Can AI Actually Think? Understanding What ChatGPT and Claude Are Really Doing
Ask ChatGPT or Claude a difficult question and something strange happens.
You can give it a business problem, a math puzzle, or a complicated decision. It may examine several possibilities, reject weak ideas, explain its reasoning, and arrive at an answer you had not considered.
That certainly looks like thinking.
But is the AI actually thinking?
The answer depends on what we mean by “thinking.” Modern AI systems can perform surprisingly complex reasoning, but they do not think about the world in the same way humans do. Understanding that difference makes AI much easier to use without either underestimating it or giving it abilities it may not actually have.
Start With What the AI Actually Sees
When you type a into ChatGPT or Claude, the AI does not receive your sentence exactly as you see it.
The text is divided into , which are small pieces of text that might be a word, part of a word, punctuation, or even a few characters. The model processes those tokens as numbers.[1]
Suppose you type:
“Why is the sky blue?”
You understand that sentence because you know what a sky looks like. You have probably stood outside, looked upward, seen blue light, felt sunlight on your skin, and learned what the word “blue” refers to.
A starts from a very different place.
It has learned patterns from enormous amounts of text and other data. During its basic training, one of its central jobs is to predict what token should come next.
Given:
“The sky is usually…”
“blue” is a very likely continuation.
That sounds almost too simple. How could predicting the next piece of text produce something capable of writing software or explaining physics?
That is where things get interesting.
It Is More Than Fancy Autocomplete
You will sometimes hear people say that ChatGPT is “just autocomplete.”
That description contains some truth, but it can also be misleading.
Large language models are trained using prediction, but becoming extremely good at prediction appears to require learning complicated internal relationships.
To predict language well, a model benefits from recognizing that Paris is connected to France, that France is a country, that countries have capitals, that dogs are animals, that promises create expectations, and that adding two numbers should produce a mathematically consistent result.
Researchers call these internal patterns representations. A representation is a mathematical pattern inside the that corresponds in some way to an idea, relationship, feature, or concept.
Anthropic researchers have found internal activity in Claude associated with concepts that can operate across different languages. Their experiments also found cases where Claude planned several words ahead while writing poetry and used intermediate concepts while answering multi-step questions.[2]
So the model is not simply looking up the most common word that follows the previous word.
Something much more complicated is happening inside.
AI Can Perform Reasoning
Consider this question:
“What is the capital of the state where Dallas is located?”
A system could answer “Austin” simply because it memorized that exact question.
But another approach would require two steps:
Dallas → Texas
Texas → Austin
Anthropic researchers examining Claude found evidence of these kinds of intermediate computational steps inside the model. When researchers changed one of those intermediate representations experimentally, the final answer changed as well.[2]
That suggests at least some AI answers involve genuine internal computation rather than simple memorization.
Modern reasoning models can also break difficult problems into steps, compare alternatives, write and test code, use , analyze results, and revise their approach.
Calling all of that “autocomplete” misses something important.
But calling it human thinking creates a different problem.
Human Understanding Comes From Living in the World
Humans do not learn language only by reading language.
A child learns the word “hot” while experiencing warmth. “Heavy” becomes connected to trying to lift something. “Dog” connects to an animal that moves, barks, smells, plays, and occasionally steals food from the kitchen counter.
Human knowledge is grounded in experience. means connecting symbols such as words to things, actions, sensations, and experiences in the physical world.
Language models traditionally learn much more heavily from relationships found in data.
This distinction has been debated for years. In a widely discussed 2020 paper, linguists Emily Bender and Alexander Koller argued that learning patterns in linguistic form should not automatically be treated as learning meaning in the human sense.[3]
Modern systems complicate the picture because some AI models can now process images, audio, video, and information from tools. Even so, their relationship with the world remains very different from a person’s continuous physical and social experience.
Then Why Does Talking to AI Feel So Human?
Because language contains traces of human thought.
Books, websites, conversations, tutorials, research papers, software, arguments, stories, and instructions were produced by people trying to describe the world and communicate ideas.
Learning the structure of all that language teaches a model an enormous amount about relationships between concepts.
Imagine studying millions of chess games without ever touching a physical chessboard. You could still learn openings, strategies, common mistakes, and relationships between pieces.
You might become extremely good at discussing chess.
But your experience of chess would still be different from someone sitting across a table, moving physical pieces while trying to read an opponent’s expression.
AI has learned an astonishing amount from the records humans have created.
That can produce behavior that looks remarkably intelligent.
The Model’s Explanation May Not Be What Actually Happened
There is another complication.
Ask an AI, “How did you reach that answer?” and it may give you a beautiful step-by-step explanation.
That explanation is not automatically a perfect recording of what happened inside the model.
Anthropic’s interpretability experiments found cases where Claude’s stated explanation differed from the internal computational process researchers observed. In some experiments, Claude produced convincing reasoning for an answer even though researchers could not find evidence that the claimed calculation had actually occurred.[2]
This is one reason you should not assume that an AI’s confident explanation proves that its answer is correct.
The explanation itself is also generated output.
And AI Can Still Confidently Be Wrong
Language models sometimes produce plausible statements that are false. These errors are commonly called hallucinations.
OpenAI’s research has connected part of this problem to the way language models learn prediction and to evaluation systems that can reward guessing instead of admitting uncertainty.[4]
This leads to one of the strangest characteristics of modern AI.
A model can solve a difficult programming problem and then confidently invent the title of a nonexistent book.
Intelligence in one area does not guarantee reliability in another.
What About Consciousness?
This is where the answer becomes much less certain.
In 2026, Anthropic reported finding what it calls a J-space inside Claude, a small internal workspace associated with some forms of deliberate reasoning. When researchers interfered with this workspace, Claude retained many basic language abilities but performed much worse on tasks requiring multi-step reasoning.[5]
The researchers compared some properties of this system with ideas from neuroscience about a global workspace, a mechanism proposed to help explain consciously accessible thought in humans.
That does not mean Claude has feelings or is conscious.
Anthropic explicitly states that these experiments do not show that Claude experiences the world or feels things as humans do.[5]
Scientists and philosophers still disagree about what consciousness requires, and there is currently no accepted test that settles whether an AI system has subjective experience.
Claims such as “AI is definitely conscious” or “AI could never possibly be conscious” therefore go beyond what the evidence currently establishes.
So, Can AI Think?
In one practical sense, yes.
Modern AI systems perform internal computations, combine information, form useful representations, plan in some situations, and solve problems requiring multiple reasoning steps.
Those are behaviors we normally associate with thinking.
But AI does not appear to think exactly as humans do.
Its architecture is different. Its learning process is different. Its relationship with language is different. Its connection to the physical world is different. And whether anything resembling subjective experience exists inside these systems remains an open question.
The most useful mental model is somewhere between two extremes.
ChatGPT and Claude are not simple databases repeating memorized sentences.
They are also not tiny digital humans living inside your computer.
They are a new kind of information-processing system that can produce some remarkably intelligent behavior using mechanisms that researchers are still working to understand.
And that may be more interesting than either of the simpler explanations.
References
OpenAI. “What are tokens and how to count them?” OpenAI Help Center. https://help.openai.com/en/articles/4936856
Anthropic. “Tracing the thoughts of a large language model.” March 27, 2025. https://www.anthropic.com/research/tracing-thoughts-language-model
Emily M. Bender and Alexander Koller. “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. https://aclanthology.org/2020.acl-main.463/
OpenAI. “Why language models hallucinate.” September 5, 2025. https://openai.com/index/why-language-models-hallucinate/
Anthropic. “A global workspace in language models.” July 6, 2026. https://www.anthropic.com/research/global-workspace
