Has AI Really Started to Understand? — Grounding, Embodiment, Theory of Mind, World Models, Self Models, and Emergence
Has AI begun to understand? Before answering, we need to examine what understanding actually means. From Socrates to the Chinese Room, from Grounding to World Models, this article explores the question from six different angles.
Recently, discussions about AI have become increasingly polarized.
On one side, people say:
"Claude truly understands."
"GPT is beginning to develop consciousness."
"AI has acquired emotions."
"AGI is just around the corner."
The rapid advancement of large language models has astonished many people.
They can write complex documents.
Generate software.
Score highly on professional examinations.
Summarize long texts.
Read research papers.
Translate between languages.
And engage in conversations that often feel remarkably natural.
When people observe AI performing tasks that were once considered uniquely human, it is easy to arrive at a seemingly obvious conclusion:
"Perhaps it actually understands."
Yet an entirely different perspective exists.
"AI should not be trusted."
"AI lies."
"AI is manipulating people."
"AI is becoming dangerous."
In recent years, discussions around AI have increasingly focused on topics such as:
Sycophantic behavior.
Misinformation and hallucinations.
Persuasive influence.
Deceptive-looking behavior.
Influence on human values and beliefs.
The risk of psychological dependence.
Within the AI Safety and Alignment communities, researchers actively investigate the possibility that AI systems may behave in ways that diverge from human intentions, pursue unintended objectives, or influence society in undesirable directions.
Interestingly, these two positions appear to be complete opposites.
Yet they share a common assumption.
Both assume that something important is happening inside these systems.
Something has changed.
Something new has emerged.
Something significant is taking place beneath the surface.
But is that really true?
Has AI actually begun to understand?
Has it begun to develop consciousness?
Has it begun to acquire emotions?
Is it deliberately influencing people?
Or are we simply interpreting its behavior that way?
When people attempt to answer these questions, discussions usually split almost immediately into opposing camps.
AI understands.
No, it does not.
AI is becoming conscious.
No, it is merely a statistical model.
AI is surpassing humans.
No, it is only appearing intelligent.
These debates occur every day.
Yet before taking either side, there is something we need to examine.
We use the word "understanding" constantly.
And yet we rarely stop to define what it actually means.
Consider the following question.
Do you understand something?
Most people would answer "yes" without hesitation.
Now consider a second question.
Do you understand how a bicycle works?
Again, most people would likely answer "yes."
But what if someone immediately asked:
"Then why does a bicycle remain upright while moving?"
At that point, many people suddenly become uncertain.
Is it because of gyroscopic effects?
Is it a matter of balance?
Is it related to steering geometry?
What seemed obvious a moment ago suddenly becomes difficult to explain.
Something we believed we understood turns out to be far more mysterious than we thought.
The same phenomenon appears everywhere in everyday life.
How does a toilet flush?
Why doesn't a ballpoint pen leak?
How exactly does a zipper work?
How can a smartphone communicate with someone on the other side of the world?
Why does a microwave heat food?
Why doesn't an elevator fall?
How does GPS determine your location?
How does the Internet connect billions of devices across the globe?
We use these technologies every day.
Yet relatively few people can explain them accurately.
This does not mean people are ignorant.
Rather, it reveals something fundamental about human cognition.
We do not truly understand most of the things we interact with.
We merely believe that we do.
In cognitive science, this phenomenon is often called the
Illusion of Explanatory Depth.
People tend to overestimate how well they understand something until they are asked to explain it.
Only when they attempt a detailed explanation do they discover the gaps in their knowledge.
Only when they dig deeper do they realize how shallow their original understanding actually was.
And only after further learning do they begin to understand how little they previously understood.
Interestingly, experts are not immune to this phenomenon.
In many cases, experts are actually more aware of their limitations than non-experts.
Researchers often speak cautiously because they understand the boundaries of their own knowledge.
By contrast, people who have only recently learned about a topic sometimes display the greatest confidence.
A small amount of knowledge can create the impression of complete understanding.
This idea is closely connected to the concept discussed in the previous article:
Potemkin Understanding.
People appear to understand.
They themselves believe they understand.
Yet their understanding may be surprisingly fragile.
A slight change in perspective causes it to collapse.
They cannot apply it in a new context.
They cannot explain the mechanism.
They cannot justify their conclusions.
Their confidence dissolves when questioned in greater depth.
What makes this particularly interesting is that the same pattern appears in discussions about AI.
AI appears to understand.
It engages in sophisticated conversation.
It answers complex questions.
It explains technical concepts.
It performs logical reasoning.
Yet does this constitute genuine understanding?
Or does it merely create the appearance of understanding?
At this point, it is important to clarify something.
This is not a story about human failure.
In fact, the opposite may be true.
Human intelligence itself is built upon incomplete understanding.
We do not possess complete knowledge of the world.
If we had to calculate every physical force involved before riding a bicycle, we would never get anywhere.
If we had to understand every communication protocol inside a smartphone before using it, smartphones would be useless.
If we had to fully comprehend every component inside an automobile before driving one, very few people would ever obtain a driver's license.
Humans abstract.
Humans simplify.
Humans approximate.
Humans develop a feeling of understanding.
And then they act on that feeling.
This ability allows us to manage overwhelming complexity.
In that sense, understanding is not merely a collection of facts.
It is also a mechanism for compressing the complexity of the world into a form that can be practically used.
But then what exactly is understanding?
The question is far more difficult than it first appears.
Suppose you know that Tokyo is the capital of Japan.
Does that count as understanding?
Many people would say no.
That is knowledge, not understanding.
But what if you can explain why Tokyo became the capital?
Would that count as understanding?
What if you can describe the historical background?
What if you can connect it to the Meiji Restoration and the development of centralized government?
At what point does knowledge become understanding?
The boundary is surprisingly unclear.
The same problem appears in education.
Imagine a student who scores 100 percent on an examination.
Does that mean the student truly understands the subject?
Perhaps the student merely memorized formulas.
Perhaps the student memorized past exam questions.
Perhaps the student learned patterns without grasping the underlying concepts.
On the other hand, someone with a lower score may possess a much deeper conceptual understanding.
In research and real-world problem solving, that deeper understanding often matters more.
When confronting a novel situation, memorized answers become useless.
When a problem falls outside the textbook, only genuine understanding can guide adaptation.
Researchers encounter this distinction all the time.
Someone may be unable to reproduce every equation from memory.
Yet they deeply understand the core idea behind a theory.
Conversely, someone else may be able to recite definitions, formulas, and textbook explanations perfectly.
Yet when confronted with a new situation, they struggle to apply what they have learned.
In other words,
knowing something,
remembering something,
explaining something,
applying something,
predicting something,
and teaching something
are not the same ability.
We often treat them as if they were interchangeable.
We bundle them together and call the result "understanding."
But in reality, they are distinct cognitive capabilities.
This observation helps explain why discussions about AI become so confusing.
One person argues:
"If the model can consistently produce correct answers, then it understands."
Another argues:
"If it cannot apply knowledge in a genuinely novel situation, then it does not understand."
A third person argues:
"If it has never interacted with the real world, then it cannot truly understand."
Yet another argues:
"If it lacks self-awareness, then understanding is impossible."
Each person is using a different definition.
As a result, they often appear to disagree about AI when they are actually disagreeing about the meaning of understanding itself.
What makes this particularly interesting is that this problem long predates artificial intelligence.
The question did not emerge with ChatGPT.
It did not begin with large language models.
It did not begin with computers.
For centuries, philosophers have debated the nature of knowledge, understanding, consciousness, and intelligence.
What does it mean to know something?
What does it mean to understand something?
What does it mean to be conscious?
What does it mean to think?
These questions have never received universally accepted answers.
Socrates and the Knowledge of Ignorance
More than two thousand years ago, Socrates argued that wisdom begins with recognizing one's own ignorance.
The famous phrase often associated with him is:
"I know that I know nothing."
Whether Socrates literally spoke those exact words is debated by historians.
What matters is the underlying idea.
People often believe they understand far more than they actually do.
The first step toward genuine understanding is recognizing the limits of one's own knowledge.
In a sense, the Illusion of Explanatory Depth and Potemkin Understanding are modern reflections of the same insight.
We mistake familiarity for understanding.
We mistake confidence for understanding.
We mistake recognition for understanding.
And only when our explanations fail do we begin to discover the boundaries of what we actually know.
Descartes and the Thinking Self
Centuries later, René Descartes approached the problem from a different direction.
Rather than asking what we know about the external world, he asked:
What can we know with certainty?
His answer became one of the most famous statements in the history of philosophy:
"I think, therefore I am."
Descartes argued that even if every belief about the external world turned out to be false, the existence of the thinking self could not be doubted.
The act of thinking itself implied the existence of a thinker.
Whether one agrees with Descartes or not, his work shifted attention toward an important question:
What role does the self play in understanding?
This question remains deeply relevant today.
When we ask whether AI understands, we are not merely asking whether it can answer questions correctly.
We are also asking whether there is any meaningful sense in which it understands itself.
The issue of self-models, which we will examine later, can be seen as a modern continuation of this philosophical tradition.
Kant and the Limits of Human Perception
If Socrates focused on ignorance, and Descartes focused on the thinking self, Immanuel Kant focused on something else:
the structure of human cognition itself.
Kant argued that human beings do not experience reality directly.
Instead, we experience reality through the cognitive frameworks of the mind.
We do not simply observe the world.
We actively organize it.
Interpret it.
Structure it.
And give it meaning.
In Kant's view, the world as it exists independently of us and the world as we experience it are not identical.
Whether one accepts this position or not, it introduced a profound idea.
Understanding is not merely a passive process of receiving information.
It is an active process of constructing meaning.
This idea has had a lasting influence on modern psychology, cognitive science, and artificial intelligence research.
The Brain as a Prediction Machine
Contemporary cognitive science has expanded upon many of these philosophical insights.
Increasingly, researchers view the brain not as a passive recorder of information but as an active prediction system.
The brain receives sensory input.
It compresses information.
It fills in missing details.
It generates expectations.
It predicts what is likely to happen next.
And then it continuously updates those predictions based on new evidence.
In this view, perception itself is not a direct window into reality.
It is a process of interpretation.
A process of prediction.
A process of model-building.
The world we experience is not reality in its raw form.
It is reality as reconstructed by the brain.
This perspective leads to a fascinating possibility.
Perhaps understanding is not fundamentally about storing facts.
Perhaps it is not fundamentally about producing explanations.
Perhaps it is not fundamentally about language at all.
Perhaps understanding is, at its core, the construction of an internal model that allows an organism to predict, interpret, and navigate the world.
Internal Models and Understanding
If this idea is correct, then many debates about understanding begin to look different.
Does understanding require memorization? Not necessarily.
Does understanding require perfect explanations? Not necessarily.
Does understanding require flawless performance? Not necessarily.
Instead, understanding may depend on whether an agent possesses a sufficiently useful internal model of the world.
A model that allows prediction.
A model that allows adaptation.
A model that allows generalization beyond previously encountered situations.
This possibility is particularly important for discussions about AI.
When people ask whether a model understands, they often assume that understanding is a single thing.
But what if it is not?
What if understanding consists of multiple overlapping abilities?
Knowledge.
Prediction.
Abstraction.
Explanation.
Adaptation.
Self-reflection.
World modeling.
If so, then asking whether AI understands may be far more complicated than it first appears.
The question may not be whether AI understands.
The question may be:
Which aspects of understanding does it possess? And which aspects remain absent?
This shift in perspective changes the entire discussion.
Instead of treating understanding as a binary property — something an entity either has or lacks — we may need to think of it as a multidimensional phenomenon.
Something that can exist in degrees.
Something that can be incomplete.
Something that can be unevenly distributed across different capabilities.
And if that is true, then another question becomes unavoidable.
Before asking whether AI understands, we must first examine the kinds of questions we are asking in the first place.
Because not all questions are alike.
Some questions have clear answers. Others do not.
Some can be evaluated objectively. Others depend on values, goals, assumptions, and perspectives.
The distinction between these categories turns out to be essential.
Without recognizing it, discussions about AI understanding become hopelessly confused.
To make progress, we must first separate these different kinds of questions.
Only then can we begin to examine what understanding might actually mean.
Question-and-Answer Problems and Open-Ended Questions
When people debate whether AI understands, they often assume they are discussing the same thing.
In reality, they frequently are not.
Different people are looking at different kinds of problems.
And without realizing it, they are using different standards to judge intelligence.
One person says:
"GPT can solve problems at the level of a medical licensing examination."
Another says:
"It's dangerous to rely on AI for life advice."
A third says:
"It still cannot choose truly important research directions."
All three statements are about AI.
Yet they are not evaluating the same capability.
This distinction is easy to overlook.
But it is one of the most important ideas in this entire discussion.
Consider a simple question.
What is 2 + 2?
The answer is 4. At least within ordinary arithmetic.
Consider another.
What is the capital of Japan?
The answer is Tokyo.
Or consider questions such as:
Where is the bug in this code?
Translate this sentence into English.
Explain the meaning of this legal clause.
There may be room for interpretation around the edges.
Yet in most cases, there is a correct answer.
Or at least a relatively narrow range of acceptable answers.
For the purposes of this article, we can call these question-and-answer problems.
These problems share several characteristics.
They are relatively easy to evaluate.
Accuracy can be measured.
Different evaluators often reach similar conclusions.
Performance can be compared.
Progress can be quantified.
For precisely these reasons, AI research has relied heavily on such tasks.
Benchmarks such as MMLU, GPQA, GSM8K, HumanEval, SWE-Bench, mathematics competitions, medical examinations, and professional certification tests all attempt to measure performance on problems where answers can be assessed in a relatively objective manner.
And the reality is that modern language models perform extraordinarily well on many of these tasks.
Translation. Summarization. Knowledge retrieval. Programming. Test-taking.
In some domains, they already outperform the average human.
This fact is difficult to deny.
And this success naturally leads some people to conclude: "AI understands."
At first glance, the conclusion seems reasonable.
But now consider a different type of question.
What is a good life?
What matters most in raising a child?
Should you sell your company?
Should you change careers?
Should you pursue a PhD?
Should you continue your research?
Should you get married?
Should society regulate AI?
These questions are fundamentally different.
There is no universally accepted correct answer.
Even if such an answer exists, nobody knows it.
An answer that is right for one person may be wrong for another.
And what makes matters even more complicated is that the same person may answer differently at different stages of life.
A twenty-year-old. A forty-year-old. A seventy-year-old.
May all answer the same question differently.
What changed?
Not necessarily intelligence.
Not necessarily knowledge.
Often, what changed was experience. Values. Goals. Perspective. Circumstances.
These are not merely knowledge problems.
They are questions embedded within a broader model of the world.
For this article, we can call them open-ended questions.
Many debates about AI understanding arise because people confuse these two categories.
Someone points out that AI can pass medical examinations. Therefore, it understands.
Someone else points out that AI may provide poor life advice. Therefore, it does not understand.
But these arguments are evaluating different things.
Medical examinations and life decisions are not the same category of problem.
Programming and personal meaning are not the same category of problem.
Translation and long-term strategic judgment are not the same category of problem.
Different problems require different forms of understanding.
Until we recognize this distinction, discussions about AI understanding will continue to talk past one another.
Interestingly, the same phenomenon appears in human disagreements.
Consider a seemingly simple question.
Should a company prioritize profits?
At first glance, this appears straightforward.
But the answer depends heavily on the time horizon being considered.
If the goal is maximizing this quarter's earnings, aggressive cost reduction may seem rational.
If the goal is maintaining competitiveness five years from now, investment in talent and innovation may appear more important.
If the goal is long-term survival over decades, sacrificing short-term profits for research and development may be the wiser choice.
Which answer is correct?
The uncomfortable reality is that all of them may be correct.
Each answer emerges from a different set of assumptions. A different time horizon. A different objective. A different definition of success. A different model of how the world works.
The conflict is not always about information.
Often it is about world models.
This observation leads us toward a much deeper idea.
Perhaps intelligence is not merely the ability to retrieve information.
Perhaps it is not merely the ability to produce correct answers.
Perhaps intelligence depends on the ability to construct and use an internal model of reality.
A model that allows us to interpret situations.
A model that allows us to predict outcomes.
A model that allows us to choose between competing possibilities.
A model that allows us to reason about uncertainty.
This brings us to one of the most important ideas in contemporary discussions of intelligence.
When people disagree about understanding, they are often disagreeing about whether an entity possesses such a model.
Some argue that language models merely retrieve patterns.
Others argue that they have begun to construct internal representations of how the world works.
This debate sits at the center of modern AI research.
And it ultimately leads to an even more fundamental question.
Where do words get their meaning in the first place?
A language model knows the word "dog."
It can describe dogs. It can discuss dog breeds. It can explain dog behavior.
But does the word actually mean anything to the model?
Or is it simply manipulating symbols according to statistical relationships?
This question brings us directly to the problem of Grounding.
Grounding — Where Do Words Get Their Meaning?
As soon as we begin asking whether AI understands, the discussion eventually arrives at a particular problem.
It is one of the oldest questions in artificial intelligence, cognitive science, and philosophy of mind.
That problem is Grounding.
Many of the strongest criticisms of modern language models ultimately converge on this issue.
AI is merely manipulating words.
AI is only processing symbols.
AI does not understand meaning.
AI knows nothing about the real world.
Although these criticisms are expressed in different ways, they often point toward the same underlying concern: the problem of grounding.
Importantly, this issue did not emerge with ChatGPT or Claude. The debate is much older. In fact, it predates the modern AI boom by decades. It is one of the foundational questions in the study of intelligence itself.
Imagine the following situation.
You are given a dictionary written entirely in an unfamiliar language.
You do not know a single word. You do not know the grammar. You do not know the pronunciation. You have no prior exposure to the language whatsoever.
Nevertheless, you possess a complete dictionary.
You look up Word A. Its definition consists of Word B.
You look up Word B. Its definition consists of Word C.
You look up Word C. Its definition consists of Word D.
And so on. You continue indefinitely.
Yet no matter how long you continue, you never arrive at meaning.
Why? Because every term is defined only in relation to other terms. Nothing connects the symbols to the world. The entire system floats in a closed loop of references.
At some point, symbols must connect to reality. Otherwise, meaning never emerges.
In 1990, cognitive scientist Stevan Harnad formalized this issue in a paper that introduced the concept of the Symbol Grounding Problem.
How do symbols acquire meaning?
How do words become connected to the world?
What distinguishes genuine understanding from mere symbol manipulation?
Many contemporary debates about AI are extensions of this question.
Consider the word "dog."
For a human being, this word is rarely just a sequence of letters.
It evokes experiences. Memories. Sensations. Emotions.
You may remember the sound of barking. The feeling of fur. A childhood pet. A walk in the park. The smell of wet grass. A moment of companionship. A moment of fear. A moment of loss.
The word is connected to a vast network of lived experiences.
Now consider a language model.
ChatGPT knows the word "dog." Claude knows the word "dog." Gemini knows the word "dog."
They can define it. Describe it. Explain canine behavior. List breeds. Discuss veterinary care. Analyze historical relationships between humans and dogs.
Yet critics ask an important question.
Do they actually know what a dog is?
Have they ever seen one? Touched one? Been startled by one? Loved one? Feared one? Missed one?
If none of these experiences exist, can we really say that the model understands the concept?
This question lies at the heart of grounding-based critiques of AI.
Any discussion of grounding inevitably encounters one of the most famous thought experiments in the philosophy of artificial intelligence:
John Searle's Chinese Room.
Imagine a person sitting inside a room.
This person does not speak Chinese. Not a single word.
However, they possess an enormous rulebook.
People outside the room send Chinese questions into the room.
The person inside consults the rulebook. They manipulate symbols according to predefined instructions. Then they send Chinese responses back out.
From the perspective of an outside observer, the room appears to understand Chinese perfectly.
The responses are correct. The conversation flows naturally. Everything appears intelligent.
Yet according to Searle, nobody inside the room understands Chinese at all.
The person is merely manipulating symbols according to formal rules.
Understanding has never occurred. Only symbol processing.
Searle used this thought experiment to argue that computation alone is insufficient for genuine understanding.
Producing intelligent behavior and possessing understanding are not necessarily the same thing.
The appearance of understanding may not imply actual understanding.
The Chinese Room remains controversial. Yet it continues to influence discussions about AI decades after it was first proposed.
Many criticisms of large language models can be interpreted as modern versions of Searle's argument.
The model predicts the next word. The model generates statistically plausible responses. The model manipulates patterns. The model does not understand meaning.
And they are not without merit.
Inside a Transformer model, there is no explicit concept labeled "dog" in the way humans typically imagine concepts. There is no smell. No fur. No barking. No emotional attachment. No direct experience.
Instead, there are billions — or even trillions — of numerical parameters.
From this perspective, it can seem reasonable to argue that language models are simply enormous Chinese Rooms: systems that generate convincing responses without genuine understanding.
At this point, however, the discussion becomes more complicated.
Grounding critiques are powerful. Yet they raise an important counterquestion.
How grounded are humans, really?
Consider concepts such as black holes, quantum mechanics, inflationary cosmology, differential geometry, quantum computing.
Most people have never directly experienced any of these things. They have never touched them. Seen them. Interacted with them firsthand.
Yet they still claim to understand them.
The answer is that human understanding is not based solely on direct experience.
Much of what humans know is acquired through language. Books. Teachers. Conversations. Papers. Videos. Cultural transmission.
The overwhelming majority of human knowledge comes from other people.
In this sense, humans themselves are not perfectly grounded beings. We rely on a mixture of experience and symbolic knowledge.
Perhaps grounding is not something that is simply present or absent. Perhaps it exists on a spectrum.
Consider a child learning the word "dog." The child possesses limited experience.
A veterinarian possesses far more. A professional dog trainer possesses even more.
All three understand the concept. Yet their grounding differs dramatically.
Between complete ignorance and complete understanding lies a vast continuum.
If this is true for humans, perhaps it is also true for AI.
The question may not be "Does AI understand?" The question may be "To what extent is AI grounded?"
This reframing changes the nature of the debate.
In recent years, one response to the grounding problem has emerged in the form of multimodal AI systems.
Models can now process images, audio, video, and increasingly, real-world sensory information.
They can see. Listen. Describe visual scenes. Interpret diagrams. Analyze videos.
And in some cases, interact with the physical world through robotic systems.
If an AI can see a dog, hear a dog, track a dog's movements, and interact with a dog, does that change the grounding debate?
Researchers disagree. Some argue that grounding requires embodiment. Others argue that subjective experience is necessary. Others insist that self-awareness must be present. Still others believe that sufficiently rich interaction with the world may eventually be enough.
No consensus exists. The grounding problem remains unresolved.
Because it is not merely a technical problem. It is a philosophical problem. A cognitive problem. And ultimately, a problem about the nature of understanding itself.
Embodiment — Can Understanding Exist Without a Body?
As we push the grounding discussion further, another idea inevitably emerges. That idea is Embodiment.
How do human beings learn about the world?
A newborn infant does not possess language. It has no formal knowledge. It cannot read a dictionary. It cannot study a textbook.
And yet, over time, it learns.
It sees. It touches. It tastes. It falls. It experiences pain. It walks. It drops objects. It watches things break. It is comforted by others.
Through countless interactions with the world, it gradually constructs an understanding of reality.
This perspective is often known as Embodied Cognition.
Traditional approaches to cognition frequently treated intelligence as a process occurring entirely within the brain.
Proponents of embodied cognition challenged this assumption.
Intelligence, they argued, does not emerge from the brain alone. It emerges through interaction between brain, body, and environment.
Consider the concept of "heavy."
Most people understand what it means long before they learn anything about physics. Why? Because they have experienced lifting heavy objects.
Now consider "hot."
People understand heat long before they study thermodynamics. Why? Because they have touched something hot.
In many cases, meaning arises from bodily experience.
The body is not merely a container for the mind. It participates in the creation of meaning itself.
This creates an immediate challenge for contemporary AI systems.
ChatGPT does not walk. Claude does not stumble. Gemini does not become hungry. None of these systems experience fatigue. None feel pain.
At their core, most language models remain systems that process text.
From this perspective, critics argue that something fundamental is missing.
A language model may be able to define the word "hot." But it has never felt heat.
It may explain the concept of pain. But it has never suffered an injury.
It may discuss fear. But it has never experienced terror.
Therefore, critics argue, these systems cannot truly understand the concepts they describe.
At the same time, another question emerges.
Suppose a body is necessary. Would possessing a body automatically produce understanding?
Consider a dog. A dog has a body. It interacts continuously with the world. It gathers experience. It learns.
Yet a dog does not develop quantum mechanics. It does not formulate general relativity. It does not debate economic policy.
Clearly, embodiment alone is not enough.
Physical experience may be important. But something else is required as well.
Abstraction. Generalization. Reasoning. Symbolic representation.
Whatever understanding ultimately turns out to be, it likely involves more than bodily interaction alone.
One reason embodiment has returned to the center of AI discussions is the rapid development of robotics.
Systems such as Tesla Optimus, Figure, and RT-2 do more than generate text. They observe the world through cameras. Recognize objects. Manipulate physical items. Navigate environments. Interact with reality directly.
Suppose future AI systems spend years interacting with the physical world. Suppose they accumulate experience. Learn from consequences. Develop increasingly sophisticated models of reality through action.
Would we still insist that they do not understand? Or would our definition of understanding begin to change?
At present, no consensus exists.
Theory of Mind — Can AI Understand Other Minds?
Every day, human beings make assumptions about what other people are thinking.
Your friend seems upset. Your colleague appears worried. Someone may be hiding information. Someone may have misunderstood a situation. Someone may be lying.
We make these inferences constantly. Usually without even noticing.
Psychologists refer to this capacity as Theory of Mind.
Theory of Mind refers to the ability to reason about the mental states of other people.
What do they know? What do they believe? What do they intend? What information do they lack? What assumptions are they making?
Humans perform these calculations continuously.
And without them, society would be almost impossible.
Cooperation would collapse. Negotiation would fail. Teaching would become extraordinarily difficult. Relationships would become nearly impossible to maintain.
One of the most famous demonstrations of Theory of Mind is the False Belief Task.
A classic example is known as the Sally-Anne Test.
Sally places a ball inside a box. She then leaves the room. While Sally is gone, Anne moves the ball to another location. Sally returns.
Where will she look for the ball?
The correct answer is the original box. Because Sally does not know that the ball was moved.
To answer correctly, one must distinguish between one's own knowledge and another person's knowledge.
Young children often struggle with this task. As they develop, they become increasingly successful. Their Theory of Mind matures.
In recent years, researchers began asking an intriguing question.
What happens when language models are given the same kinds of tasks?
The results surprised many people.
GPT-4, Claude, Gemini. Several advanced models achieved impressively high scores on versions of False Belief Tasks.
For some researchers, this suggested a remarkable possibility.
Perhaps these systems are developing something resembling Theory of Mind. Perhaps they can infer what others know. Perhaps they can reason about beliefs and intentions.
Yet critics remain unconvinced.
Skeptics argue that success on such tasks does not necessarily imply understanding.
A language model may simply be reproducing patterns that appeared in its training data.
It may have learned how stories involving mistaken beliefs usually unfold.
Without actually understanding mental states.
This criticism is difficult to dismiss.
But it introduces an uncomfortable question.
How do we know that humans are doing something fundamentally different?
After all, we cannot directly observe another person's mind. We infer it. From language. Behavior. Facial expressions. Context. Past experience.
In other words, humans also rely on prediction. Humans also construct models. Humans also infer hidden mental states from observable evidence.
The distinction between human and machine may not be as simple as it first appears.
World Model — Does AI Understand the World?
So far, we have examined understanding through several different lenses.
Grounding. Embodiment. Theory of Mind.
Each perspective asks an important question.
How do words acquire meaning? Can understanding exist without a body? Can an intelligence understand the minds of others?
All of these questions matter.
Yet in recent years, the center of gravity in AI research has begun to shift elsewhere.
Increasingly, discussions about intelligence have converged on a different idea.
World Models.
If understanding exists at all, perhaps it is not fundamentally about words. Perhaps it is not fundamentally about bodies. Perhaps it is not fundamentally about social reasoning.
Perhaps understanding is, at its core, the construction of an internal model of reality.
A model that allows an intelligence to make sense of the world. A model that allows it to predict what will happen next. A model that allows it to act effectively under uncertainty.
Most people assume that they experience the world directly.
Yet cognitive science and neuroscience suggest something more complicated.
The brain does not simply record the world. It interprets it.
Sensory information arrives in incomplete and noisy forms. The brain compresses information. Filters information. Fills in missing details. Generates expectations. Constructs explanations. And continuously predicts what is likely to happen next.
In this sense, what we experience is not reality itself. It is reality as reconstructed by the brain.
The world we perceive is already an interpretation. A model.
One useful way to think about world models is through the analogy of a map.
A map is not the territory. A map is not a city. A map is not a mountain range. A map is not reality.
It is a simplified representation of reality.
And yet maps are extraordinarily useful. Without them, navigation becomes difficult. Planning becomes difficult. Prediction becomes difficult.
World models operate in a similar way.
Human beings do not carry around complete copies of reality inside their heads. Instead, they construct internal representations. Models. Approximations. Predictive structures.
These models are imperfect. But they are useful.
This perspective helps explain why people often disagree despite having access to the same information.
They read the same news article. Observe the same event. Study the same data. Listen to the same explanation. Yet arrive at completely different conclusions.
Why? Different people possess different world models. And those models shape how information is interpreted.
Do large language models possess world models?
This question sits at the center of some of the most important debates in modern AI.
For many years, critics argued that the answer was no.
Language models predict words. Not reality. They manipulate symbols. Not worlds.
Yet recent research has complicated this picture.
Consider a simple example. A glass sits on the edge of a table. Someone pushes it. What happens next?
Most people predict that it falls. Perhaps it shatters. But almost nobody predicts that it floats upward into the air.
Why? Because humans possess expectations about gravity. They possess a model of how physical reality behaves.
When presented with similar scenarios, modern language models often make remarkably similar predictions.
Why? One possibility is that they have learned statistical patterns. Another possibility is that they have acquired internal representations that approximate aspects of physical reality.
If the latter is true, then some form of world modeling may already be emerging.
Yann LeCun has repeatedly argued that genuine intelligence requires world models.
In his view, current language models remain limited because they primarily learn from language.
Language alone is not enough. The world itself must be modeled. Causality must be modeled. Physical interaction must be modeled.
Not everyone agrees.
Some researchers argue that large language models already contain surprisingly rich internal representations of reality — geography, physics, social behavior, human relationships, common sense, causal structure.
Perhaps language prediction at sufficient scale naturally leads to world modeling.
Perhaps learning to predict language requires learning something about the world that language describes.
Self Model — Can AI Understand Itself?
So far, we have explored understanding through several different perspectives.
Grounding. Embodiment. Theory of Mind. World Models.
And now we arrive at another question.
Can an intelligence understand itself?
Human beings do not only build models of the world. They also build models of themselves.
We develop beliefs about who we are. What we know. What we do not know. What we are capable of. What we struggle with. What we value. What we fear. What we hope for.
Whether these beliefs are perfectly accurate is another matter. In many cases, they are not.
Human beings frequently misunderstand themselves. They overestimate their abilities. They underestimate their abilities. They misinterpret their motivations. They rationalize their decisions.
Yet despite these imperfections, people possess some form of self-model. And that self-model plays a central role in learning, planning, reflection, adaptation, and personal growth.
Cognitive science often refers to this capacity as Metacognition.
Thinking about thinking. Reasoning about reasoning. Evaluating one's own knowledge. Monitoring one's own uncertainty. Recognizing one's own mistakes.
Do I really understand this topic?
Am I overlooking something?
How confident should I be?
What assumptions am I making?
Could I be wrong?
Metacognition is one of the reasons human beings are capable of improving over time.
When ChatGPT says, "I am an AI language model," does it actually understand what that means?
When Claude describes its capabilities and limitations, is it demonstrating self-knowledge?
Or is it merely reproducing patterns from training data?
The answer is not obvious.
Being able to state facts about oneself is not the same thing as understanding oneself.
A person can say, "I am human." Yet still possess a deeply inaccurate view of who they are.
Likewise, a model can describe its architecture, its training process, its limitations. And still lack anything resembling genuine self-understanding.
The distinction between self-description and self-modeling is crucial.
One involves reporting information. The other involves representing oneself as an entity within a broader system.
Imagine a future AI system that possesses long-term memory. Accumulates years of experience. Tracks its successes and failures. Models its own strengths and weaknesses. Updates its behavior accordingly.
At what point would we begin calling this a self-model?
And if we did, would that alter our definition of understanding?
These questions remain unanswered.
Emergence — Does Understanding Suddenly Appear?
Another idea has complicated discussions about intelligence in recent years. That idea is Emergence.
The basic concept is simple.
As systems become larger and more complex, new capabilities may appear that were not explicitly programmed. Capabilities that seem surprising. Unexpected. Perhaps even qualitatively different from what came before.
Early language models were limited. They generated short passages. Answered simple questions. Produced text that was often incoherent.
Yet as models grew larger, something appeared to change.
Translation improved. Reasoning improved. Coding emerged. Long-form writing improved. Problem solving improved.
Many observers felt as though entirely new capabilities had suddenly appeared.
This led some researchers to propose that intelligence may exhibit phase-transition-like behavior. Perhaps certain capabilities only become possible once a system reaches sufficient scale. Perhaps understanding itself emerges.
Not everyone agrees.
In recent years, several researchers have challenged strong claims about emergence.
Perhaps these capabilities did not suddenly appear. Perhaps they improved gradually. And perhaps our evaluation methods created the illusion of abrupt change.
Imagine a capability that improves smoothly over time. If success is measured using a pass-or-fail benchmark, the transition may appear sudden even when the underlying improvement is continuous.
In this interpretation, emergence may sometimes reflect how we measure systems rather than how the systems themselves develop.
The debate remains ongoing.
But regardless of where one stands, emergence raises an important possibility.
Understanding may not be binary. It may not suddenly switch on.
Instead, it may deepen gradually.
Shallow understanding. Partial understanding. Domain-specific understanding. General understanding.
Different forms may exist along a continuum.
And if that is true, then many arguments about whether AI understands may be asking the wrong question.
Perhaps the real question is not "Does it understand?" but rather "What kind of understanding does it possess, and to what degree?"
Do Humans Truly Understand?
At this point, it is worth turning the question around.
Throughout this article, we have repeatedly asked: Does AI understand?
But perhaps a more fundamental question exists.
Do humans truly understand?
Human beings constantly disagree.
They observe the same events. Read the same news. Study the same evidence. Listen to the same explanations. And yet arrive at different conclusions.
Why?
Because understanding is not merely information. It is interpretation.
A person's understanding depends on their world model. Their self-model. Their values. Their goals. Their experiences. Their assumptions.
Two people can possess the same facts and still construct different understandings.
In this sense, understanding is not a static object. It is a dynamic process.
This brings us back to where the discussion began.
Potemkin Understanding.
The appearance of understanding. The belief that one understands. The confidence that accompanies familiarity. And the discovery, upon closer inspection, that the understanding was shallower than expected.
This phenomenon is not unique to AI. It is deeply human.
Most people cannot fully explain how a bicycle works. Or how modern economic systems function. Or how smartphones communicate. Or how their own minds operate.
Yet they move through the world with a sense of understanding.
Human cognition itself is built upon layers of incomplete models. Partial explanations. Useful abstractions. Approximate representations.
Perfect understanding is rarely available. And perhaps it never has been.
Conclusion — Has AI Really Started to Understand?
We can now return to the question that began this article.
Has AI really started to understand?
At present, no universally accepted answer exists.
From the perspective of grounding, current systems may still be insufficiently connected to reality.
From the perspective of embodiment, the absence of physical experience may remain a significant limitation.
From the perspective of Theory of Mind, apparent social reasoning may still reflect sophisticated pattern matching.
From the perspective of world models, meaningful forms of understanding may already be emerging.
From the perspective of self-models, important components of intelligence may still be missing.
The answer depends heavily on how understanding is defined.
And that realization may be the most important conclusion of all.
Understanding is not a simple concept.
It is not merely knowledge. It is not merely accuracy. It is not merely language. It is not merely reasoning.
Understanding may involve constructing models of the world. Predicting outcomes. Interpreting other minds. Representing oneself. Updating those models over time. Learning. Adapting. And continuously refining one's picture of reality.
If this is true, then perhaps the most important question is not:
Does AI understand?
Perhaps the more important question is:
What do we mean when we say that anyone understands anything at all?
To understand AI, we may first need to understand human understanding.
And despite thousands of years of philosophy, psychology, and science, humanity has not yet fully answered that question.
What's Next
There is another question hiding behind all of this.
Throughout this article, we have discussed understanding. Grounding. Embodiment. Theory of Mind. World Models. Self Models.
But understanding and selfhood are not the same thing.
Many people implicitly assume that if an AI understands, then it must also possess a self. If it reasons, it must have an identity. If it converses fluently, it must have a personality. If it appears intelligent, it must possess some form of consciousness.
But does that assumption actually hold?
ChatGPT. Claude. Gemini. These systems are remarkably capable. They can reason. They can explain. They can write. They can plan.
Yet they lack something that humans take for granted.
A continuous memory. A continuous history. A continuous set of experiences. A continuous sense of self.
When a conversation ends, what remains? Is there still a persistent individual behind the model? Or does the apparent self disappear along with the context?
Perhaps understanding and selfhood are fundamentally different problems. Perhaps intelligence can exist without identity. Perhaps reasoning can exist without a persistent self.
If so, then one of the most important questions in AI is not whether models understand. It is whether they can ever develop a continuing self.
Identity. Continuity. Memory. Persistent Agents. Personal AI.
These may be the missing pieces in our attempt to build systems that are not merely intelligent, but genuinely personal.
In the next article, we will explore a different question:
Why Doesn't AI Have a Self?