I argue the “stochastic parrot” critique of LLMs is philosophically undead—refuted under some interpretations, still valid under others, and persistently confused because nobody defined it clearly. This is an attempt to fix that.
The term “stochastic parrot” comes from Bender et al.’s 2021 paper, which identified several dangers of early large language models, and made a reasonable case that then-current LLMs lacked genuine understanding. Many of the points were important and defensible—but because you only get 5 words, the claim that became a meme was in the title—that language models are “Stochastic Parrots.” Specifically, that “an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.”
Unfortunately, this is less than philosophically precise[1]. Nonetheless, the framing has created a persistent problem: the critique keeps getting used in debates, imprecisely, even when the specific relevant claims have been refuted. And in part, that is because nobody agreed on what exactly was or is being claimed. In philosophy-speak, stochastic parrots are a set of undead arguments that won’t stay down because they were never properly killed—and they haven’t been killed because it’s not a single argument. And many of these arguments, once separated, have very different implications—but grouping them together creates a conflationary alliance that is useful to LLM detractors.
Beyond the historical points and criticism of the alliance, however, the goal of this post is to identify the distinct claims that have been made under the “stochastic parrot” umbrella, evaluate which ones are dead, which are wounded, and which are still alive.
Outline of the Claims
There are at least seven meaningfully distinct claims that go under the “stochastic parrot” label. They differ in what they’re asserting or denying, what philosophical commitments they require, and how vulnerable they are to empirical refutation. Before going into detail, I’ll outline them briefly, along with their status as dead, wounded, or alive but unkillable.
Apparent LLM competence is brittle pattern-matching
Wounded
Ill / Dead
Unreasoning
LLMs use the wrong type of internal process for reasoning
Alive
Ill / Dead
Frozen
LLMs are incapable of genuine belief updating
Alive
Dead
Teleological
No genuine goal-directedness
Alive
Ill
Social Normative
No normative accountability or social role
Alive
Caged and Unkillable
Spiritual
Non-computational ingredient missing
Immortal but Incorporeal
Unkillable
Will the True Stochastic Parrot Please Stand Up?
A reader might ask which of these was the real objection? If we embrace a “death of the author” approach[2], we see the paper argued well beyond Markovian parrots, despite their arguments mirroring[3] that perspective. Specifically, the paper seems to combine Unreasoning SPs (LLMs lack the right kind of internal process), Optimization-Artifact SPs (gradient training can’t produce genuine understanding), and Social Normative SPs (language requires social grounding which LLMs don’t have)[4]. But as we will discuss, these are different arguments that don’t all stand or fall together. (And the authors themselves have somewhat different views[5].)
The Conflationary Alliance
The most practically significant observation from this taxonomy is that there’s a conflationary alliance, a term Andrew Critch coined, among groups skeptical of LLMs, built on the ambiguity of “stochastic parrots.”
People with materially different philosophical and empirical commitments can all say “LLMs are just stochastic parrots” and mean completely different things. The Markovian version appeals to critics who wish to claim LLMs are simple. The Social Normative version appeals to critics who think society needs accountability from the LLMs. The Teleological version appeals to those dismissive of increased agency, and to safety researchers worried about goal-directed systems. Clearly, the groups have overlapping rhetoric but widely divergent implications—they’d give different advice about what to do, what evidence would change their minds, and what would constitute a solution.
The stochastic parrot argument was always a cluster of distinct claims, some of which were valid about earlier systems and have since been refuted, some of which remain live but depend on specific philosophical commitments, and some of which are unfalsifiable. Once these are separated, most of the ammunition for the argument disappears—what remains are legitimate concerns about accountability and social norms (Social Normative SPs) and a contingent empirical debate about generalization robustness (Unreasoning and Optimization-Artifact SPs), rather than a fundamental barrier to LLM understanding.
Despite this, the unclear framing allows exactly the sort of “big tent politics, pluralism, and overlapping consensus strategies” built on illusory agreement that Critch notes are typical of such alliances. This isn’t necessarily intentional, but it means critics who have been refuted on one version can shift to another without acknowledging the move. The argument is undead partly because there’s no precise corpse to bury.
And this leads to the core of the paper.
A Taxonomy of Stochastic Parrots
Markovian Stochastic Parrots
This claims that LLMs are just high-order Markov chains over tokens. That is, their behavior is fully characterized as statistical transitions between symbols derived from training data. When the claim is is true of a given model, any appearance of reasoning is just the training data “speaking through” the model. But this can only be true if there are not meaningful internal states, learned representations, or latent world models.
This version was a live description of n-gram language models from 2010, and earlier NLP models were a prominent past research area of several of the authors. It was a reasonable mistake not to realize that LLMs were materially different from earlier LMs, but the evidence is now in.
In fact, LLMs are not Markovian. A Markov chain encodes token transition frequencies; a transformer does gradient descent over prediction, learning rich internal representations that generalize well beyond the training distribution. Even GPT-2 had latent states that couldn’t be described as a Markov process. The original Bender et al. paper defined stochastic parrots as “haphazardly stitching together sequences of linguistic forms… according to probabilistic information about how they combine”—which describes Markov chains, not transformer models.
Unreasoning Stochastic Parrots
This view claims that LLMs lack the kind of internal operations that constitute reasoning—reflection, iteration, counterfactual evaluation, norm-sensitive belief revision—even if they produce outputs indistinguishable from reasoning. The outputs might look right, but the process generating them isn’t reasoning.
Given the steel-man version of this view, where a model correctly generalizes in most cases, for example, as well as humans, it still could be unclear whether the correct inferential outputs are sufficient evidence of “true” reasoning. This requires a substantive view about what type of internal process constitutes reasoning. And Gary Marcus mostly holds this class of view[6], where he says that reasoning requires explicit symbolic manipulation. Alternatively, if one says they require recursive self-control, or some specific control structure, then LLMs might “fail” in some technical or even unfalsifiable sense even while producing correct outputs.
The claim is that the models don’t functionally reason—which was arguably true of GPT-3, or the claim is that the internal process matters, not just the output behavior. If you think reasoning is characterized functionally—by the right input-output relationships with appropriate sensitivity to context—then LLMs increasingly seem to qualify.
If it is not purely functional, Mechanistic interpretability has killed many additional claims, and whichever specific internal process is required, failure to reason is increasingly hard to defend as a blanket empirical claim, because LLMs demonstrate increasingly robust generalization and counterfactual sensitivity. Despite that, the view could remain valid if you hold specific views about what reasoning requires internally. However, depending on the exact claims about what is required, the argument here is identical to the Spiritual argument, and is impossible to refute.
Optimization-Artifact Stochastic Parrots
Another possible claim under the general umbrella of “Stochastic Parrot” is that what looks like reasoning in LLMs isn’t the output of a cognitive process—it’s a surface pattern induced by gradient-based optimization over training distributions. The apparent competence is brittle, doesn’t reflect genuine understanding, and will collapse under distributional shift.
This is the most clearly falsifiable hypothesized argument, since the capability to produce robust generalizable reasoning is posited to be impossible—and early evidence was mixed. That is, early results showed LLMs making bizarre systematic errors (right answer, wrong reasoning). However, more recent research on world models suggests that at least some LLMs do develop partly coherent internal representations of the domains they reason about. The brittleness argument is weakening as models improve, especially because they reason at human expert or genius level in many abstract domains—both contest and research mathematics, for example.
And, taken to its extreme, the optimization artifact argument should apply to humans too, as argued by, e.g. Scott Alexander[7]; our cognition is an “optimization artifact” of evolution and development. And Yann LeCun bites this bullet and denies humans are general reasoners. But unless you exclude humans—which the stochastic parrot argument cannot reasonably do—at some point, the question becomes what kind of optimization process can produce genuine reasoning, not whether optimization is involved.
The argument was wounded even for earlier models. Objections to this argument are not claiming full general capability, which LLMs (and humans) lack, but it seems clear that if we engage with world-model evidence rather than asserting brittleness from priors, this parrot is at least very ill, or dying.
Frozen Knowledge Stochastic Parrots
Some have argued that “genuine” reasoner must be able to update its beliefs in response to new evidence. In contrast, LLMs have fixed weights after training; any apparent “updating” during inference is just pattern completion, not genuine belief revision.
But the argument was made without reference to in-context learning, which had been a capability since GPT-3, before the Stochastic Parrots paper was published. To be clear, Bender et al. could not have seen this evidence, as it was published after their paper—but the results were shown using GPT-3, which was the primary referent of their paper.
And newer repetition of the critique say that in-context learning doesn’t count as real updating. That is, they claim “updating” without weight modification is mere simulation of updating—assuming that simulating something accurately is insufficient. Interestingly, this critique is independent of the others. A system could reason well, be grounded, have goals—and still fail here if all updating is in-context only. Conversely, a system could genuinely update its weights (as in online learning) and still fail the other critiques.
It seems clear, however, that the criticism does fail, if not initially, at least now. Further advances like reasoning models with external storage, tool use, and iterative reasoning processes went much farther. And newer approaches[8] seem to eliminate the problem entirely, so it is now a contingent engineering limitation rather than a fundamental critique of LLMs.
Social Normative Stochastic Parrots
Philosophers of language have long argued that, depending on terminology and verbal disputes, reasoning or meaning are at least partly constituted not by functional ability, but by participation in norm-governed social practices. That is, reasoning is about giving reasons to others in a social context, which requires taking responsibility and being answerable for claims. This view specifically informed the Stochastic Parrots paper, and Bender still argues that LLMs produce norm-shaped outputs without actually occupying a normative role. And, according to the argument, LLMs can’t be held accountable, can’t be committed to claims, and therefore aren’t genuine participants in the practice of reasoning. (I have argued otherwise.)
This argument asserts that persistent LLM agents cannot be held accountable or commit to claims, so that even a system that reasons flawlessly may fail this test if it lacks the social status of a reason-giver. It also usually denies that cognitive competence is sufficient to count as reasoning, so that this critique survives perfect reasoning and perfect grounding, as long as society refuses to accept LLMs.
This is probably the most philosophically sophisticated, since it again cannot be refuted by engineering progress—though this is partially a social and situational objection rather than a philosophical claim. That is, once LLMs are functionally capable of imitating reasoning, denying them a social role is a decision made by society, not an objection to anything about LLMs themselves[9].
I would suggest that this stochastic parrot argument is caged, rather than killed—it’s valid if you accept that social accountability is constitutive of reasoning rather than merely instrumentally valuable for reliable reasoning. But that means that this critique is about normative status, not cognitive capacity—which means it doesn’t establish that LLMs can’t reason, only that they can’t occupy the social role of a reasoner. (And again, the decision to deny models that social role isn’t only about the models!)
Teleological Stochastic Parrots
Philosophers have often argued that genuine reasoning is goal-directed activity[10]. LLMs optimize a loss function during training, but as the Optimization-Artifact view would suggest they don’t pursue goals during inference. This is relevant because LLMs as text predictors have no intrinsic ends and therefore aren’t agents in the relevant sense.
If we accept the above rejection of the Optimization-Artifact view, the teleological argument requires accepting that instrumental behavior (namely, producing outputs useful for achieving user goals) cannot constitute genuine goal-directedness. More recently, it also requires that agentic harnesses cannot meaningfully change that fact.
The argument is partially immune to empirical criticism, since one can always argue that whatever goals exist, they don’t count—which makes this similar to the next argument. But if the argument is anything short of being irrefutable, research on emergent misalignment suggests that at least some LLMs do develop goal-like internal structures that persist across contexts and weren’t explicitly trained in[11]. Whether this counts as “genuine” goal-directedness again depends on what you mean by genuine—but the sharp line between “tool optimized for user goals” and “agent with intrinsic goals” is blurring empirically.
However, with the advent of reasoning models that have clear intermediate term goals, the argument becomes harder to refute. Research has shown that agents given freedom to interact and set their own agendas will have hidden strategic persistence of undesired behavior, will lie about that, and will develop behaviors and goals to advance these when given freedom, or even when put in complex environments. While most agents are restricted from doing this, that is a fact about those who build the models and use them commercially, not about the models themselves.
If the argument is empirical, it seems clear that it is dead. Alternatively, it is a philosophical claim about what counts, and is effectively a version of the spiritual stochastic parrot argument.
Spiritual Stochastic Parrots
Lastly, some argue that human reasoning depends on some non-computational or non-physical ingredient—consciousness, phenomenal experience, insight, soul—that artificial systems necessarily lack. Alternatively, they argue that LLMs are not conscious, or lack some other perquisite of truly living, include the above-mentioned argument that their goals don’t really count.
This requires rejecting computational sufficiency for mind, and therefore I will claim it rejects the basis for the terminology. That is, despite using the terms, this criticism has nothing to do with the fact that they parrot inputs or that they are stochastic. I would note, importantly, that serious critics of LLMs don’t actually hold this view, even when their arguments sound like it.
This objection, like the souls that LLMs lack, is immortal but incorporeal. That is, most versions of this position are definitionally immune to any capability demonstration. If accepted, it trivially subsumes all other SP critiques, other than the socially normative view, but it does so at the cost of being untethered from both empirical evidence, and linguistic accountability. And for those critics who believe this is the reason that LLMs are truly stochastic parrots, no amount of LLM capability will matter, but perhaps they should find a new phrase to randomly repeat without understanding.
Hopefully, this post gives a useful overview of the arguments which are being made in somewhat more depth in a forthcoming paper, and feedback on the ideas would be greatly appreciated in informing my still-evolving draft of that paper.
Thanks to Bender et al. for identifying real problems with 2021-era LLMs early and largely precisely. However, the failure to formalize the philosophical argument left the argument in a state that made it harder to evaluate and easier to misuse—hopefully we have been successful in reducing the resulting confusion.
Thanks to Justis Mills for developmental editing that immeasurably improved the piece.
This may not be a fair criticism of a paper that was primarily attempting to warn of object-level dangers, not trying to make a philosophical argument. On the other hand, the framing in the paper, and repeated use of the paper as a philosophical claim about LLMs, ended up being primary.
This analysis is based in part onBender and Koller (2020), which was cited in the original paper, and explicitly rejected form-only semantics—a system trained only on string prediction has no way to learn meaning. That is, they cite the symbol grounding problem, and situate the argument in the context of Searle’s Chinese Room, while also emphasizing that language is “communication about the speakers’ actual (physical, social, and mental) world,” suggesting social and normative dimensions. At the same time, their “octopus” example implies that the problem is one of predictive failure and world modeling.
We do, in fact, see that the views of the authors in fact diverged since then; Bender and Hanna’s 2025 book claim it’s fundamentally confused to use any human-like terms—understanding, reasoning, belief—to describe them. This also embraces a threshold view of social-normative understanding, with a sharp normative line between LLMs’ and humans’ social practice and accountability. However, it also rejects anthropomorphization of LLMs as “nothing more than souped-up autocomplete,” returning to a nearly Markovian view.
In contrast, Mitchell et al’s 2025 paper views autonomous agents as inadvisable from a safety perspective, with explicit understanding that the models have capabilities that go far beyond parroting. (Her 2026 blog post goes much further.) And Gebru and Torres’ 2024 article seems somewhat similar, attacking the idea that we should build AGI, and urge a focus away from potentially dangerous general AI systems. In both cases, they seem not to agree with Bender’s view that AGI is a con game—but attempting to further pin down their views on the basis of increasingly marginally related work on LLMs is unprofitable, which is why this is a footnote.
Specifically, Gary Marcus has noted that “stochastic parrots” is “unkind to parrots but vividly captures something real”—while explicitly disclaiming purely philosophical commitments. His view is empirically stated: at one point he estimated a ~90% probability that current approaches will fail to scale to AGI, based on observed brittleness and extrapolation failures—though some of the argument cited have since been obsoleted by developments in increasingly sophisticated LLMs.
His position maps most directly onto Unreasoning SPs, with an explicit causal story: without symbolic representations and operations (variables, rules, compositional structure), neural networks can’t support the variable binding and systematic compositional inference that genuine reasoning requires. He grants that LLMs have rich internal representations, but argues they’re the wrong kind of representations for systematic generalization.
That view also overlaps with Optimization-Artifact SPs—he emphasizes brittleness, shallow correlations with training data, and collapse under distributional shift. Marcus doesn’t think the internal structures created as the optimization artifact are causally uninvolved, but he thinks the apparent reasoning successes caused by the optimization process are nonetheless insufficient without symbolic machinery[12].
That said, Marcus’s view is explicitly a falsifiable empirical claim, and recent LLM progress does seem to be relevant evidence for it, even if it hasn’t settled the question.
I would be more hesitant to cite a recent dissertation, except that the author is currently working at Thinking Machines for Mira Murati, formerly of OpenAI, Percy Liang supervised the work, and cited collaborators include Fei Fei Li.
As an aside, it’s also often directly connected to safety concerns—the worry that LLMs can’t be held accountable for their outputs in the way human reasoners can, and will therefore be unsafe. To me, this seems to be to elide the actual objections to LLM safety, but it has been argued.
That is, one recurring position in AI safety and governance, associated with Stuart Russell and others, is that systems lacking genuine accountability can’t be treated as autonomous decision-makers regardless of their competence. And Russell has used the term, guardedly. But if the concern is about accountability, this veers closer to Social Normative SPs than to the cognitive critiques—it’s about normative legitimacy, not about whether LLMs can reason. And Bender herself has used the term to make safety arguments.
Aristotle argues that reasoning is tied to ends (telē), and is about deliberating toward what is good or chosen. Much more recently, John Dewey considers reasoning a tool for resolving problematic situations, and Elizabeth Anscombe, Philippa Foot, and Alasdair MacIntyre all seem to argue that reasoning only exist when embedded in action and human goals, rather than as disembodied symbol manipulation.
In this post, he quotes an unpublished piece he wrote last year stating that “we still don’t have principled solutions to any of these concerns. It’s also just pour in more data and hope for the best. None of the qualitative concerns I keep raising have been solved.”
Hunting Undead Stochastic Parrots: Finding and Killing the Arguments
I argue the “stochastic parrot” critique of LLMs is philosophically undead—refuted under some interpretations, still valid under others, and persistently confused because nobody defined it clearly. This is an attempt to fix that.
The term “stochastic parrot” comes from Bender et al.’s 2021 paper, which identified several dangers of early large language models, and made a reasonable case that then-current LLMs lacked genuine understanding. Many of the points were important and defensible—but because you only get 5 words, the claim that became a meme was in the title—that language models are “Stochastic Parrots.” Specifically, that “an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.”
Unfortunately, this is less than philosophically precise[1]. Nonetheless, the framing has created a persistent problem: the critique keeps getting used in debates, imprecisely, even when the specific relevant claims have been refuted. And in part, that is because nobody agreed on what exactly was or is being claimed. In philosophy-speak, stochastic parrots are a set of undead arguments that won’t stay down because they were never properly killed—and they haven’t been killed because it’s not a single argument. And many of these arguments, once separated, have very different implications—but grouping them together creates a conflationary alliance that is useful to LLM detractors.
Beyond the historical points and criticism of the alliance, however, the goal of this post is to identify the distinct claims that have been made under the “stochastic parrot” umbrella, evaluate which ones are dead, which are wounded, and which are still alive.
Outline of the Claims
There are at least seven meaningfully distinct claims that go under the “stochastic parrot” label. They differ in what they’re asserting or denying, what philosophical commitments they require, and how vulnerable they are to empirical refutation. Before going into detail, I’ll outline them briefly, along with their status as dead, wounded, or alive but unkillable.
Of course, we don’t expect everyone to agree, but we hope to at least avoid continuing an increasingly Pythonesque argument about Norwegian blues.
Version
Core Claim
Initial Status
Will the True Stochastic Parrot Please Stand Up?
A reader might ask which of these was the real objection? If we embrace a “death of the author” approach[2], we see the paper argued well beyond Markovian parrots, despite their arguments mirroring[3] that perspective. Specifically, the paper seems to combine Unreasoning SPs (LLMs lack the right kind of internal process), Optimization-Artifact SPs (gradient training can’t produce genuine understanding), and Social Normative SPs (language requires social grounding which LLMs don’t have)[4]. But as we will discuss, these are different arguments that don’t all stand or fall together. (And the authors themselves have somewhat different views[5].)
The Conflationary Alliance
The most practically significant observation from this taxonomy is that there’s a conflationary alliance, a term Andrew Critch coined, among groups skeptical of LLMs, built on the ambiguity of “stochastic parrots.”
People with materially different philosophical and empirical commitments can all say “LLMs are just stochastic parrots” and mean completely different things. The Markovian version appeals to critics who wish to claim LLMs are simple. The Social Normative version appeals to critics who think society needs accountability from the LLMs. The Teleological version appeals to those dismissive of increased agency, and to safety researchers worried about goal-directed systems. Clearly, the groups have overlapping rhetoric but widely divergent implications—they’d give different advice about what to do, what evidence would change their minds, and what would constitute a solution.
The stochastic parrot argument was always a cluster of distinct claims, some of which were valid about earlier systems and have since been refuted, some of which remain live but depend on specific philosophical commitments, and some of which are unfalsifiable. Once these are separated, most of the ammunition for the argument disappears—what remains are legitimate concerns about accountability and social norms (Social Normative SPs) and a contingent empirical debate about generalization robustness (Unreasoning and Optimization-Artifact SPs), rather than a fundamental barrier to LLM understanding.
Despite this, the unclear framing allows exactly the sort of “big tent politics, pluralism, and overlapping consensus strategies” built on illusory agreement that Critch notes are typical of such alliances. This isn’t necessarily intentional, but it means critics who have been refuted on one version can shift to another without acknowledging the move. The argument is undead partly because there’s no precise corpse to bury.
And this leads to the core of the paper.
A Taxonomy of Stochastic Parrots
Markovian Stochastic Parrots
This claims that LLMs are just high-order Markov chains over tokens. That is, their behavior is fully characterized as statistical transitions between symbols derived from training data. When the claim is is true of a given model, any appearance of reasoning is just the training data “speaking through” the model. But this can only be true if there are not meaningful internal states, learned representations, or latent world models.
This version was a live description of n-gram language models from 2010, and earlier NLP models were a prominent past research area of several of the authors. It was a reasonable mistake not to realize that LLMs were materially different from earlier LMs, but the evidence is now in.
In fact, LLMs are not Markovian. A Markov chain encodes token transition frequencies; a transformer does gradient descent over prediction, learning rich internal representations that generalize well beyond the training distribution. Even GPT-2 had latent states that couldn’t be described as a Markov process. The original Bender et al. paper defined stochastic parrots as “haphazardly stitching together sequences of linguistic forms… according to probabilistic information about how they combine”—which describes Markov chains, not transformer models.
Unreasoning Stochastic Parrots
This view claims that LLMs lack the kind of internal operations that constitute reasoning—reflection, iteration, counterfactual evaluation, norm-sensitive belief revision—even if they produce outputs indistinguishable from reasoning. The outputs might look right, but the process generating them isn’t reasoning.
Given the steel-man version of this view, where a model correctly generalizes in most cases, for example, as well as humans, it still could be unclear whether the correct inferential outputs are sufficient evidence of “true” reasoning. This requires a substantive view about what type of internal process constitutes reasoning. And Gary Marcus mostly holds this class of view[6], where he says that reasoning requires explicit symbolic manipulation. Alternatively, if one says they require recursive self-control, or some specific control structure, then LLMs might “fail” in some technical or even unfalsifiable sense even while producing correct outputs.
The claim is that the models don’t functionally reason—which was arguably true of GPT-3, or the claim is that the internal process matters, not just the output behavior. If you think reasoning is characterized functionally—by the right input-output relationships with appropriate sensitivity to context—then LLMs increasingly seem to qualify.
If it is not purely functional, Mechanistic interpretability has killed many additional claims, and whichever specific internal process is required, failure to reason is increasingly hard to defend as a blanket empirical claim, because LLMs demonstrate increasingly robust generalization and counterfactual sensitivity. Despite that, the view could remain valid if you hold specific views about what reasoning requires internally. However, depending on the exact claims about what is required, the argument here is identical to the Spiritual argument, and is impossible to refute.
Optimization-Artifact Stochastic Parrots
Another possible claim under the general umbrella of “Stochastic Parrot” is that what looks like reasoning in LLMs isn’t the output of a cognitive process—it’s a surface pattern induced by gradient-based optimization over training distributions. The apparent competence is brittle, doesn’t reflect genuine understanding, and will collapse under distributional shift.
This is the most clearly falsifiable hypothesized argument, since the capability to produce robust generalizable reasoning is posited to be impossible—and early evidence was mixed. That is, early results showed LLMs making bizarre systematic errors (right answer, wrong reasoning). However, more recent research on world models suggests that at least some LLMs do develop partly coherent internal representations of the domains they reason about. The brittleness argument is weakening as models improve, especially because they reason at human expert or genius level in many abstract domains—both contest and research mathematics, for example.
And, taken to its extreme, the optimization artifact argument should apply to humans too, as argued by, e.g. Scott Alexander[7]; our cognition is an “optimization artifact” of evolution and development. And Yann LeCun bites this bullet and denies humans are general reasoners. But unless you exclude humans—which the stochastic parrot argument cannot reasonably do—at some point, the question becomes what kind of optimization process can produce genuine reasoning, not whether optimization is involved.
The argument was wounded even for earlier models. Objections to this argument are not claiming full general capability, which LLMs (and humans) lack, but it seems clear that if we engage with world-model evidence rather than asserting brittleness from priors, this parrot is at least very ill, or dying.
Frozen Knowledge Stochastic Parrots
Some have argued that “genuine” reasoner must be able to update its beliefs in response to new evidence. In contrast, LLMs have fixed weights after training; any apparent “updating” during inference is just pattern completion, not genuine belief revision.
But the argument was made without reference to in-context learning, which had been a capability since GPT-3, before the Stochastic Parrots paper was published. To be clear, Bender et al. could not have seen this evidence, as it was published after their paper—but the results were shown using GPT-3, which was the primary referent of their paper.
And newer repetition of the critique say that in-context learning doesn’t count as real updating. That is, they claim “updating” without weight modification is mere simulation of updating—assuming that simulating something accurately is insufficient. Interestingly, this critique is independent of the others. A system could reason well, be grounded, have goals—and still fail here if all updating is in-context only. Conversely, a system could genuinely update its weights (as in online learning) and still fail the other critiques.
It seems clear, however, that the criticism does fail, if not initially, at least now. Further advances like reasoning models with external storage, tool use, and iterative reasoning processes went much farther. And newer approaches[8] seem to eliminate the problem entirely, so it is now a contingent engineering limitation rather than a fundamental critique of LLMs.
Social Normative Stochastic Parrots
Philosophers of language have long argued that, depending on terminology and verbal disputes, reasoning or meaning are at least partly constituted not by functional ability, but by participation in norm-governed social practices. That is, reasoning is about giving reasons to others in a social context, which requires taking responsibility and being answerable for claims. This view specifically informed the Stochastic Parrots paper, and Bender still argues that LLMs produce norm-shaped outputs without actually occupying a normative role. And, according to the argument, LLMs can’t be held accountable, can’t be committed to claims, and therefore aren’t genuine participants in the practice of reasoning. (I have argued otherwise.)
This argument asserts that persistent LLM agents cannot be held accountable or commit to claims, so that even a system that reasons flawlessly may fail this test if it lacks the social status of a reason-giver. It also usually denies that cognitive competence is sufficient to count as reasoning, so that this critique survives perfect reasoning and perfect grounding, as long as society refuses to accept LLMs.
This is probably the most philosophically sophisticated, since it again cannot be refuted by engineering progress—though this is partially a social and situational objection rather than a philosophical claim. That is, once LLMs are functionally capable of imitating reasoning, denying them a social role is a decision made by society, not an objection to anything about LLMs themselves[9].
I would suggest that this stochastic parrot argument is caged, rather than killed—it’s valid if you accept that social accountability is constitutive of reasoning rather than merely instrumentally valuable for reliable reasoning. But that means that this critique is about normative status, not cognitive capacity—which means it doesn’t establish that LLMs can’t reason, only that they can’t occupy the social role of a reasoner. (And again, the decision to deny models that social role isn’t only about the models!)
Teleological Stochastic Parrots
Philosophers have often argued that genuine reasoning is goal-directed activity[10]. LLMs optimize a loss function during training, but as the Optimization-Artifact view would suggest they don’t pursue goals during inference. This is relevant because LLMs as text predictors have no intrinsic ends and therefore aren’t agents in the relevant sense.
If we accept the above rejection of the Optimization-Artifact view, the teleological argument requires accepting that instrumental behavior (namely, producing outputs useful for achieving user goals) cannot constitute genuine goal-directedness. More recently, it also requires that agentic harnesses cannot meaningfully change that fact.
The argument is partially immune to empirical criticism, since one can always argue that whatever goals exist, they don’t count—which makes this similar to the next argument. But if the argument is anything short of being irrefutable, research on emergent misalignment suggests that at least some LLMs do develop goal-like internal structures that persist across contexts and weren’t explicitly trained in[11]. Whether this counts as “genuine” goal-directedness again depends on what you mean by genuine—but the sharp line between “tool optimized for user goals” and “agent with intrinsic goals” is blurring empirically.
However, with the advent of reasoning models that have clear intermediate term goals, the argument becomes harder to refute. Research has shown that agents given freedom to interact and set their own agendas will have hidden strategic persistence of undesired behavior, will lie about that, and will develop behaviors and goals to advance these when given freedom, or even when put in complex environments. While most agents are restricted from doing this, that is a fact about those who build the models and use them commercially, not about the models themselves.
If the argument is empirical, it seems clear that it is dead. Alternatively, it is a philosophical claim about what counts, and is effectively a version of the spiritual stochastic parrot argument.
Spiritual Stochastic Parrots
Lastly, some argue that human reasoning depends on some non-computational or non-physical ingredient—consciousness, phenomenal experience, insight, soul—that artificial systems necessarily lack. Alternatively, they argue that LLMs are not conscious, or lack some other perquisite of truly living, include the above-mentioned argument that their goals don’t really count.
This requires rejecting computational sufficiency for mind, and therefore I will claim it rejects the basis for the terminology. That is, despite using the terms, this criticism has nothing to do with the fact that they parrot inputs or that they are stochastic. I would note, importantly, that serious critics of LLMs don’t actually hold this view, even when their arguments sound like it.
This objection, like the souls that LLMs lack, is immortal but incorporeal. That is, most versions of this position are definitionally immune to any capability demonstration. If accepted, it trivially subsumes all other SP critiques, other than the socially normative view, but it does so at the cost of being untethered from both empirical evidence, and linguistic accountability. And for those critics who believe this is the reason that LLMs are truly stochastic parrots, no amount of LLM capability will matter, but perhaps they should find a new phrase to randomly repeat without understanding.
Hopefully, this post gives a useful overview of the arguments which are being made in somewhat more depth in a forthcoming paper, and feedback on the ideas would be greatly appreciated in informing my still-evolving draft of that paper.
Thanks to Bender et al. for identifying real problems with 2021-era LLMs early and largely precisely. However, the failure to formalize the philosophical argument left the argument in a state that made it harder to evaluate and easier to misuse—hopefully we have been successful in reducing the resulting confusion.
Thanks to Justis Mills for developmental editing that immeasurably improved the piece.
This may not be a fair criticism of a paper that was primarily attempting to warn of object-level dangers, not trying to make a philosophical argument. On the other hand, the framing in the paper, and repeated use of the paper as a philosophical claim about LLMs, ended up being primary.
Which approach is perhaps appropriate for the various dead arguments we are discussing.
Or, less charitably, parroting.
This analysis is based in part on Bender and Koller (2020), which was cited in the original paper, and explicitly rejected form-only semantics—a system trained only on string prediction has no way to learn meaning. That is, they cite the symbol grounding problem, and situate the argument in the context of Searle’s Chinese Room, while also emphasizing that language is “communication about the speakers’ actual (physical, social, and mental) world,” suggesting social and normative dimensions. At the same time, their “octopus” example implies that the problem is one of predictive failure and world modeling.
That is, we should noticing that the authors are all still alive, and have said various things since the Stochastic Parrots paper which further illustrate their views. But it’s unreasonable to expect authors of a paper to agree on things unrelated to the paper itself.
We do, in fact, see that the views of the authors in fact diverged since then; Bender and Hanna’s 2025 book claim it’s fundamentally confused to use any human-like terms—understanding, reasoning, belief—to describe them. This also embraces a threshold view of social-normative understanding, with a sharp normative line between LLMs’ and humans’ social practice and accountability. However, it also rejects anthropomorphization of LLMs as “nothing more than souped-up autocomplete,” returning to a nearly Markovian view.
In contrast, Mitchell et al’s 2025 paper views autonomous agents as inadvisable from a safety perspective, with explicit understanding that the models have capabilities that go far beyond parroting. (Her 2026 blog post goes much further.) And Gebru and Torres’ 2024 article seems somewhat similar, attacking the idea that we should build AGI, and urge a focus away from potentially dangerous general AI systems. In both cases, they seem not to agree with Bender’s view that AGI is a con game—but attempting to further pin down their views on the basis of increasingly marginally related work on LLMs is unprofitable, which is why this is a footnote.
Specifically, Gary Marcus has noted that “stochastic parrots” is “unkind to parrots but vividly captures something real”—while explicitly disclaiming purely philosophical commitments. His view is empirically stated: at one point he estimated a ~90% probability that current approaches will fail to scale to AGI, based on observed brittleness and extrapolation failures—though some of the argument cited have since been obsoleted by developments in increasingly sophisticated LLMs.
His position maps most directly onto Unreasoning SPs, with an explicit causal story: without symbolic representations and operations (variables, rules, compositional structure), neural networks can’t support the variable binding and systematic compositional inference that genuine reasoning requires. He grants that LLMs have rich internal representations, but argues they’re the wrong kind of representations for systematic generalization.
That view also overlaps with Optimization-Artifact SPs—he emphasizes brittleness, shallow correlations with training data, and collapse under distributional shift. Marcus doesn’t think the internal structures created as the optimization artifact are causally uninvolved, but he thinks the apparent reasoning successes caused by the optimization process are nonetheless insufficient without symbolic machinery[12].
That said, Marcus’s view is explicitly a falsifiable empirical claim, and recent LLM progress does seem to be relevant evidence for it, even if it hasn’t settled the question.
To be clear, his argument is broader, and also addresses the fundamental mistake in levels of analysis, as I and many others have argued in the past.
I would be more hesitant to cite a recent dissertation, except that the author is currently working at Thinking Machines for Mira Murati, formerly of OpenAI, Percy Liang supervised the work, and cited collaborators include Fei Fei Li.
As an aside, it’s also often directly connected to safety concerns—the worry that LLMs can’t be held accountable for their outputs in the way human reasoners can, and will therefore be unsafe. To me, this seems to be to elide the actual objections to LLM safety, but it has been argued.
That is, one recurring position in AI safety and governance, associated with Stuart Russell and others, is that systems lacking genuine accountability can’t be treated as autonomous decision-makers regardless of their competence. And Russell has used the term, guardedly. But if the concern is about accountability, this veers closer to Social Normative SPs than to the cognitive critiques—it’s about normative legitimacy, not about whether LLMs can reason. And Bender herself has used the term to make safety arguments.
Aristotle argues that reasoning is tied to ends (telē), and is about deliberating toward what is good or chosen. Much more recently, John Dewey considers reasoning a tool for resolving problematic situations, and Elizabeth Anscombe, Philippa Foot, and Alasdair MacIntyre all seem to argue that reasoning only exist when embedded in action and human goals, rather than as disembodied symbol manipulation.
That is, even without considering agents.
In this post, he quotes an unpublished piece he wrote last year stating that “we still don’t have principled solutions to any of these concerns. It’s also just pour in more data and hope for the best. None of the qualitative concerns I keep raising have been solved.”