P-0007 ● Active

“Large language models understand language.”

Community stance
No votes yet
§
Meaning & context

This is partly an empirical question and partly a fight over the word "understand". Sceptics argue that a system trained only on text predicts word sequences without access to what words refer to, so it cannot understand them. Others point to internal representations of space, time and game states that look like world models, and to behaviour that would count as understanding in a human.

Interpretability research has moved the debate. We can now look inside models and find structured representations. The disagreement is shifting to whether those representations are the right kind, and whether brittle failures reveal pattern-matching beneath the competence.

UnderstandContested
▶

Strong reading: grasping meaning the way humans do, with grounding in perception and intention. Functional reading: using language in ways that track meaning and the world, including generalising to new situations. Most of the dispute comes from mixing these.

Large language models
▶

Neural networks trained on large text corpora to predict tokens, possibly further trained with human feedback. Multimodal models trained on images and audio complicate the "text only" argument.

Filter by type:
↕
Arguments
6 total
↑ Supporting3
Empirical

Models trained only to predict the next move in a board game build an internal representation of the board state, which researchers can read out and even edit to change the model's predictions. Prediction at scale produces world models, not just surface statistics.

Li and colleagues trained a GPT model on Othello move sequences with no knowledge of the rules; probes recovered the board state from its activations, and intervening on that representation changed its moves consistently.

Li et al. (2023). Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task. ICLR 2023.

Limitation

A board game is a tiny closed world. Showing a model can represent an 8×8 grid says little about whether it represents meaning in open-ended language.

Empirical

Language models contain linear representations of real-world space and time: the latitude and longitude of places and the dates of events can be read directly from their activations.

Gurnee and Tegmark found such representations in Llama-2 models, consistent across entity types (cities, landmarks, events), suggesting a learned map of the world rather than isolated facts.

Gurnee & Tegmark (2024). Language Models Represent Space and Time. ICLR 2024.

Limitation

Coordinates can be correlated with word co-occurrence, so this might reflect structure in text rather than a model of the world. The disagreement here is about what "really" representing means.

Definitional

By the behavioural standards we apply to people (paraphrasing, answering novel questions, explaining, translating, catching ambiguity, reasoning through new cases), these systems often perform at or above human level. Denying them understanding requires a standard no behaviour could meet.

Mitchell and Krakauer's review notes that the field is split; a 2022 survey of NLP researchers found them divided about evenly on whether models trained only on text could understand language in some non-trivial sense.

Mitchell & Krakauer (2023). The debate over understanding in AI's large language models. PNAS, 120(13), e2215907120.

Limitation

Behavioural tests can be passed by systems that exploit statistical shortcuts; benchmark contamination is a known problem.

No arguments of this type on this side.
↓ Opposing3
Definitional

Meaning is the relation between language and something outside it: objects, intentions, communicative goals. A system trained only on linguistic form has no access to that relation, however fluent it becomes.

Bender and Koller's "octopus" thought experiment: an octopus intercepting telegraph messages between two islanders could learn to mimic their replies, but would fail as soon as a reply required knowing what the words refer to (e.g. how to build a weapon against a bear).

Bender & Koller (2020). Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. ACL 2020, 5185–5198.

Limitation

Multimodal models trained on images, audio and actions weaken the "form only" premise. Humans also learn much of their knowledge from language alone.

Empirical

Models fail in ways that reveal reliance on the statistics of their training data. They do much worse on versions of tasks that are rarer online, even when the logic is identical, and models trained on "A is B" often fail to infer "B is A".

McCoy and colleagues showed GPT-4 decodes a common cipher (shift of 13) far more accurately than a rare one (shift of 12). Berglund and colleagues found models trained on "X is the parent of Y" often could not answer "who is Y's parent?".

McCoy, Yao, Friedman, Hardy & Griffiths (2024). Embers of autoregression. PNAS, 121(41), e2322420121; Berglund et al. (2024). The Reversal Curse. ICLR 2024.

Limitation

Humans also show frequency effects and asymmetric recall. Newer reasoning models reduce many such failures, which suggests a limitation of particular models rather than of the approach.

Definitional

"Understanding" is being redefined to whatever models do. Fluent output is exactly what a system optimised to produce plausible text would generate whether or not it understands, so fluency cannot be the evidence.

Bender, Gebru and colleagues described language models as "stochastic parrots" that stitch together linguistic forms according to probabilistic information about how they combine, without reference to meaning.

Bender, Gebru, McMillan-Major & Shmitchell (2021). On the Dangers of Stochastic Parrots. FAccT 2021, 610–623.

Limitation

The argument is unfalsifiable if no behaviour could count as evidence. Interpretability findings offer non-behavioural evidence.

No arguments of this type on this side.
≈
Areas of agreement
3 points
Empirical

Language models form internal representations that are richer than simple word co-occurrence counts.

Empirical

Models still make errors no competent human would make, especially on rare variants of familiar tasks.

Definitional

Much of the disagreement turns on how "understanding" is defined.

P-0007 · Last edited Oct 26, 2026 · 1 revision