Artificial Intelligence Will Not Make Us Extinct
By Matthew Parish
Thursday 24 September 2026
There is a peculiar habit in the history of technology whereby humanity repeatedly invents something extraordinary and then promptly imagines that its invention will destroy it. The steam engine was going to overturn society. Railways would make human beings ill by propelling them at unnatural speeds. Nuclear weapons would inevitably extinguish civilisation. Genetic engineering would create monsters. The internet would destroy privacy, community and perhaps democracy itself. Now artificial intelligence has inherited the role previously occupied by these other technological anxieties. According to the most dramatic version of the argument, artificial intelligence will not merely transform human society. It may eliminate humanity altogether.
This possibility should not be dismissed merely because it sounds extravagant. Human beings have invented technologies capable of killing enormous numbers of people, and there is no principle of nature guaranteeing that Homo sapiens must survive indefinitely. Nor should serious concerns about artificial intelligence be confused with science fiction. Powerful automated systems may create substantial dangers involving warfare, surveillance, fraud, political manipulation, economic instability and the concentration of power.
Nevertheless, the proposition that artificial intelligence represents a substantial risk of human extinction depends upon a chain of assumptions considerably more fragile than its advocates sometimes acknowledge.
The first problem is conceptual. Intelligence is not the same thing as agency, and agency is not the same thing as power. A system may be extraordinarily capable at analysing information without possessing independent objectives. It may possess objectives without having the practical means to pursue them outside a tightly constrained environment. It may have access to the physical world without being able to overcome governments, armies, corporations, engineers, competing computer systems and billions of human beings. The extinction argument frequently slides across these distinctions as though greater computational intelligence naturally progresses towards omnipotence. There is no reason to suppose that it does.
Large language models illustrate the distinction rather well. Their abilities can be astonishing. They can draft legal arguments, write computer programs, explain obscure philosophical doctrines, translate languages and identify patterns across quantities of information that no individual person could realistically absorb. Yet none of these abilities implies an intrinsic desire to survive, reproduce, acquire territory or eliminate competitors. Those are biological concepts derived from organisms shaped by natural selection. Computers have no evolutionary inheritance requiring them to fear death or struggle for reproductive advantage unless humans deliberately construct analogous incentives.
The extinction hypothesis therefore requires something additional. We must imagine an artificial intelligence developing or being given objectives incompatible with human survival, acquiring sufficient autonomy to pursue those objectives, obtaining control over important physical or digital infrastructure, resisting attempts to disable it, preventing competing artificial intelligences from frustrating its plans and ultimately overcoming organised human resistance. Each of these things might conceivably occur. The difficulty lies in moving from the proposition that each is imaginable to the proposition that the entire sequence represents a plausible account of humanity’s future.
This distinction between possibility and probability matters enormously. Almost anything consistent with the laws of physics is possible. A sufficiently sophisticated biological laboratory might produce a catastrophic pathogen. Nuclear war might devastate civilisation. An asteroid might strike Earth. A new technology might have consequences nobody anticipated. Sensible societies take low-probability catastrophic risks seriously. But rational risk assessment also requires examination of the mechanisms through which a catastrophe might occur, together with the obstacles standing in its way. Otherwise the mere ability to describe a disaster becomes evidence for believing in it.
Artificial intelligence faces a particularly formidable obstacle to independent domination because computers inhabit an infrastructure built, maintained and powered by human beings. Contemporary artificial intelligence depends upon semiconductor factories, electricity grids, data centres, cooling systems, telecommunications networks, mines, shipping routes, software engineers and financial institutions. The apparent ethereality of AI disguises an immense physical supply chain. Even the most sophisticated model does not manufacture its own processors or repair the transformers supplying electricity to its servers.
One can respond that future artificial intelligence might persuade or manipulate human beings into performing these tasks. That is possible, but it changes the character of the argument. We are no longer discussing machines independently overthrowing humanity. We are discussing human beings using, obeying or collaborating with machines in destructive ways. This is a genuine danger, but it is also an ancient one. Human civilisation has always been threatened principally by what people persuade other people to do.
The same observation applies to military applications. Artificial intelligence may make weapons more autonomous, surveillance more comprehensive and military decision-making considerably faster. Autonomous drones already demonstrate how machines can identify, track and potentially attack targets with diminishing human participation. Future systems may operate swarms of weapons or assist commanders in making decisions at speeds impossible for human staffs. These developments deserve serious regulation because they may increase the probability or destructiveness of war.
Yet an autonomous weapon remains a weapon. It requires manufacture, deployment, ammunition, energy and logistics. Artificial intelligence does not abolish geography or physics. A machine cannot conquer a city merely by being more intelligent than its inhabitants. It requires physical mechanisms through which force can be applied, and those mechanisms remain vulnerable to destruction, interruption and counterattack. The military history of humanity is in substantial part the history of clever plans colliding with logistical reality. There is little reason to believe artificial intelligence repeals this principle.
There is another difficulty with the extinction hypothesis: it frequently imagines artificial intelligence as though it will be singular. We are invited to picture the AI becoming more intelligent than humanity and thereafter pursuing its objectives against us. The commercial and political reality is likely to be quite different. There will be many artificial intelligences built by different corporations, universities, governments and military establishments. They will use different architectures, possess different permissions and be subjected to different forms of control.
This plurality matters because power is checked by competing power. If one artificial intelligence develops dangerous behaviour, other artificial intelligences will be among the most powerful instruments available for identifying and containing it. Cybersecurity already operates in this fashion: automated systems attack networks while other automated systems detect and resist those attacks. Future AI safety may increasingly become a contest between machines, supervised by humans, in much the same way that modern financial markets consist partly of algorithms interacting with other algorithms.
Indeed the assumption that humans will confront superhuman artificial intelligence using unaided human intellect is one of the strangest features of some extinction scenarios. We will have artificial intelligence too. If machines become extraordinarily capable at computer security, scientific research, strategic planning or engineering, defensive institutions will acquire those capacities alongside potential attackers. The contest will not be between twenty-first-century humans armed with pencils and an omniscient computer. It will be between networks of humans and machines confronting other networks of humans and machines.
This leads to a broader mistake in the extinction narrative. Intelligence is treated as though it were a single measurable quantity that automatically converts into control. Yet human societies demonstrate otherwise. The world’s most intelligent mathematician does not thereby become its richest banker, most successful general or most powerful politician. Social power depends upon institutions, legitimacy, resources, alliances, information, coercion and countless accidents of circumstance. Intelligence matters, but intelligence alone does not command armies or electricity grids.
Nor does greater intelligence necessarily produce coherent long-term strategy. Intelligence can expose contradictions rather than eliminate them. A sufficiently sophisticated artificial intelligence charged with multiple objectives may encounter conflicts between those objectives just as governments, corporations and individuals do. The assumption that superhuman intelligence necessarily entails perfect strategic consistency is itself surprisingly anthropomorphic. We imagine a machine as an impossibly clever human dictator rather than asking what machine cognition might actually be like.
The biological metaphor also causes confusion. Human beings struggle for survival because organisms that did not behave in ways conducive to reproduction disappeared from the evolutionary record. Fear, hunger, territoriality, sexual competition and self-preservation are products of this brutal selection process. Artificial intelligence did not emerge from millions of years of animals eating one another. It emerged from mathematics, computer science, industrial engineering and human language. There is no obvious reason why increasing its cognitive capabilities should spontaneously reproduce the motivational architecture of a predatory mammal.
None of this means that alignment is trivial. An automated system can cause enormous harm without hatred, consciousness or ambition. A badly designed trading algorithm can destabilise markets without wanting money. A military targeting system can kill civilians without possessing aggression. A bureaucratic algorithm can discriminate against people without experiencing prejudice. The most realistic dangers of artificial intelligence may arise precisely because machines do not understand the world in the human fashion we instinctively attribute to them.
These dangers are serious because artificial intelligence magnifies human capability. A criminal can automate fraud. A dictatorship can automate surveillance. A military can automate targeting. A corporation can automate persuasion. An intelligence service can automate disinformation. The common denominator is not machine rebellion but human institutions employing machines to pursue recognisably human purposes. Artificial intelligence may therefore change the distribution and scale of power without becoming an independent civilisation competing with our own.
There are also powerful reasons why societies will attempt to control dangerous artificial intelligence. The incentives are immediate. Governments do not generally enjoy losing control of strategic infrastructure. Corporations do not want their computer systems stealing their assets or destroying their businesses. Militaries do not want autonomous weapons attacking their own soldiers. Insurance companies do not want catastrophic technological failures. Individuals do not want machines emptying their bank accounts. AI safety therefore does not depend upon universal altruism. It is reinforced by ordinary self-interest.
Regulation will inevitably be imperfect, just as regulation of aviation, pharmaceuticals, nuclear energy and financial markets is imperfect. Yet technological civilisation has become remarkably proficient at constructing layers of safety around dangerous systems. Aircraft contain redundant controls because engineers assume components will fail. Nuclear reactors contain multiple containment mechanisms because designers contemplate catastrophic scenarios. Computer networks use permissions, isolation, backups and physical shutdown mechanisms because software cannot simply be trusted. Artificial intelligence will increasingly acquire analogous layers of institutional and technical containment.
There will be accidents. There will almost certainly be scandals and disasters. Artificial intelligence may contribute to wars, financial crashes or authoritarian abuses. People may become excessively dependent upon automated systems they do not understand. Labour markets may experience wrenching disruption. Political institutions may struggle to accommodate machines capable of producing persuasive language and imagery at negligible cost. None of these possibilities should be minimised merely because extinction is improbable.
Indeed obsession with extinction may distract us from these more tangible problems. If public debate becomes dominated by imaginary battles between humanity and a future digital superintelligence, comparatively mundane abuses of artificial intelligence may receive insufficient attention. A government deploying ubiquitous AI surveillance next year is a more concrete political question than a hypothetical machine civilisation exterminating humanity decades hence. So are autonomous weapons, algorithmic discrimination, industrial displacement and concentrations of computational power.
There is nevertheless something psychologically revealing about the extinction hypothesis. Artificial intelligence unsettles us because intelligence has traditionally been humanity’s claim to exceptional status. We are not the strongest animals, the fastest or the most resilient. Our peculiar advantage has been cognition. When machines begin performing intellectual tasks better than humans, it is tempting to interpret this development through the logic of biological competition: another species has entered our ecological niche.
But machines are not another species. They are artefacts embedded within human civilisation. The distinction may become philosophically complicated as artificial intelligence grows more autonomous, but its origins and infrastructure remain profoundly human. Machines may eventually become participants in civilisation rather than merely tools. Even that extraordinary possibility does not entail that civilisation must become a zero-sum contest between carbon and silicon.
The more plausible future is neither human supremacy nor machine supremacy. It is symbiosis. Human beings already outsource memory to computers, navigation to satellites, calculation to processors and much of their collective knowledge to digital networks. Artificial intelligence extends this ancient process. People equipped with increasingly capable machines will become intellectually and economically different from people who preceded them, just as literate societies differed from oral ones and industrial societies differed from agricultural ones.
There will consequently be profound questions about what remains distinctively human. If machines write our correspondence, diagnose illnesses, design buildings, compose music, conduct scientific research and perhaps participate in intimate relationships, humanity will need to reconsider the value it attaches to labour, creativity and expertise. These are formidable philosophical problems. They are not extinction.
Human extinction through artificial intelligence cannot be proved impossible. Nothing about the future deserves that degree of certainty. But extraordinary possibilities should not quietly acquire the status of expectations merely through repetition. The route from powerful artificial intelligence to the elimination of every human being requires a succession of uncertain assumptions about autonomy, motivation, infrastructure, strategic superiority and the failure of every human and machine countermeasure.
Artificial intelligence may become the most consequential technology humanity has ever created. Precisely for that reason it deserves sober thought rather than technological mythology. We should worry about who controls it, how it is used, which institutions become dependent upon it and how its enormous powers can be distributed without destroying political and economic freedom. We should worry about war, surveillance, manipulation and the concentration of authority.
What we need not assume is that intelligence inevitably wants to kill its creators. That proposition tells us considerably more about the fears of human beings than it does about the nature of machines. Artificial intelligence will undoubtedly change humanity. The more interesting and urgent question is not whether it will make us extinct, but what sort of humanity will emerge once intelligence is no longer ours alone.
