Why don't humans get stuck in infinite loops? Neuroscientist Anil Seth on the infinite-loop limitations of AI

**Key points**

**1. Algorithms cannot guarantee escape from infinite loops**

In 1936, Turing proved that no algorithm can determine whether an arbitrary program will halt; the frame problem shows that machines struggle to exhaust irrelevant information, and layer upon layer of self-checking can always fail.

**2. Human intelligence is rooted in time and entropy**

Life relies on metabolism and the free energy principle to reduce entropy and statistical surprise; survival pressure becomes a selection mechanism, enabling humans to almost always find a way out in an open world, rather than relying on recursive self-checking.

**3. Time in classical computation is thin and energy-hungry**

Digital computation concerns only state transitions and ignores the thermodynamic arrow; maintaining 0s and 1s requires error correction, leading to enormous AI energy consumption, and it may terminate because time or battery power runs out first.

**4. Consciousness may support open-ended intelligence**

Intelligence is about doing, consciousness is about being or feeling; the author speculates that some of the intelligence needed to escape infinite loops depends on consciousness embedded in time, and related theories also link consciousness with open-ended learning.

**5. Alternative computational paths are promising but not guaranteed**

Analog computation, mortal computation, neuromorphic computing and cybernetics place more emphasis on time and may improve energy efficiency; but the author doubts they can still achieve biological-style open-ended intelligence, and the imminence of artificial general intelligence and AI consciousness seems more like a pipe dream. 2026-10-23 12:33 Why don't humans get stuck in infinite loops? NeuroReality Quick Read

This article is from the WeChat public account: NeuroReality, author: Anil Seth, translation: EY

We landed on time at Madrid-Barajas Airport, but could not disembark for a long time. It turned out that the boarding bridges here were controlled by a new artificial intelligence system. No human intervention was needed—until intervention became necessary. Through the porthole, I saw the AI-controlled bridge repeatedly approach the aircraft, sway gently in the morning air, make futile attempts to dock, and then retreat. Over and over, it fell into an infinite loop. Finally, a staff member came to the rescue, spotted the problem at a glance, and fixed it immediately.

Humans do not endlessly repeat the same behavior, but AI sometimes does: it may get stuck on the apron, or trapped in the purgatory of an automated call center. These examples may seem like mere insufficient intelligence, but I believe they point to a deeper problem: a fundamental limitation that may make even the smartest AI forever unable to escape the risk of falling into an infinite loop, neither aware of the problem nor able to get out.

Imagine a smarter boarding bridge: a system able to monitor its own performance, detect when it is making futile attempts and when it has failed to complete its task. With this internal model of the environment and its own behavior, it could do what today's automated bridges cannot. But it too would eventually malfunction. So you add another, higher-order feedback loop: a model within the model that monitors its own monitoring process. What if that fails too? Then add another layer. And so on recursively; each added layer adds some intelligence and makes the system more robust. But there is always the possibility of another failure, causing the layer upon layer of self-monitoring system to err. Unless you build an infinite hierarchy of recursive self-checking mechanisms, the system will eventually fall into an infinite loop, never able to satisfy the stopping condition, and everything collapses.

Computer scientists have grappled with this problem for nearly a century. In 1936, Alan Turing proved that no algorithm can determine, in all cases, whether another algorithm will halt or run forever given a certain input. The following example helps make this intuitive. Suppose you are writing code to direct a robot's actions. Some programs are simple: the robot completes the task, and the algorithm ends. But other programs are complex, containing loops and conditions: "Keep moving until you see a red object." If the robot never sees a red object, or the program logic is flawed, it may execute instructions endlessly. No matter how advanced your code, it may be impossible to predict whether it will halt.

Turing's insight is abstract, but in the so-called "frame problem," the limitations of algorithms become more concrete. John McCarthy and Patrick J. Hayes proposed this problem in 1969. Broadly speaking, the frame problem refers to the difficulty of getting a machine to make sensible decisions based on relevant information without having to explicitly consider every irrelevant detail. In any situation, irrelevant information may be endless, so it is hard for a machine to know which information should not be taken into account. This opens the door to another kind of infinite loop.

These problems were once important milestones in thinking about natural and artificial intelligence. But as wave after wave of deep neural networks and generative models swept through society, they mostly receded into the background. That does not mean the problems have disappeared. The boarding bridge in Madrid shows that when silicon chips actually arrive on the apron, there is still much that can go wrong. AI systems may always face the risk of infinite loops, not because of insufficient computing power, but because of a deeper difference between such systems and our embodied biological brains.

This difference begins with the relationship between computers and brains and time, and ultimately leads to the everyday miracle of consciousness.

Intelligence and consciousness are two different concepts, although some technology leaders mistakenly believe that an increase in the former will necessarily bring about the latter. Put simply, intelligence is about doing, while consciousness is about being or feeling. Still, the two are connected, at least in humans and other animals. My hypothesis is this: certain forms of intelligence—specifically, the intelligence needed to completely escape the threat of infinite loops in a world where possibilities have no preset boundaries—depends on the capacities of a conscious mind, and that mind is deeply embedded in the flow of time.

Rooted in time and entropy

Continuously stacking layers of self-monitoring may be enough to build a satisfactory boarding bridge, but humans do not cope with complexity this way. Living, conscious creatures like you and me can survive and even thrive in environments far less predictable than an airport without adding ever more bizarre circuits layer by layer. In fact, the human mind, wondrous as it is, is not good at handling recursion. Most of us can manage at most three levels: knowing that we know that we know. For example, sometimes I am confident that I am capable of judging when I am right and when I am wrong. Our astonishing adaptability in open worlds must seemingly rest on something else.

What might this "something else" possibly be? One possibility is that our minds, brains, and conscious experiences are in some way rooted in time and entropy, whereas algorithms are by design cut off from such roots.

We humans are also animals, and like other living things, we must constantly resist the tendency toward decay and disorder dictated by the second law of thermodynamics. This law states that the entropy of an isolated system can only increase or stay the same; entropy is a measure of disorder or randomness. A drop of ink diffuses through a glass of water and never spontaneously gathers itself back together; a broken egg does not restore itself to wholeness. This law is awe-inspiring, as the physicist Arthur Eddington said: "If your theory is found to be against the second law of thermodynamics I can give you no hope; there is nothing for it but to collapse in deepest humiliation."

A key point about the second law of thermodynamics is that it applies to isolated systems, such as a sealed thermos, where there is no energy exchange between the system and its surroundings. In open systems that can exchange energy with their environment, entropy can decrease. Living systems are a classic example of open systems: they use energy from the environment to keep themselves in the statistically improbable state of being alive.

This process unfolds in different ways and at different levels. The cells in our bodies constantly recreate the conditions that sustain their own existence through metabolic energy flows. According to some important views in neuroscience, especially the free energy principle, neural circuits continuously work to minimize the statistical surprise of sensory input, reducing the entropy of that input and thereby exerting a kind of predictive control over the body's physiological state.

Conscious experience also directly reflects this fundamental drive to sustain life. Every conscious experience gathers a large amount of survival-related information to guide our behavior, whether or not that behavior is voluntary. Moreover, arguably, all these experiences carry a trace of valence, that is, a more or less positive or negative feeling that may prompt us to take certain actions. As I wrote in the book Being You, we experience the world around us, and ourselves within it, through, by means of, and because of our living bodies.

These processes across different scales are tightly interwoven, and each is closely tied to time, though in different ways. Metabolic processes unfold on the biochemical timescale: perhaps a billion reactions occur every second in each cell of our bodies. Neural activity has a slower timescale, but it is likewise constrained by neurobiological limits and adaptive mechanisms: from myelin-accelerated action potential transmission to the slow diffusion of neurotransmitters. In the stream of consciousness, the passage of time is continuous, irresistible, and complex, encompassing both sequence and flow, duration and persistence. Time in the body, brain, and mind is rich, dynamic, and multidimensional; every level, from biochemistry to personal identity, is deeply interdependent. Our entire way of being is inevitably embedded in physical time; and many physicists believe that physical time itself is grounded in the second law of thermodynamics.

The contrast between brains and computers is stark and thought-provoking. For a digital computer, time is thin and one-dimensional, and its thermodynamic arrow is abstracted away. Computation is only about state transitions: A becomes B, 0 becomes 1. In Turing's classic formulation of computation, only order matters, not the actual dynamic changes of the underlying material substrate. Every algorithm is nothing but one damn state after another. Between two steps of an algorithm, that is, between A and B, there may be a microsecond or a million years; it is still the same algorithm, the same computation.

This disregard for time comes at a high cost. The computation formalized by Turing can exist outside time, but computers themselves cannot. Today's AI's insatiable appetite for energy stems largely from the need for error correction: 1 must be kept as 1, and 0 as 0, because even silicon chips made of inanimate sand cannot escape the tentacles of entropy. Pretending that time does not exist is an expensive business.

This is precisely the key reason why I believe algorithms can get stuck in infinite loops, while intelligent animals almost never do. Algorithms are frozen within the space of sequences, detached—at least in principle—from the pull of entropy. Therefore, there will always be some unforeseen and unforeseeable circumstance that causes time itself to end before the computation does—or, more likely, that causes the electricity to run out first.

Unlike computers, we are beings that exist in time: we have bodies, we are embedded in the world, and we are embedded in the time within it. We can never be trapped in an infinite loop, because we have never existed outside of time. For creatures like us, the constant, time-pressured demand—to minimize statistical surprise and maintain physiological viability—is the ultimate relevance-filtering mechanism, and it is also the reason we almost always find a way out.

I say "almost always" because many repetitive behaviors do indeed carry a certain loop-like character, though not infinite loops. These examples include perseverative repetitive behavior after frontal lobe damage, obsessive-compulsive disorder, repetitive movements in severe autism and Tourette syndrome, and even self-destructive addictive behaviors. In a sense, these examples actually support my argument, because they all reflect disturbances of "normal" neurocognitive function.

If these ideas are on the right track, then AI systems that rely on Turing computation will forever lack certain intelligent behaviors rooted in the body and in time. If this holds, then artificial general intelligence—intelligence capable of matching or surpassing human performance on any cognitive task—is even more of a pipe dream than its already rather illusory prospects suggest.

Can new forms of AI connect with the flow of time in richer ways than classical computation allows? There is both a long history and an exciting future in this regard. The earliest computers, such as the 2,000-year-old Antikythera mechanism used for astronomical prediction, were analog and operated in continuous time. At the other end of history, researchers are exploring approaches such as "mortal computation," which allow the computational process to depend on the specific hardware that executes it: when the hardware "dies," the algorithm dies with it. This approach works with, rather than against, the changes brought by the second law of thermodynamics, and promises substantial improvements in energy efficiency. In addition, neuromorphic computing takes many forms, which imitate the workings of the human brain to varying degrees and are more deeply rooted in time than Turing's time-independent computational paradigm.

Research approaches to intelligence based on dynamical systems avoid computation altogether and instead focus on concepts such as attractors and phase space; these concepts place greater emphasis on time. These research approaches can be traced back to the early days of AI development and to the long-neglected tradition of cybernetics. Cybernetics focuses on feedback and control rather than symbolic processing by algorithms.

These approaches are highly promising, and there are still many new forms of AI waiting to be discovered. But I still suspect that even time-embedded computation or cybernetic engineering cannot completely escape the shadow of infinite loops, or fully achieve the open, highly adaptive intelligence possessed by biological systems.

I believe that this kind of intelligence is not only closely related to time but also closely related to consciousness—perhaps even necessarily so. At least, this is the case when you understand consciousness, as I do, as a process deeply connected to the drive for survival. According to this view, conscious experience integrates processes at different levels, from metabolism's resistance to entropy all the way to the opportunities for action and planning available in perception, and it is always inseparably embedded in the flow of time. What really matters is this cross-scale integration. It may make conscious experience a currency with sufficient credit to help us escape almost any infinite loop and to reframe any frame problem.

Some other views also link consciousness with open-ended intelligence, though without focusing on time and entropy. Twenty years ago, cognitive scientist Murray Shanahan, inspired by the global workspace theory of consciousness, proposed that consciousness can combine elements of cognition and perception in different ways, thereby providing the resources needed to solve, or at least bypass, the frame problem. Theoretical biologists Eva Jablonka and Simona Ginsburg argue that consciousness is related to what they call unlimited associative learning. This is a special form of learning that enables organisms to discover new solutions without being limited by preset boundaries. Taken together, a picture is emerging: certain forms of intelligence, at least in biological organisms, and perhaps in any context, require consciousness.

We might also briefly look at this from another angle. If conscious experience depends on being deeply embedded in time, and also on the metabolic requirements of sustaining life in an entropy-dominated universe, then this will have a major impact on the future of AI. Specifically, the temporal thinness of classical digital computation—just one state after another, with the time intervals between states completely arbitrary—seems fundamentally incompatible with the nature of consciousness as a rich dynamic process. If consciousness is indeed inseparable from physical time, as it appears to be, then it cannot be merely a product of algorithms. This drives another nail into the coffin of the idea that "AI consciousness is imminent"—if another nail is still needed.

In his later years, the philosopher Daniel Dennett liked to talk about the "hard problems" in the science of consciousness, using this as a humorous contrast to David Chalmers's famous "hard problem" of consciousness. The latter asks: how exactly does consciousness arise? The problems Dennett referred to are all about the function of consciousness—not what consciousness is, but what consciousness can do. In his words, a conscious experience occurs—"and then what happens?"

Perhaps the answer lies in time. Consciousness may be nature's way of enabling the intelligent life it has created to keep moving forward and to (almost) always find a way out.

British neuroscientist and popular science writer, currently Professor of Cognitive and Computational Neuroscience at the University of Sussex, Director of the Sussex Centre for Consciousness Science, and co-director of the Canadian Institute for Advanced Research (CIFAR) program on Brain, Mind and Consciousness. His research focuses on the biological basis of consciousness and how the brain forms perception, the sense of self, and subjective experience, combining neuroscience experiments, computational modeling, and theoretical research to explore how we perceive the world and experience ourselves. He was the founding editor-in-chief of Neuroscience of Consciousness. His representative work Being You: A New Science of Consciousness introduces consciousness research to the general public and was selected for annual best book lists by media including The Economist. AI industry signal channel: Frontier Technology. This content comes from the internet, original link; the views represent only the author and do not represent the position of Huxiu. If there are copyright issues, please contact hezuo@huxiu.com, and we will verify and handle them promptly. Everyone who is changing and wants to change the world is on the Huxiu APP [ ]