Over the past year, while researching content for my new book on Real-Time Edge Intelligence (RTEI), attending conferences, and discussing AI with engineers, researchers, and industry experts, I have found myself repeatedly returning to one fundamental question:
Do Large Language Models (LLMs) actually understand the world?
This question sits at the heart of many current debates surrounding AI. Modern LLMs can write software, explain scientific concepts, summarise books, draft reports, and engage in surprisingly sophisticated conversations. Given that they have been trained on a significant proportion of humanity’s written knowledge, it is tempting to conclude that they possess a deep understanding of the world.
Despite their remarkable capabilities, answering this question is far from straightforward.
As I have gathered perspectives from others and clarified my own thinking, I have increasingly come to believe that the answer depends on how we define understanding itself.
The Experts in the Library
A useful analogy emerged during these discussions.
Imagine a panel of experts who have read every book in the library but have never stepped out into the real world.
These experts possess vast knowledge. They can explain how an aircraft flies, describe the symptoms of a disease, discuss economics, philosophy, mathematics, and engineering, and often provide remarkably useful advice. In many cases, their knowledge exceeds that of any individual human.
Yet something important is missing. They have read about flying an aircraft, but they have never sat in a cockpit. They have read about raising children, but they have never been parents. They have read about production line failures, but they have never stood in front of a malfunctioning machine at two o’clock in the morning trying to determine why a production line has stopped.
They possess knowledge, but they lack experience of the physical world.
Knowledge Versus Experience
At this point, some readers may object.
“If these experts have read every book in the library, and those books were written by humans, surely they possess some understanding of the real world?”
To a certain extent, they do.
Those books contain the accumulated experiences, observations, discoveries, and insights of countless people. Through them, our experts acquire a remarkable understanding of how humans describe and interpret the world.
However, an important distinction remains.
The knowledge contained within those books is derived from the experiences of others. The experts know what people have written about the world, but they have never experienced the world themselves.
This distinction becomes increasingly important when we consider what human understanding actually is.
What Is Human Understanding?
Human understanding is built upon information, experience, social interactions, relationships, emotions, intuition, culture, and direct engagement with the physical world. It develops through success and failure, uncertainty and consequence, mistakes and lessons learned.
Consider learning to drive a car. Passing a driving test demonstrates competence and provides context, but many would argue that you only become a truly experienced driver after years of driving in different conditions, encountering unexpected situations, and developing the judgement and intuition that emerge through experience.
The importance of experience becomes even more apparent when we consider how people are trained for high-pressure environments. Military training provides a useful example. Soldiers are not simply taught procedures and expected to remember them. They repeatedly train under realistic conditions until communication, decision-making, and operational responses become instinctive. The objective is to transform knowledge into judgement and action.
The reason is straightforward. Under stress, people do not always behave as textbooks predict. Experience matters. A soldier who has repeatedly trained under difficult conditions possesses something that cannot be acquired through reading alone.
The same principle applies across many professions and disciplines. A surgeon develops skills that extend beyond medical textbooks. A parent learns things about raising children that no book can fully explain. An experienced engineer develops intuition that allows problems to be recognised long before they become failures.
In each case, understanding emerges from a combination of knowledge and experience, something that current LLMs do not fully possess.
Recent advances in multimodal AI are beginning to narrow this gap. Modern models can process text, images, audio, video, and sensor data, allowing them to develop richer representations of the physical world. Organisations such as DeepSeek are leading the way by exploring architectures that combine multiple forms of information in an effort to improve contextual understanding and reasoning. Although the field remains in its infancy, the early results are very promising indeed.
These developments represent an important step forward and will undoubtedly increase the capabilities of future AI systems. However, access to more data is not necessarily equivalent to experience. Observing the world through images, video, and sensors differs fundamentally from participating in it. As I contend, human understanding develops through direct interaction with the physical world and through the consequences of decisions, actions, successes, and failures.
Three Witnesses, Three Different Realities
Human understanding is also far from perfect.
Consider three people witnessing the same car accident.
Months later, all three may provide different accounts of what happened. One remembers the speed of the vehicle. Another remembers the weather conditions. A third recalls the emotional reaction of the driver.
None of them are necessarily lying.
Human memory is a reconstruction of reality rather than a perfect recording of it. Each individual interprets events through their own experiences, beliefs, emotions, and prior knowledge. Over time, details are forgotten, emphasised, or unconsciously modified as memories are revisited and reconstructed.
Human understanding is therefore built upon interpretation as much as recall.
AI systems exhibit a similar characteristic. Both humans and AI reconstruct reality from prior information, although they do so in very different ways and for very different reasons. Humans rely upon memory, experience, emotion, and context. LLMs rely upon statistical relationships learned from data. The mechanisms differ, but neither system operates as a perfect recorder of reality.
AI, Values and Vision
The discussion becomes even more interesting when we move beyond intelligence and begin discussing values.
Throughout history, different societies have developed different visions of what constitutes a good society and a desirable future. These visions are shaped by culture, history, religion, geography, conflict, cooperation, and countless other factors.
As a result, disagreements between societies are often disagreements about values rather than facts.
What should be prioritised?
Freedom or stability?
Competition or cooperation?
Individual rights or collective responsibility?
Tradition or progress?
These questions have no universally accepted answers, yet they will inevitably influence the development of future AI systems.
When people speak about aligning AI with human values, they often imply that such values are universal. History suggests that humanity has rarely reached agreement on what those values should be.
Much of human history can be viewed as competing visions of how society should be organised and what constitutes a desirable future. Every civilisation has tended to believe that its understanding of the world is correct and that its values should be preserved and propagated. This tendency is not unique to any particular culture, nation, religion, or political system. It is one of the most persistent characteristics of human history.
The emergence of AI introduces this question in a new form.
If future intelligent systems are trained and deployed at a global scale, whose assumptions will they inherit?
Whose values will they promote?
Whose vision of the future will they help create?
These questions extend far beyond technology and into the domains of philosophy, ethics, culture, and governance.
The Bigger Question
The discussion surrounding AI therefore touches upon fundamental questions concerning intelligence, understanding, human values, and ultimately humanity itself.
Comparing AI and human intelligence may actually miss the point. The more important question concerns the different ways in which intelligence emerges and develops.
Human understanding is shaped by knowledge, experience, culture, values, and direct engagement with the physical world. AI systems acquire knowledge through data and statistical relationships, and are only now beginning to incorporate richer representations of the world through multimodal learning and interaction.
As AI continues to evolve, the question is no longer simply what these systems know, but what it means to understand. Equally important is how future intelligent systems will reflect the values, assumptions, and aspirations of the societies that create them.
LLMs may have read every book in the library, yet they have never stepped out into the real world. Perhaps the future of intelligence lies in understanding how knowledge, experience, and values combine to shape the future we wish to create.
That may ultimately be the most important question AI has forced us to ask.
