language model resembles:
input → generated text
An agentic system increasingly resembles:
goal → reasoning → action → observation → correction → action
The system has entered a feedback loop with an environment.
In medicine, this could eventually mean a model that does not merely suggest that a patient’s renal function should be checked. It could retrieve the latest creatinine, calculate the trend, compare it with previous values, check the relevant drug doses and return a structured recommendation for clinician review. That is still not autonomous medical understanding, and it raises major safety and governance questions. It is, however, clearly more than the production of plausible sentences. The “parrot” can now act, inspect the consequences of its actions and adjust its behaviour.
What the stochastic-parrot argument still gets right
More than five years after the paper was published, several of its warnings have aged remarkably well.
Fluency is not truth
This is perhaps still the most important lesson. A beautifully written answer can be false, and the better language models become at sounding authoritative, the more independent verification matters. In healthcare the principle should be stronger still: polished language must never be treated as a substitute for validated evidence, clinical context or professional accountability.
Human beings anthropomorphize language
We instinctively attribute minds to entities that communicate fluently, which is why words such as knows, believes, wants and understands call for care. Sometimes they are useful functional shorthand. They should not silently become claims about consciousness.
Training data shape behaviour
Large-scale training does not magically remove the biases, distortions and cultural asymmetries present in the information from which models learn. In medicine this matters especially, because unequal representation of populations, healthcare systems and languages can translate into unequal model performance.
Scale has costs
Large models require considerable computational and economic resources. The original paper’s concerns (who has the resources to build frontier systems, who controls them, how their costs are distributed) remain highly relevant.
Linguistic performance is not sufficient evidence of human-like understanding
A system that performs well in conversation has demonstrated conversational competence. Anything beyond that requires additional evidence. As scientific principles go, it is still a valuable one.
What now looks less convincing
The strongest interpretation of the stochastic-parrot metaphor is much harder to defend. If by “stochastic parrot” we mean
“a probabilistic system trained from human-generated data,”
then the description is technically reasonable, but not very informative. If instead we mean
“a system that merely memorizes and recombines fragments of its training data without developing internal representations,”
then the evidence accumulated since 2021 makes the claim harder and harder to sustain.
Language models clearly do more than reproduce stored sentences. They generalize and manipulate abstractions. They sometimes develop discoverable internal representations of structures that were not explicitly supplied during training, and they can combine information in novel ways. Embedded in larger systems, they can also interact with external environments and modify their behaviour according to feedback.
None of this proves human-like understanding. It does mean, though, that “mere imitation” is no longer an adequate scientific description either.
Perhaps we are asking the wrong question
The debate is frequently framed as a yes-or-no question: do language models understand? Perhaps that framing assumes too much. Human cognition is not a single ability; we distinguish perception, memory, reasoning, language, planning, self-awareness and consciousness. There is little reason to assume that artificial intelligence must either possess all of these in human form or possess none of them.
A more productive framework might ask:
- What kinds of internal representations does the system construct?
- Can it generalize beyond familiar patterns?
- Can it distinguish appearance from reality?
- Can it use causal relationships?
- Can it revise conclusions when confronted with contradictory evidence?
- Can it learn from interaction with an environment?
- Can it maintain goals over time?
- Can it recognize its own uncertainty?
- Which abilities require language alone, and which require interaction with the world?
For medicine, the list can be made even more concrete:
- Does the system know when information is missing?
- Does it distinguish guideline knowledge from patient-specific judgement?
- Can it calibrate uncertainty rather than merely sound confident?
- Does it recognize when a case lies outside the distribution on which it performs well?
- Can a clinician inspect the evidence on which an answer depends?
- Does human oversight genuinely detect the model’s failures?
These questions can be tested experimentally. “Does it really understand?”, on the other hand, may ultimately remain partly philosophical.
The parrot has left the cage
The history of the metaphor contains an irony. The original stochastic parrot lived entirely inside language: it consumed text and produced text.
Today’s AI systems are beginning to leave that cage. They can perceive images and process audio and structured data. They retrieve external information, run software and interact with digital environments, observing the consequences of their actions and changing what they do next. Memory and persistent agency add another dimension, with systems whose present behaviour can depend not only on their training and current prompt but also on accumulated interaction history. At that point, the conceptual object we are trying to describe has moved considerably beyond the language models of 2021.
Medicine offers a useful warning against both extremes. The impressive ability of modern systems to manipulate medical knowledge should prevent us from dismissing them as trivial pattern-matching machines. Their capacity to produce persuasive errors, at the same time, should prevent us from equating performance with understanding.
Perhaps, then, the most useful conclusion is neither “the stochastic parrot argument was right” nor “the stochastic parrot argument was wrong”. It is that the metaphor identified a genuine danger, confusing fluent language with understanding, while becoming less and less adequate as a complete theory of modern AI.
The paradox is worth preserving. Language models are statistical machines, yet statistical machines can learn internal structure. They predict tokens, and predicting tokens may require learning something about the world that produced those tokens. The world is not something they experience as we do; increasingly, though, they can perceive and interact with parts of it. They can be spectacularly competent and spectacularly wrong.
Perhaps that tension is precisely what makes the philosophical question so interesting.
The real challenge is no longer to decide whether the machine is a parrot. It is to understand what kind of thing a stochastic system can become when prediction, representation, reasoning, perception, memory and action begin to converge.
References and Further Reading
- Bender EM, Gebru T, McMillan-Major A, Mitchell M. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21). 2021:610–623. https://doi.org/10.1145/3442188.3445922
- Bender EM, Koller A. Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020:5185–5198. https://doi.org/10.18653/v1/2020.acl-main.463
- Gurnee W, Tegmark M. Language Models Represent Space and Time. International Conference on Learning Representations (ICLR). 2024. Official publication
- Li K, Hopkins AK, Bau D, Viégas F, Pfister H, Wattenberg M. Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task. International Conference on Learning Representations (ICLR). 2023. OpenReview
- Hazineh DS, Zhang Z, Chiu J. Linear Latent World Models in Simple Transformers: A Case Study on Othello-GPT. arXiv:2310.07582. 2023. arXiv
- Schulze Buschoff LM, Akata E, Bethge M, et al. Visual cognition in multimodal large language models. Nature Machine Intelligence. 2025;7:96–106. https://doi.org/10.1038/s42256-024-00963-y
- Chowa SS, Alvi R, Rahman SS, et al. From language to action: a review of large language models as autonomous agents and tool users. Artificial Intelligence Review. 2026;59:71. https://doi.org/10.1007/s10462-025-11471-9
- Hager P, Jungmann F, Holland R, et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature Medicine. 2024;30:2613–2622. https://doi.org/10.1038/s41591-024-03097-1
- Kwong JCC, Wang SCY, Nickel GC, et al. The long but necessary road to responsible use of large language models in healthcare research. npj Digital Medicine. 2024;7:177. https://doi.org/10.1038/s41746-024-01180-y
- Gilbert S, Kather JN, Hogan A. Augmented non-hallucinating large language models as medical information curators. npj Digital Medicine. 2024;7:100. https://doi.org/10.1038/s41746-024-01081-0
- Ye J, Tang H. Multimodal Large Language Models for Medicine: A Comprehensive Survey. arXiv:2504.21051. 2025. arXiv
- Savage T, Nayak A, Gallo R, et al. Diagnostic reasoning prompts reveal the potential for large language model interpretability in medicine. npj Digital Medicine. 2024;7:20. https://doi.org/10.1038/s41746-024-01010-1
- Wang X, Xiong Z, Zou K, et al. Reasoning-driven large language models in medicine: opportunities, challenges, and the road ahead. The Lancet Digital Health. 2026;8:100931. https://doi.org/10.1016/j.landig.2025.100931
Editorial note
Throughout this article, the term understanding is used with deliberate caution. Evidence that a model builds useful internal representations, reasons across multiple steps or interacts effectively with tools does not, on its own, establish consciousness, subjective experience or human-equivalent semantic understanding.