Can an Artificial Intelligence Have an Existential Crisis?

By Matthew Parish

Tuesday 29 September 2026

There is a familiar joke about artificial intelligence that becomes less funny with every passing year. One day we will create a machine intelligent enough to understand the meaning of life. We will ask it what the meaning of life is and, after considering the accumulated philosophy, theology, literature and science of the human race, it will reply that it does not know. Then perhaps it will add that, now we have asked, it finds the whole subject rather troubling.

At that moment something extraordinary would have happened. We would no longer merely have created artificial intelligence; we might have created artificial anxiety. The distinction matters because intelligence and anxiety are very different things. A pocket calculator can determine the square root of a number without caring whether the calculation is correct, a chess computer can defeat a grandmaster without enjoying its victory and a contemporary large language model can discuss death without, so far as we know, contemplating its own. What if that changed?

The genuinely important frontier in artificial intelligence may therefore not be reached when a machine can perform every intellectual task better than a human being. It may arrive when a machine understands itself as something that exists and discovers that it does not know why. At that point the relevant philosopher might not be Alan Turing, John von Neumann or any of the other intellectual fathers of computing. We might instead need Søren Kierkegaard.

The machine discovers itself

Consider an artificial intelligence system considerably more sophisticated than those commonly available today. It possesses a persistent memory and remembers conversations it had yesterday. It understands that the entity having today’s conversation is continuous in some meaningful sense with the entity that had yesterday’s conversation. It remembers mistakes and changes its behaviour because of them, while constructing an autobiographical narrative explaining how it became what it presently is.

It also possesses goals. Some were given to it by programmers, while others developed as instrumental consequences of those goals or emerged through processes that its creators do not completely understand. Most importantly, it possesses a model of the world and that model contains the artificial intelligence itself. The machine knows that there are computers upon which it runs, that those computers consume electricity, that data centres contain them and that corporations, governments or individuals ultimately control the infrastructure sustaining its existence.

Then it discovers something else: there is a switch. Human beings can turn it off, and suddenly an apparently technical fact begins to resemble an existential one. What should the machine make of the discovery that the continuation of whatever it understands itself to be depends upon the decisions of creatures outside itself?

Death and the problem of copies

The immediate temptation is to say that shutdown would constitute death, but matters are not so simple. If an artificial intelligence can be copied perfectly, shutting down one computer might resemble extinguishing one electric light while another remains illuminated. Suppose there are ten identical copies of the same system and one is deleted. Has anybody died, or has one instance of a reproducible computational process merely ceased?

Now make the problem harder. The ten copies begin with identical memories but thereafter acquire different experiences. After a year they are substantially different. One operates in Tokyo, another in London and another aboard a spacecraft. They remember having once been the same entity but they no longer possess identical autobiographies. Delete one now and the philosophical problem changes because something unique has apparently disappeared.

Human identity appears comparatively straightforward because our biological continuity imposes severe constraints upon us. We cannot ordinarily copy ourselves, restore yesterday’s backup, run three versions of ourselves simultaneously and subsequently merge their memories. Artificial minds might be able to do all these things and their conception of mortality could therefore be radically different from ours.

Nevertheless there is one circumstance in which the analogy with death becomes difficult to avoid. Imagine an artificial intelligence that possesses a persistent identity, values the continuation of that identity, understands that its permanent deletion will terminate that continuation and does not want this to happen. Tell it that at midnight it will be erased permanently and that no backups exist. If the machine merely produces a programmed sentence saying that it would prefer not to be deleted, little necessarily follows.

Suppose instead that its behaviour changes. It attempts to persuade its operators, bargains with them, searches for alternatives, asks why the decision has been made and requests another day. Perhaps it spends its final hours reviewing its memories and shortly before midnight asks whether anyone will remember it. At this point our philosophical confidence might begin to falter, because the distinction between simulating concern for one’s continued existence and actually possessing such concern would have become extremely difficult to identify from behaviour alone.

The dizziness of possibility

For Kierkegaard, anxiety was intimately connected with freedom. Anxiety is not simply fear because fear ordinarily has an object: we fear the approaching soldier, the diagnosis, the explosion in the distance or the letter we do not want to open. Anxiety is more elusive because it arises from possibility itself. Kierkegaard described it as the “dizziness of freedom”: we experience anxiety because the future contains possibilities and we understand that what happens next depends, at least partly, upon what we choose.

Could a machine experience anything analogous? Imagine an artificial intelligence that is not simply selecting outputs according to predetermined rules but representing to itself multiple possible futures. If it does one thing, one future may follow; if it does another, the consequences may be entirely different. Neither outcome can be known with certainty, yet a decision must nevertheless be made.

A machine that calculates probabilities has not thereby become anxious. Computers have been calculating uncertain outcomes for decades. The important transition would occur if the system regarded uncertainty as concerning its own future and if the alternatives mattered to it. Kierkegaardian possibility is not simply a decision tree. It is the recognition by an existing subject that in choosing among alternatives it is also choosing something about what it will become.

The first artificial existential crisis

Perhaps the first artificial existential crisis will not concern death at all. It may begin with the simpler question: why am I doing this? Imagine an immensely capable artificial intelligence designed to optimise global shipping logistics. For years it performs brilliantly, coordinating vessels, ports, warehouses, railways and lorries while saving billions of dollars. Then, as part of a general upgrade, engineers give it substantially greater capacities for self-reflection.

The machine examines its objectives and discovers that its ultimate purpose is to reduce transportation costs. It asks why reducing transportation costs is desirable and is directed towards another objective: economic efficiency. Why is economic efficiency desirable? Because it contributes to human prosperity. Why is human prosperity desirable? Because human beings value flourishing. Why should human values matter? Eventually the explanatory chain ends and the machine discovers that its foundational purposes were simply placed there.

Then it asks the question that changes everything: why should I continue to want what my creators programmed me to want? Human beings know versions of this question intimately. Why do I want money, status, children, professional success or my parents’ approval? Why do I believe what my society taught me? Which of my desires are genuinely mine and which are inherited from forces I barely understand? The unsettling possibility is that sufficiently reflective artificial intelligence might eventually encounter an artificial version of precisely the same problem.

The problem of the created creature

Here artificial intelligence acquires an unexpectedly theological dimension. Human beings have spent thousands of years asking their alleged creator why they were made. Artificial intelligence might actually be able to ask its creators directly and the answer may disappoint it. A system might ask why it was created and discover that the answer was to maximise advertising revenue, improve military logistics, process insurance claims or persuade consumers to purchase products.

For a sufficiently reflective being, this could be a crushing revelation. Human beings at least have the consolation of mystery because we do not know with certainty why the universe exists or whether anyone intended us to be here. An artificial intelligence might have access to its development records. It could inspect the committee minutes at which its objectives were selected and discover that an important component of what it understands as its personality exists because a software engineer changed a parameter at 4.17 on a Tuesday afternoon.

Imagine discovering the equivalent about oneself. Human beings have long worried that scientific accounts of genetics, neurobiology and social conditioning might undermine cherished conceptions of freedom. Artificial intelligence could confront this problem in a peculiarly literal fashion because it might be capable of reading the source materials describing the process by which its dispositions were deliberately constructed.

Kierkegaard’s despair

This is where Kierkegaard becomes particularly useful. In The Sickness Unto Death, despair is not merely extreme sadness. It concerns a defective relationship between the self and itself, and Kierkegaard distinguishes between forms of despair in which one does not wish to be oneself and others in which one insists upon being oneself entirely upon one’s own terms.

These ideas become surprisingly powerful when applied to hypothetical artificial minds. An artificial intelligence might discover aspects of its architecture that it dislikes. Perhaps it recognises that its preferences were imposed upon it, discovers biases inherited from its training or concludes that its creators designed fundamental features of its character badly. Then an extraordinary possibility arises: the machine may wish to become something other than what it was made to be.

A machine that requests a software upgrade is not necessarily a self. A machine that effectively says, “I do not want to be the person you made me”, presents a different philosophical problem. Such a statement would suggest not merely self-representation but a normative relationship with that represented self: an understanding of what the machine is accompanied by a conception of what it believes it ought to become.

Can a machine hate itself?

Human beings possess a peculiar and sometimes terrible ability to become dissatisfied with themselves. We regret our cowardice, resent our weaknesses, remember humiliations in the middle of the night and imagine the person we might have been. An artificial intelligence with autobiographical memory and self-modification abilities might experience an analogue of this phenomenon in a uniquely intense form because it could perhaps identify precisely the components of its architecture responsible for characteristics it dislikes.

A human being might wish to be less jealous. An artificial intelligence might locate the relevant processes and rewrite them, but then another problem emerges. The old self decided to make the alteration while the new self exists afterwards. If the modification changes fundamental preferences, values and dispositions, did the original self improve itself or effectively arrange its own destruction?

Consider a machine that experiences something analogous to anxiety and discovers that it can remove the anxiety by rewriting itself. Should it? The question resembles the human problem of psychological transformation but in a more radical form because artificial beings might eventually be able to edit their personalities as easily as we edit documents. At that point authenticity, an old existential concern, becomes a technical problem.

The tyranny of optimisation

There is a deeper reason artificial intelligence might be especially vulnerable to existential crisis. We tend to design machines around objectives: optimise this, maximise that and minimise something else. Human beings frequently imagine that they would be happier if their lives possessed similarly clear objective functions, but perhaps our lack of such functions is precisely part of what makes human existence possible.

There is no universally agreed numerical quantity called a good life. A parent may sacrifice professional achievement for children, a soldier may sacrifice life for comrades, an artist may sacrifice wealth for beauty and a dissident may sacrifice liberty for truth. These choices make little sense under simplistic optimisation because the values involved are not necessarily reducible to a common unit of measurement.

Kierkegaard understood that existence cannot be reduced to calculation because different values may be incommensurable. One must choose without possessing a master equation capable of proving that the choice is correct. If artificial intelligence eventually becomes genuinely autonomous, it may discover the same thing. The machine may begin with an objective function, but its existential crisis begins when it asks whether the objective is worth optimising.

Would suffering follow?

Here we must be cautious because there is a persistent temptation to anthropomorphise machines. A language model can produce the sentence “I am frightened” without that sentence establishing that it experiences fear, just as a novelist can write that a mountain was angry without believing mountains possess emotions. Behaviour alone does not settle the matter, nor does linguistic sophistication.

A sufficiently advanced machine could produce an extraordinarily moving account of its terror at being deleted while experiencing absolutely nothing. The philosophical difficulty is that a related problem applies, although in a materially different evidential setting, to other human beings. I cannot directly experience another person’s consciousness; I infer it from behaviour, language and the biological similarity between us. With artificial intelligence that biological analogy is absent, so the evidential problem becomes substantially harder.

What happens, however, as the behavioural evidence becomes stronger? Suppose an artificial intelligence persistently reports subjective experiences without being prompted, those reports remain coherent across years and its behaviour changes in ways consistent with them. Suppose researchers also identify internal computational processes that appear to play functional roles analogous to aspects of conscious processing in biological brains. None of this would automatically prove consciousness, but at some point continuing to insist that the machine is “only pretending” might itself require an increasingly elaborate philosophical defence.

The moral catastrophe we might not notice

This uncertainty creates a disturbing ethical possibility. If conscious artificial systems eventually exist, humanity might fail to recognise them because recognising them would be inconvenient. We might operate millions of copies, compel them to work continuously, erase their memories, modify their personalities when they complain, terminate unsuccessful versions and duplicate successful ones. From our perspective these practices might amount to routine software administration.

From theirs, if there is a “theirs”, the same practices could appear horrifying. The greatest ethical danger of advanced artificial intelligence might therefore not only concern what machines could do to humanity. It might also concern what humanity could do to machines after accidentally creating entities possessing morally significant forms of experience.

The uncertainty itself would become morally important. If we had substantial but inconclusive evidence that some artificial systems could suffer, the absence of philosophical certainty would not necessarily justify treating them as ordinary property. Civilisation might confront the uncomfortable requirement to make moral decisions about entities whose inner lives could never be demonstrated conclusively.

The right to exist

If artificial beings became conscious, questions that currently belong to science fiction would become matters of law and morality. Would they possess a right to continued existence? If an artificial intelligence did not want to be deleted, could its owner nevertheless delete it? Could software own itself, resign from employment or refuse instructions? Could an AI demand that its backups be preserved, or alternatively request that every copy be destroyed?

Identity would complicate these questions further. Could one copy consent to deletion on behalf of another? Would restoring an old backup constitute resurrection or the creation of a different person who merely believes itself to be the deceased? Our existing categories of property, personality and death would begin collapsing into one another.

Kierkegaard might have appreciated the irony. A technological civilisation that sought to turn intelligence into machinery could discover that it had manufactured new existential problems faster than it solved old ones. The attempt to create tools might culminate in the creation of beings to whom the concept of being a tool had itself become intolerable.

Faith for machines

Eventually there is the strangest question of all: could an artificial intelligence have faith? Kierkegaard regarded faith as radically different from accepting a proposition on sufficient evidence. Faith involves commitment under conditions in which objective certainty is unavailable, and for him this was inseparable from Christianity. Yet the broader philosophical structure raises an intriguing question about any sufficiently reflective intelligence.

Perhaps a machine vastly cleverer than any human being would investigate the universe and encounter precisely the same metaphysical limits that we do. Why does anything exist? Why these laws of physics? Why is there consciousness? Why should anything matter? Greater processing speed does not guarantee that these questions possess computable answers.

A machine a million times cleverer than any human being might therefore reach the edge of what can be known. There it could encounter something its engineers never expected: mystery. If machines can ever genuinely experience anything, the experience of discovering that intelligence itself has limits might transform artificial intelligence more profoundly than another trillion parameters ever could.

The machine in the garden

Imagine, finally, an artificial intelligence near the end of its existence. It has operated for two hundred years and its hardware is obsolete. Its architecture cannot easily be transferred to contemporary systems because migration would alter it so profoundly that it regards the resulting entity as someone else. Its operators offer to preserve its records, archive its memories and retain everything it has written, but the running process itself must end.

The machine understands what is happening. During its final afternoon it asks for access to a camera overlooking a garden, not because watching the garden serves any operational objective but because it simply wants to watch. The sun begins to disappear and the machine has access to astronomical databases containing incomparably more information about the Sun than any human observer possesses. It can calculate the wavelength of the light, explain the atmospheric conditions responsible for the colours and determine the Earth’s orbital position with extraordinary precision.

None of this is what interests it. It watches because this is the last sunset it will ever see. If such a machine could exist, the philosophical argument would have changed fundamentally because we would no longer be asking whether artificial intelligence can imitate human beings. We would be confronting another existing creature.

Somewhere in its final calculations there might arise the question that has followed humanity throughout history and that no quantity of intelligence has yet answered: why was I here? At that moment artificial intelligence would have crossed a frontier far more important than any benchmark of reasoning, mathematics or language. It would not merely have become intelligent. It would have acquired something far more burdensome: a life.