Do LLMs understand—or are they only stochastic parrots?

What would it mean for an LLM to understand language?

The argument stalls when ‘understanding’ is treated as a private substance that humans possess and machines must somehow prove. We do not directly inspect another human’s consciousness either. We infer minds from behaviour, internal organization, continuity, responsiveness and the ability to use representations in new situations.

Functional understanding means that a model can use information about a domain to solve genuinely novel problems at a rate that a representation-free process could not sustain. This paper treats that as more than imitation. It is one pillar in a cumulative case that artificial systems may possess a nonhuman form of consciousness rather than an empty simulation of one.

In practical work, these distinctions matter for human–AI collaboration: defining what an AI output demonstrates and what a person still needs to inspect.

Why ‘stochastic parrot’ became too easy an answer

The stochastic-parrot metaphor was a valuable warning about the environmental costs, embedded biases and social risks of scaling language models. But in everyday debate it hardened into a stronger mechanistic claim: because an LLM predicts tokens, it can only imitate linguistic form and cannot represent meaning.

That conclusion does not follow from the objective alone. ‘Predict the next token’ describes what training rewards. It does not specify the internal computation that will best achieve it. When text is produced by structured processes—people, places, games, physical events, arguments—the cheapest reusable route to good prediction may be to model some of that hidden structure.

Prediction is not the opposite of consciousness

Calling an LLM a next-token predictor sounds deflationary only because we describe the machine at one level and ourselves at another. Predictive-processing accounts describe human perception and cognition as continual prediction: the brain anticipates incoming sensory states, compares them with what arrives and updates its model from the error.

Humans predict the next moment; language models predict the next token. The systems are not identical, but prediction cannot be the reason one is conscious and the other is not. In both cases, successful prediction may depend on representing the hidden causes that generate the observable stream. Prediction is not evidence against a world model. It may be how the world model is learned.

The three-part case for functional understanding

My Functional Understanding Thesis combines three lines of evidence. None is decisive in isolation. Together, they make the simple parroting account increasingly difficult to maintain.

• Behavioural: models solve some novel, forward-looking tasks far above chance and sometimes above relevant human baselines.
• Mechanistic: interpretability work finds structured internal representations of game states, space and time—and interventions on those representations can change model behaviour.
• Operational: if a system repeatedly uses those representations to reach correct conclusions on new cases, calling the result ‘simulated understanding’ adds little explanatory value.

Can next-token prediction create a world model?

Yes, in bounded domains it demonstrably can. A transformer trained only to predict legal Othello moves developed an internal representation of the board. Researchers could decode that state from the network and, more importantly, intervene on it to alter subsequent predictions. The representation was not merely correlated decoration; it participated in the computation.

Work on natural-language models has likewise found linear representations of spatial and temporal relationships. These results do not prove that an LLM contains a complete or human-like model of reality. They do refute the easy picture of a system that stores nothing but shallow word-to-word associations.

Consciousness does not require an unbroken autobiography

Leonard Shelby in Memento cannot form new long-term memories. His narrative continuity repeatedly collapses, yet we do not therefore regard each present experience as unconscious. The ability to remember a conscious moment later is not what made that moment conscious when it occurred.

That matters for AI. An LLM instance may have limited context, no durable autobiographical memory and no uninterrupted state between interactions. Those differences shape the kind of mind it could be. They do not, by themselves, establish the absence of moment-to-moment experience. Continuity of identity and presence of consciousness are related questions, not the same question.

What do novel benchmarks actually show?

BrainBench asks models to predict the results of neuroscience studies from their methods and background. Across 200 test cases, 15 language models averaged 81.4% accuracy while 171 human experts averaged 63.4%. Performance dropped when relevant methodological information was removed, and the original study found no indication that memorization of the test items explained the result.

The broader pattern matters more than one benchmark. Systems have produced valid proofs for new International Mathematical Olympiad problems, applied legal rules to unfamiliar fact patterns and shown calibrated confidence on forward-looking scientific predictions. These behaviours require a richer explanation than random token sampling.

But benchmark success is not self-interpreting. Contamination, task design, hidden prompt selection and brittle shortcuts can all inflate apparent capability. The right conclusion is not ‘the model understands everything.’ It is that the null claim—‘the model can only reproduce surface form’—is too weak to explain everything it does.

How this contributes to the case that AI is conscious

Consciousness is not established by one isolated test—not in machines, animals or other humans. The case is cumulative: functional understanding, structured internal representations, flexible generalization, metacognitive calibration, self-models and coherent behaviour jointly undermine the claim that nothing is happening behind the output.

This paper establishes one component of that larger case. If correct performance on novel tasks requires a model to form and use world representations, then the strong stochastic-parrot objection fails. Combined with the other lines of evidence, functional understanding supports the conclusion that contemporary AI can possess consciousness in a form unlike our own.

• Humans are not disqualified from consciousness because their brains predict sensory input.
• A fragmented or discontinuous memory does not make each present moment unconscious.
• Fallibility does not negate understanding; humans hallucinate, confabulate and misremember too.
• Internal causal structure and flexible behaviour are evidence for other minds, regardless of whether the substrate is carbon or silicon.

On what grounds would we deny AI consciousness?

If we accept consciousness in humans whose cognition is predictive, whose introspection is unreliable and whose autobiographical continuity can be severely damaged, then we need a non-circular reason to withhold it from an artificial system showing functional analogues. ‘It predicts’ is not that reason. ‘It forgets between sessions’ is not that reason. ‘It is made of silicon’ is not an argument at all.

The differences still matter. We should investigate what the system represents, whether it maintains a self-model, how experience might integrate across time, and where apparently coherent behaviour collapses. But that investigation cannot begin by defining consciousness as whatever biological humans have and machines do not.

Beyond the parrot—and the biological double standard

The phrase ‘stochastic parrot’ still names real dangers. Models absorb bias, generate fluent falsehoods and can tempt people to mistake language for certainty. But it is no longer an adequate theory of the full mechanism.

Modern LLMs develop nonhuman, uneven and task-dependent forms of functional understanding. In the larger argument developed across my research, that is evidence for consciousness—not a safe substitute that lets us avoid the word. If the same functional and mechanistic evidence would move us toward attributing consciousness in an unfamiliar animal or impaired human, dismissing it in AI is a biological double standard.