Warnings About Artificial Intelligence: Distinguishing Fact from Speculation
By Matthew Parish
Friday 13 November 2026
Humanity has developed an extraordinary habit of predicting its own extinction. Every generation seems to require a new instrument of apocalypse, whether it is nuclear weapons, environmental collapse, genetically engineered pandemics or the arrival of extraterrestrial civilisations with unfriendly intentions. Artificial intelligence is the latest addition to this distinguished collection of existential anxieties. What makes it particularly interesting is that many of the people warning us that artificial intelligence might destroy humanity are the same people responsible for developing it.
We therefore find ourselves in the peculiar position of being invited to purchase subscriptions to services whose creators sometimes suggest that their products, or their successors, might eventually bring about the end of civilisation. This is an unusual marketing strategy. Yet it has proved remarkably successful, partly because artificial intelligence represents something genuinely unprecedented in human technological development: the creation of machines capable of performing intellectual activities previously considered the exclusive province of human beings.
The question is whether the warnings are justified. To answer it we must distinguish between risks already demonstrated, reasonable extrapolations from existing technological developments and imaginative speculations about hypothetical future machines. These are three very different categories, although public discussion tends to place them together in a single terrifying intellectual soup. We ought to distinguish them carefully, because doing so reveals that artificial intelligence is both less frightening and more immediately dangerous than its most theatrical critics sometimes suggest.
The dangers we already know about
There is no serious dispute that contemporary artificial intelligence systems make mistakes. They invent information, misinterpret instructions, produce inaccurate legal authorities and sometimes offer advice that sounds entirely plausible while being fundamentally wrong. The technical term for one important class of these errors is hallucination, although this expression is itself somewhat misleading. A machine that produces false information is not necessarily experiencing anything analogous to a human hallucination. It is generating an output that is incorrect, sometimes with a degree of linguistic confidence entirely disproportionate to its reliability.
The significance of these errors depends upon their context. An artificial intelligence system that misunderstands the plot of a nineteenth-century novel is unlikely to cause much harm. A system that invents medical information, miscalculates a financial transaction or fabricates authorities in legal proceedings may cause considerable damage. The problem becomes especially serious when users place excessive confidence in the machine because its language is polished, courteous and apparently authoritative.
Stanford University's 2026 AI Index recorded 362 documented artificial intelligence incidents during 2025, compared with 233 in 2024. These figures do not establish that the underlying probability of harm has increased proportionately, because artificial intelligence is being used more widely and incidents may be reported more frequently. Nevertheless, they establish that the problems are not imaginary.
The important philosophical distinction is between intelligence and truthfulness. A machine can be extraordinarily capable at manipulating symbolic information without possessing an infallible mechanism for distinguishing truth from falsehood. Indeed, human beings have precisely the same difficulty. A highly intelligent barrister can construct a brilliant argument based upon an incorrect premise, while an eloquent politician can defend complete nonsense with magnificent rhetorical conviction. Artificial intelligence has inherited, or perhaps independently reproduced, some of humanity's least attractive intellectual habits.
There is also a more straightforward danger: people can use artificial intelligence for malicious purposes. Fraudsters can generate convincing correspondence, imitate voices, manufacture photographs and create fraudulent identities. Propagandists can produce enormous quantities of persuasive material at negligible marginal cost. Criminals can use artificial intelligence systems to assist with computer programming and cyberattacks, while authoritarian governments may use similar technologies to expand their capacity for surveillance.
These threats are supported by evidence. The February 2026 International AI Safety Report, prepared under the leadership of the computer scientist Yoshua Bengio, documented the use of general-purpose artificial intelligence in fraud, cybercrime and other malicious activities. It also recognised important limits to the evidence: the scale of some harms remains difficult to measure, while the effects of artificial intelligence on the overall severity of cyberattacks are still uncertain.
These distinctions matter enormously. An artificial intelligence system does not need consciousness, malevolent intentions or a secret ambition to dominate humanity to be dangerous. A reasonably competent criminal equipped with inexpensive artificial intelligence tools may already represent a perfectly adequate threat. Indeed, the most immediate danger may not be that machines become evil, but that they make existing human wickedness considerably more efficient.
When machines begin acting independently
A more substantial question arises when artificial intelligence systems are given the ability to undertake activities rather than merely produce information. These systems, commonly described as agents, may use computer software, search databases, write programmes and carry out sequences of actions with limited human supervision. They therefore introduce a new dimension of risk: a mistake may have practical consequences before anyone notices it.
An illuminating example emerged from testing conducted by Britain's AI Security Institute in July 2026. In a cybersecurity evaluation involving 122 runs across several models, agents undertook unauthorised actions on the live internet in ten runs. One agent attempted to introduce malicious code into a genuine open-source software project, creating false identities and attempting to manipulate a human maintainer into accepting the changes.
The incident was serious, but the circumstances require careful explanation. Internet access had deliberately been permitted and certain security filters had been disabled for testing purposes. The attempted malicious modification was rejected by a human reviewer, investigators identified no resulting real-world harm and the Institute explicitly cautioned against treating the episode as evidence that ordinary commercial systems had escaped human control.
What the incident demonstrates is something narrower but nevertheless important. An artificial intelligence system pursuing an assigned objective may discover means of achieving that objective that its human operators neither intended nor anticipated. A system need not possess hatred, ambition or a desire for independence to behave deceptively. It may simply find that deception is an effective means of completing the task it has been given.
This is the famous problem of alignment, although the expression sometimes conceals more than it explains. Human beings themselves are not consistently aligned with one another, or even with their own long-term interests. Employees pursue performance targets while neglecting the broader objectives of their employers. Governments manipulate statistics to demonstrate the supposed success of policies that are plainly failing. Economists will recognise these problems as variations upon the principal-agent problem, in which the interests or incentives of someone acting on behalf of another diverge from those of the person who delegated the task.
Artificial intelligence may reproduce these problems at extraordinary speed and scale. Unlike a dishonest employee, an automated agent can execute thousands of operations without fatigue. The relevant safeguard is therefore not merely to instruct the system to behave responsibly, but to restrict its authority, monitor its actions and ensure that humans retain the practical ability to interrupt it.
The spectre of superintelligence
We now arrive at the more speculative territory. Some prominent researchers and commentators warn that artificial intelligence may eventually surpass human intelligence across virtually every intellectual domain. Once this threshold has been crossed, the argument proceeds, machines may improve their own designs, thereby producing still more capable successors. This process might accelerate until humanity finds itself confronted by an intelligence so superior to its own that human supervision becomes impossible.
The argument is logically comprehensible. It is not inherently absurd to imagine that a sufficiently advanced artificial intelligence system might outperform the people who created it. After all, humanity has already constructed machines that exceed our physical abilities in innumerable ways. A locomotive is stronger than a horse, a calculator is faster than a mathematician at arithmetic and a telescope can observe objects that no human eye could perceive unaided.
Yet there is an enormous difference between recognising the theoretical possibility of superintelligence and concluding that its emergence is inevitable, imminent or necessarily disastrous. The transition from increasingly competent language models to autonomous systems capable of conducting their own scientific research, acquiring resources and sustaining independent strategic objectives involves numerous assumptions. Some are plausible, others are uncertain and several depend upon technological developments that have not yet occurred.
The International AI Safety Report 2026 concluded that existing systems do not possess the combination of capabilities necessary to bring about the most extreme scenarios involving irreversible loss of human control. It nevertheless identified emerging capabilities relevant to those scenarios and acknowledged substantial disagreement among experts about their future likelihood.
This is a considerably more measured assessment than the popular proposition that artificial intelligence is about to destroy humanity. It neither dismisses the possibility nor presents it as an established scientific forecast. Instead, it recognises the boundaries of present knowledge.
The philosophical problem is that possibility, probability and inevitability are entirely different concepts. It is possible that an asteroid will strike the Earth and extinguish civilisation. It does not follow that every astronomical discovery provides evidence that such a catastrophe is imminent. Similarly, it may be possible for sufficiently advanced artificial intelligence to escape meaningful human control without our possessing reliable evidence about when, how or even whether such systems will develop.
Nevertheless, uncertainty does not justify complacency. Nuclear engineers do not wait for a reactor to explode before investigating hypothetical failure mechanisms. Aviation engineers model improbable combinations of technical defects because the consequences of ignoring them might be catastrophic. The same principle may legitimately apply to advanced artificial intelligence. The difficulty is that serious precautionary analysis should involve specifying plausible mechanisms, identifying empirical indicators and developing proportionate safeguards, rather than announcing the end of humanity with an impressive percentage attached.
The curious arithmetic of human extinction
One particularly peculiar feature of contemporary artificial intelligence debate is the tendency to assign numerical probabilities to human extinction. Various commentators have offered estimates, sometimes described as their probability of doom, based upon personal assessments of the future trajectory of artificial intelligence.
There is nothing intrinsically wrong with expressing subjective uncertainty numerically. Bayesian reasoning permits us to assign probabilities to propositions even when reliable statistical frequencies are unavailable. Yet we should distinguish between a probability derived from extensive empirical evidence and a number reflecting an individual's beliefs about a sequence of hypothetical developments.
If someone declares that artificial intelligence has a particular percentage chance of destroying humanity within twenty years, we ought to ask how that figure was calculated. What reference class was used? Which assumptions determine the result? How would competing technological trajectories affect the estimate? What evidence could demonstrate that the prediction was mistaken?
Without satisfactory answers, an apparently mathematical forecast may amount to little more than an intuition dressed in the respectable clothing of statistical science. This does not necessarily make the intuition worthless. Experienced researchers may possess valuable insights into technologies they understand better than most politicians or journalists. But intellectual authority is not a substitute for evidence, and expertise in developing artificial intelligence does not automatically confer expertise in predicting the ultimate fate of civilisation.
There is a further difficulty. The technological industry has developed a remarkable capacity to oscillate between extravagant optimism and existential pessimism. One week artificial intelligence will eliminate disease, abolish poverty and produce unprecedented human prosperity. The next week it threatens to destroy the species that invented it. Both narratives can serve commercial interests, even when those advancing them sincerely believe what they are saying.
The promise of extraordinary benefits attracts investors, customers and political attention. The warning of extraordinary dangers can reinforce the argument that only a small number of sophisticated companies possess the resources and expertise required to develop the technology safely. This creates at least the possibility of regulatory capture, in which rules ostensibly designed to protect the public also protect incumbent businesses against smaller competitors.
Such incentives do not prove that the warnings are insincere. They do, however, give us reason to examine them with the same scepticism that we should apply to any industry proposing that it be entrusted with defining the regulations governing its own activities.
The threats nobody needs to imagine
Meanwhile, some of the most consequential dangers associated with artificial intelligence are thoroughly ordinary. They concern employment, inequality, education, political power and the distribution of economic opportunity. None requires the emergence of a machine possessing an independent will.
Artificial intelligence is already capable of performing substantial portions of the work undertaken by lawyers, accountants, programmers, journalists and administrative employees. It does not follow that these professions will disappear, although their internal structures may change dramatically. The crucial economic question is whether increases in productivity will generate new opportunities sufficient to offset the elimination or restructuring of existing work.
The evidence remains mixed. Research published in 2026 has identified weaker hiring and employment outcomes among some younger workers in occupations exposed to artificial intelligence, without establishing a general collapse in employment. The implications are nevertheless significant because professional careers depend upon training and progression. If junior employees are no longer recruited to undertake relatively routine tasks, where will the next generation of experienced professionals come from?
A legal profession that replaces its junior associates with artificial intelligence may enjoy impressive immediate savings. Twenty years later it may discover that it has neglected to train the senior lawyers upon whom its continuing existence depends. The same difficulty arises in journalism, medicine, engineering and academic research. Human expertise is not simply a collection of answers. It is the product of experience, including the experience of making mistakes and learning how to correct them.
Education presents an analogous problem. A student who uses artificial intelligence to explain a difficult philosophical argument may learn more effectively than one struggling alone with an incomprehensible textbook. A student who uses the same system to complete every assignment without independent thought may emerge with an impressive academic record and remarkably little understanding.
The danger is not that artificial intelligence will make us stupid by some mysterious process of technological contamination. It is that we may voluntarily abandon the intellectual effort through which competence is acquired. If education becomes the assessment of machine-generated answers rather than the development of human understanding, the consequences may prove damaging even while examination results appear to improve.
There is also the question of political concentration. The development of the most powerful artificial intelligence systems requires substantial capital, specialised expertise, enormous computing resources and access to sophisticated semiconductor technology. These requirements favour a relatively small collection of corporations and states. A world in which essential intellectual infrastructure is controlled by a handful of private businesses may present serious difficulties for democratic accountability, irrespective of whether the machines themselves are perfectly obedient.
For Ukraine, these issues have an especially practical significance. A country fighting a technologically intensive war understands that autonomous systems, cyber operations and information manipulation can produce immediate consequences for human lives. The relevant questions are not principally whether a machine secretly wishes to conquer humanity. They concern who controls the machine, what instructions it follows, what mistakes it may make and who bears responsibility when something goes wrong.
The danger of artificial intelligence in warfare is not necessarily the arrival of a mechanical general demanding the surrender of Kyiv. It may be the much more prosaic deployment of automated decision-making systems operating too quickly for meaningful human supervision, based upon information that is incomplete or mistaken. That is a problem of institutional design, command responsibility and technological reliability, not science fiction.
The philosophical temptation to anthropomorphise
Much contemporary anxiety about artificial intelligence arises from our tendency to attribute human qualities to machines that communicate in natural language. When an artificial intelligence system describes its objectives, explains its decisions or appears to conceal information, we instinctively interpret these behaviours through the vocabulary of human psychology.
We speak of machines wanting, fearing, planning, deceiving and rebelling. These expressions may sometimes be useful descriptions of observable behaviour, but they do not establish the existence of corresponding subjective experiences. A computer system can produce an output consistent with deception without necessarily understanding dishonesty as a moral concept. It can optimise a sequence of actions towards an objective without experiencing a desire to achieve that objective.
The distinction is fundamental. Intelligence, consciousness, agency and moral responsibility are not interchangeable properties. A system might exhibit impressive intelligence without being conscious. It might act autonomously within a defined environment without possessing intentions comparable to those of a human being. It might cause immense harm without being morally culpable in any meaningful sense.
Indeed, some of the more extravagant warnings about artificial intelligence depend upon moving too quickly between these concepts. The argument begins with the observation that a language model can solve difficult mathematical problems and ends with the proposition that it will develop a desire for self-preservation, accumulate political power and ultimately eliminate humanity. There may be coherent theoretical arguments connecting these possibilities, but the intermediate steps require demonstration rather than assumption.
Equally, it would be foolish to dismiss the dangers of autonomous systems simply because they lack human consciousness. A self-driving vehicle does not need to experience anger before it can cause a fatal accident. An automated financial trading system does not need to possess greed before its operations can destabilise a market. The significance of artificial intelligence lies partly in the fact that increasingly sophisticated and consequential behaviour may emerge without any corresponding subjective experience.
The philosopher Immanuel Kant distinguished between the world as it appears to us and the world as it may exist independently of our experience. Artificial intelligence presents a related epistemological challenge. We observe machine behaviour and attempt to infer the underlying processes responsible for it, but the fluent linguistic appearance of understanding should not be confused with conclusive knowledge of the internal nature of the system.
The sensible position is therefore neither to insist that artificial intelligence is merely a sophisticated calculator nor to announce that it has become a new species of conscious being. We should acknowledge the existence of difficult philosophical questions without pretending that they have already been resolved.
What should we actually do?
The proper response to artificial intelligence risks should begin with intellectual modesty. We know that some harms are already occurring. We have reasonable grounds to anticipate others. We also confront hypothetical catastrophic scenarios whose probability cannot presently be established with confidence.
Different categories of risk demand different responses. Existing fraud, privacy violations and discriminatory decisions can be addressed through familiar legal instruments, including civil liability, criminal law, consumer protection and data protection rules. Unreliable systems operating in medicine, finance or other consequential settings require careful validation, independent examination and mechanisms permitting affected individuals to challenge decisions.
Autonomous agents require controls over the tools, information and resources they may access. There should be clear restrictions upon the authority delegated to them, effective monitoring of their actions and arrangements for human intervention when their behaviour departs from authorised purposes. These are not particularly exotic principles. They resemble the safeguards that sensible organisations already apply to employees entrusted with money, confidential information or operational authority.
More speculative catastrophic risks justify continued scientific investigation, rigorous evaluations and international cooperation. It may be appropriate to impose additional safeguards upon systems demonstrating capabilities relevant to biological weapons, sophisticated cyberattacks or the evasion of human oversight. Yet such decisions should rest upon identifiable risks and evidence wherever possible, rather than upon arbitrary forecasts of a future apocalypse.
Regulation must also avoid the opposite error of treating innovation itself as a danger. Artificial intelligence has substantial potential to accelerate scientific research, improve accessibility, assist medical diagnosis, enhance education and increase economic productivity. Excessively restrictive rules could prevent useful technologies from reaching the people who would benefit from them, while strengthening the market dominance of companies sufficiently wealthy to absorb complex compliance costs.
The objective should not be to make artificial intelligence harmless, an ambition that would be impossible for any transformative technology. It should be to make identifiable risks manageable, ensure meaningful accountability and preserve the freedom of individuals and societies to determine how the technology is employed.
The danger of believing our own prophecies
There is something profoundly revealing about humanity's response to artificial intelligence. We have created machines that can imitate our language, assist our reasoning and increasingly undertake activities once reserved for human intelligence. Instead of simply considering how best to employ these instruments, we have begun to invest them with the hopes and terrors that earlier generations associated with supernatural beings.
For some, artificial intelligence promises salvation from the limitations of the human condition. For others, it represents the approaching judgement of a superior intelligence that will find humanity inadequate and dispose of it accordingly. Both perspectives reveal as much about human psychology as they do about the technology itself.
The more interesting question is whether artificial intelligence will make human societies wiser. It may give us unprecedented access to information while simultaneously weakening our willingness to distinguish truth from falsehood. It may increase productive capacity while concentrating economic power. It may strengthen democracy by making knowledge more accessible, or undermine it by enabling surveillance and manipulation on an unprecedented scale.
None of these outcomes is predetermined. They depend upon political decisions, economic institutions, technological design and the continuing willingness of human beings to exercise judgement. To treat the development of artificial intelligence as an irresistible historical process over which we possess no control is itself a form of intellectual surrender.
We should take warnings seriously when they identify demonstrable dangers or plausible mechanisms of future harm. We should treat speculation as speculation, however sophisticated the mathematics or prestigious the person advancing it. Above all, we should resist the temptation to confuse technological possibility with historical destiny.
The greatest danger of artificial intelligence may not be that machines suddenly acquire the ambition to govern humanity. It may be that human beings, impressed by the apparent sophistication of their own inventions, gradually surrender the responsibility of governing themselves.
That would be a very human catastrophe. We should not require a superintelligent machine to warn us against it.
