From Computation to Cognition Assessing the Integration Gap Between Artificial Intelligence and Human Cognition
This review argues that while current AI systems demonstrate strong functional and behavioral task performance, they lack the mechanistic integration of memory, learning, context sensitivity, decision-making, and embodiment required for genuine human-like cognition, and it proposes a research agenda with measurable criteria to bridge this gap.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where machines can do things we once thought only humans could do. They can recognize a face in a crowd, translate a foreign language in an instant, or write a poem that moves us to tears. For decades, scientists have built computers to follow strict rules, but in the last ten years, a new kind of machine has emerged. These systems, often called large language models, learn from vast amounts of text and data to predict what comes next, whether it is the next word in a sentence or the next step in a plan. They are incredibly good at performing specific tasks, often matching or even beating human experts in narrow fields like medical diagnosis or financial forecasting. But a deeper question has begun to trouble researchers in computer science, psychology, and neuroscience: just because a machine can produce the right answer, does it actually understand the world the way we do?
This question is not about whether a robot can solve a math problem faster than a person. It is about whether that robot possesses a unified mind. Human thinking is not a collection of separate tricks; it is a seamless flow where memory, learning, context, and decision-making work together constantly. We remember a past event, use that memory to understand a new situation, and adjust our behavior based on how we feel and what is happening around us. All of this happens in a single, continuous system that changes and grows with every experience. The big question is whether our artificial creations are beginning to build this same kind of integrated mind, or if they are simply mimicking the surface of human intelligence without the underlying machinery.
A recent review by Muhammad Subhan from the University of Tasmania takes a hard, clear look at this issue. Instead of asking if machines are "smart" in a vague sense, the study breaks the problem down into three distinct levels. The first level is functional reproduction, which simply asks if a machine can do the same job as a human. The second is behavioral reproduction, which asks if the machine acts like a human not just on average, but in tricky situations where humans might stumble or change their minds. The third and most difficult level is mechanistic reproduction, which asks if the machine is using the same kind of internal processes that our brains use. The study finds that while machines are excellent at the first two levels, they are still very far from the third.
The researchers examined a wide range of modern artificial intelligence systems, from the text-generating models we use every day to robots that can see and move in the physical world. They looked at how these systems handle memory, learning, context, decision-making, and even emotion. The evidence shows that machines can certainly perform tasks that require memory. For instance, a system can look up information in a database and use it to answer a question. This is a functional match for human memory. However, the way the machine does this is fundamentally different from how a human brain works. A human brain does not store facts in a separate file cabinet and then go look them up. Instead, our memories are woven into the very fabric of our thinking; they are patterns of connections that change and strengthen as we learn. When a machine retrieves information, it is usually running a separate, engineered search process that is bolted onto its main brain, rather than having the memory be an intrinsic part of how it thinks.
This gap becomes even wider when looking at learning. Humans learn continuously. We can hear a new fact today and use it tomorrow without having to stop and reprogram our entire brain. We can also learn from mistakes without erasing everything we knew before. Current artificial intelligence systems struggle with this. When they learn something new, they often forget what they knew previously, a problem scientists call catastrophic forgetting. Engineers have built patches to fix this, such as special techniques that protect old information while new data is added, but these are external solutions. They are not the same as the natural, plastic way human neurons change and adapt over a lifetime. The machine is not learning in the same way a person does; it is being patched up to avoid a specific failure mode.
The study also looked at how these systems handle context and decision-making. Humans are incredibly sensitive to the situation they are in. We know that a word might mean one thing in a joke and something completely different in a serious argument. We make decisions based on our goals, our emotions, and our past experiences, often using shortcuts that are not perfectly logical but work well in the real world. Artificial systems can be very good at making decisions when the rules are clear and the goal is specific. But when the situation is vague, or when the wording of a question changes slightly, these systems often stumble. They tend to focus on the surface shape of the words rather than the deep meaning, showing that they do not truly grasp the context in the way a human does. They can be confident and fluent even when they are wrong, generating false information with the same tone as a true fact.
Perhaps the most significant finding of the review is that these problems are not just isolated glitches. They point to a missing piece in the entire design of current artificial intelligence. The systems we have today are built from separate parts. One part handles language, another handles memory, and another handles decision-making. These parts are connected by engineers, but they do not talk to each other naturally or adapt together as a single unit. The study argues that we have not yet created a system where memory, learning, and decision-making are integrated into one continuous, self-modifying whole. The machine can do the job, but it does not have the unified mind that allows a human to navigate the world with flexibility and understanding.
The researchers are careful to say that this does not mean machines can never achieve this kind of intelligence. It simply means that the evidence we have right now does not show it. The current approach of making models bigger and training them on more data has not solved the problem. In fact, as these models grow larger, some of the gaps, like the difficulty in remembering old information while learning new things, seem to become more pressing. The study suggests that simply scaling up what we have now is not enough. To bridge the gap, we may need to rethink the very architecture of these systems, perhaps finding ways to make memory and learning intrinsic to the machine rather than added on as tools.
This work provides a clear map of where we stand. It shows that while we have built machines that can mimic human behavior in many impressive ways, we have not yet built a machine that thinks like a human. The difference is not just a matter of getting the answer right; it is about how the answer is reached. Until we can create a system where learning, memory, and action are woven together into a single, adaptive fabric, our artificial creations will remain powerful tools that simulate intelligence, rather than true examples of it. The path forward requires more than just bigger computers; it requires a deeper understanding of how to build a mind that can truly integrate the world it experiences.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.