Does AI Really Understand What It Says?

Does AI understand what it says? Not quite yes, and not simply no: large language models build genuine internal representations of meaning, but without any experience of the world and with unpredictable gaps. The practical takeaway for you is simple: treat a fluent answer as a well-phrased hypothesis, not as proof that the AI has grasped your situation.

What a language model actually does

A large language model (LLM) is trained on a vast amount of text with a single instruction: predict what comes next. The text is split into tokens, chunks of words, and the model learns to guess the next token. By repeating this billions of times over its training data, it tunes its internal settings.

Put that way, it sounds far from understanding. That is the “giant autocomplete” argument. But there is a catch: to predict the next part of a text well, you often need to have picked up a lot. To complete “Paris is to France as Madrid is to…,” you need some notion of capitals and countries. To finish a math proof, you need to follow its logic. So the real question becomes: what structures did the model build in order to predict so well?

If you want a refresher on the basics, our article on how ChatGPT works walks through the steps, and what is an LLM explains the architecture.

The case for “no, it does not understand”

The Chinese Room

In 1980, philosopher John Searle proposed a thought experiment. Someone who does not speak Chinese is locked in a room with a rulebook. Questions in Chinese are slipped under the door; by following the rulebook, the person sends back perfect answers. From outside, the room “speaks Chinese.” Inside, nobody understands a word. For Searle, manipulating symbols by rules is not enough to understand.

The “stochastic parrot”

In 2021, researchers Emily Bender, Timnit Gebru and their co-authors published a paper describing language models as “stochastic parrots”: they stitch together likely-sounding language without access to meaning. Their core argument: the model has never seen, touched or lived through what it talks about. It only has words pointing to other words. Researchers call this the grounding problem.

Telltale mistakes

Models also make mistakes that a person who truly understood would not: confidently inventing a reference, failing a trivial variant of a well-known puzzle, contradicting themselves from one sentence to the next. These hallucinations show that fluent text is not backed by any check on meaning.

The case for “yes, partly”

Measurable internal representations

Research in interpretability, which opens up the “black box” of models, offers concrete evidence. A study presented in 2023 trained a small model only on lists of moves from the board game Othello: inside it, researchers found a representation of the state of the board that nobody had given it.

More recently, on March 27, 2025, Anthropic published work on its Claude model (“Tracing the thoughts of a large language model”, accessed October 22, 2026). According to Anthropic, when asked for the capital of the state where Dallas is located, the model first activates the idea “Dallas is in Texas,” then “the capital of Texas is Austin.” By artificially swapping that intermediate step, researchers change the final answer. The same post describes the model picking a rhyming word in advance before writing a line of poetry, and concepts shared across several languages.

Results that are hard to fake by memorizing

Test scores have also climbed sharply. GPQA is a set of PhD-level science questions designed so that the answers cannot be found with a quick web search.

Quarter	Best GPQA score (out of 100)	Model
2023-Q1	35.7	GPT-4
2024-Q2	54	Claude 3.5 Sonnet
2024-Q4	76.8	o1
2025-Q4	92.6	Gemini 3 Pro
2026-Q3	96.4	Gemini 4 Argon

Source: Epoch AI, best score per quarter, retrieved October 22, 2026.

To follow the models with the best test results, see our weekly AI ranking.

The same research adds an important caveat

Anthropic’s work also shows the other side. When the model is asked to explain how it did a piece of mental arithmetic, it describes the standard method taught in school, while its internal circuits take a different route, combining a rough estimate and a precise calculation in parallel. In other words: the explanation an AI gives of its own reasoning is not necessarily what happened inside.

What looks like understanding	What falls short of it
Concepts shared across languages	No direct experience of the world
Observable intermediate reasoning steps	Self-explanations that can be inaccurate
Planning (choosing the rhyme before writing the line)	Basic errors on simple variants
High scores on difficult tests	Inventions stated with confidence

Check yourself An assistant explains step by step how it reached a result. Does that explanation prove it reasoned that way?

No. According to Anthropic’s interpretability research, the explanation a model gives of its reasoning can differ from the computation it actually performed. Judge the result itself, by checking it, rather than the quality of the explanation.

So which word should we use?

This debate echoes the Turing test: back in 1950, Alan Turing suggested replacing the question “can machines think?” with an observable test. Today, several models hold convincing conversations, and the question has moved to what happens inside them. Our plain-English guide to artificial intelligence puts these ideas in a wider context.

What this means for how you use AI

1. Check understanding before the answer. Ask the assistant to restate your request or list its assumptions.
2. Do not trust the tone. A confident, well-written answer can be wrong; our guide to AI hallucinations suggests ways to check.
3. Provide the context it lacks. The model knows nothing about your situation, your constraints or what seems “obvious” to you: spell it out.
4. Keep the final judgment. On medical, legal or financial matters, AI helps you prepare your questions; it does not replace a qualified professional.