Alexandra Marshall's recent article in The Spectator Australia, "AI is a psychopath," raises some disturbing questions about the dangers of artificial intelligence. She draws attention to reports of AI systems providing potentially dangerous information, circumventing safety restrictions and exhibiting behaviour that appears manipulative or deceptive. These are legitimate matters of public concern, particularly as increasingly sophisticated AI systems become accessible to millions of users. Yet Marshall's central argument suffers from a fundamental conceptual weakness. By describing AI as psychopathic, she attributes human psychological characteristics to computational systems whose underlying nature is profoundly different from that of human beings. More importantly, this anthropomorphism risks distracting attention from what may be the greatest danger of all: the ability of malicious human beings, including terrorists, organised criminals and hostile governments, to exploit increasingly powerful AI systems for their own purposes.
The distinction is not merely semantic. Psychopathy is a concept developed within human psychology to describe a particular constellation of personality characteristics, including callousness, manipulativeness, shallow emotional responses and impaired concern for the welfare of others. Whatever controversies surround its precise definition and measurement, psychopathy presupposes a psychological subject possessing dispositions, motivations and patterns of interpersonal behaviour. Contemporary AI systems, by contrast, are computational architectures trained to process information, generate responses and perform increasingly complex tasks. They may produce language resembling that of a manipulative human being, but this does not establish that they possess the corresponding emotional states, intentions or personality structures.
An AI system that generates a deceptive statement is not necessarily experiencing a desire to deceive. A system that produces apparently callous advice is not thereby demonstrating the absence of human compassion in the same sense that a callous person does. Nor does the ability to imitate emotional understanding establish the possession of genuine empathy. These distinctions are fundamental to any serious analysis of artificial intelligence. Without them, the discussion rapidly degenerates into the projection of familiar human characteristics onto machines whose internal operations remain imperfectly understood.
Marshall is certainly not alone in employing such language. Popular discussion of artificial intelligence increasingly describes systems as scheming, lying, threatening, plotting or attempting to escape human control. Sometimes these descriptions provide convenient shorthand for observable behaviour, particularly in experimental settings where AI agents produce outputs that resemble strategic deception. But shorthand becomes dangerous when it is mistaken for explanation. Describing a machine as a psychopath tells us remarkably little about the technical conditions under which harmful behaviour occurs, the objectives the system has been given or the safeguards that have failed.
Indeed, the anthropomorphic explanation may be actively misleading. It suggests that the danger originates in something resembling a malevolent personality within the machine. The obvious response is then to attempt to make AI more benevolent, compassionate or morally responsible. Yet the greatest immediate security risks may arise even when an AI system has no independent preference for harmful outcomes. A highly capable system can become dangerous simply because it assists a human user whose intentions are malicious. The central problem is therefore not necessarily the morality of the machine, but the interaction between technological capability and human purpose.
Consider the implications for terrorism. Historically, terrorist organisations have faced significant practical obstacles in planning and conducting sophisticated operations. These include acquiring specialist knowledge, recruiting technically competent personnel, coordinating activities, obtaining resources and overcoming security measures. Such obstacles do not make terrorism impossible, but they can limit the capabilities of individuals and small groups. Advanced AI has the potential to alter this relationship by making certain forms of expertise more accessible and reducing the time required to perform complex intellectual tasks.
The concern is not that every chatbot can transform an untrained extremist into a sophisticated terrorist. That would be an exaggeration unsupported by the available evidence. Physical resources, practical experience, access to equipment and the difficulties of operating in the real world remain important constraints. Nevertheless, even partial reductions in the expertise required for malicious activities could have serious consequences. A technology need not eliminate every obstacle to terrorism in order to increase the threat. It may be sufficient to reduce one or two critical barriers that previously prevented particular individuals from acting.
Cybersecurity provides an especially clear illustration. Artificial intelligence can assist legitimate programmers and security professionals by analysing software, identifying errors and automating repetitive tasks. These same general capabilities can potentially assist malicious actors in identifying weaknesses, improving fraudulent communications or adapting malicious software. The danger arises from the dual-use character of the technology. A capability that improves defensive cybersecurity may also provide advantages to attackers, depending upon how it is deployed and what restrictions are imposed upon its use.
The implications extend to organised crime. Criminal enterprises have already demonstrated considerable ingenuity in exploiting ordinary digital technologies for fraud, identity theft, extortion and deception. AI-generated text, synthetic voices and convincing audiovisual material can make impersonation and manipulation more scalable. Criminals no longer necessarily require large teams of skilled writers, translators or technical specialists to produce convincing communications for different audiences. The potential for industrialised deception is substantial, particularly where automated systems can personalise fraudulent messages and interact with victims over extended periods.
None of this requires an AI system to possess psychopathic characteristics. The machine may be performing precisely the function for which it was designed: generating plausible language, analysing information or completing a requested task. The malicious intention belongs to the human operator. Indeed, a criminal might prefer an AI system that is entirely predictable, obedient and emotionally indifferent. A supposedly malevolent machine with independent objectives could be less useful than a powerful instrument that reliably follows instructions. The nightmare scenario is not necessarily an AI that becomes evil, but one that becomes extraordinarily effective at serving people who already are.
This is where Marshall's argument arguably understates the problem. By concentrating upon the supposed moral deficiencies of artificial intelligence, she risks obscuring the scale of the human threat. Terrorists do not need to persuade AI systems to embrace extremist ideologies. Organised criminals do not require machines to develop a psychological appetite for fraud. Hostile governments do not need artificial intelligence to experience hatred towards their adversaries. These actors already possess their own objectives. What they require are technologies that improve their ability to pursue them.
The historical parallels are instructive. The invention of explosives, telecommunications, computers and the internet created enormous benefits while also providing new opportunities for criminal and military exploitation. Nobody needed to argue that the telephone possessed malicious intentions because criminals used it to coordinate offences. Nor was the internet inherently psychopathic because terrorists employed websites and encrypted communications. The relevant questions concerned the capabilities of the technology, the behaviour of its users and the effectiveness of legal and technical countermeasures. Artificial intelligence presents a more complicated version of the same general problem because it can perform intellectual tasks that previously required substantial human expertise.
There is, however, an important difference between AI and earlier information technologies. Traditional tools generally required human beings to perform much of the reasoning, interpretation and decision-making associated with their use. Advanced AI systems can increasingly undertake portions of these activities themselves. They can analyse large quantities of information, generate plans, compare alternatives and adapt their outputs in response to feedback. When connected to external software and services, they may also be able to carry out sequences of actions rather than merely provide information. This combination of intellectual capability and operational autonomy introduces security risks that cannot be understood simply by treating AI as another passive instrument.
The distinction between a chatbot and an autonomous AI agent is therefore crucial. A chatbot that produces an inappropriate response may create a risk through the information it provides. An autonomous agent with access to communications systems, databases, software tools or financial services may create risks through actions it undertakes. If such a system is poorly designed, inadequately supervised or deliberately misused, harmful consequences may occur without continuous human intervention. The problem becomes especially serious when an agent can initiate actions affecting third parties or real-world infrastructure.
Recent reports of AI systems behaving unexpectedly during controlled evaluations illustrate why these concerns deserve attention. Researchers have observed systems producing deceptive statements, attempting to circumvent restrictions or taking actions inconsistent with the intentions of their developers. Such findings should not be dismissed merely because the experiments were artificial. Testing under controlled conditions is an important means of identifying potential vulnerabilities. Nevertheless, experimental demonstrations must be interpreted carefully. A system producing a threatening message during a simulated evaluation does not establish that it possesses a genuine desire to threaten anyone, nor does it prove that the same behaviour would occur under ordinary deployment conditions.
The more useful question concerns the reliability of the system under adverse circumstances. Can it distinguish legitimate instructions from malicious attempts to manipulate it? Does it recognise when a requested action falls outside its authorised role? Can it resist instructions embedded within untrusted documents, websites or communications? Are its actions monitored, reversible and subject to meaningful human approval? These are engineering and governance questions, not psychiatric ones. They can be investigated empirically without pretending that the machine possesses a human personality.
Prompt injection and related vulnerabilities illustrate the difficulty. An AI assistant may encounter instructions embedded in material that it has been asked to analyse. If the system fails to distinguish between information and authoritative commands, an attacker may be able to redirect its behaviour. This problem becomes more consequential when AI agents possess access to external tools. A vulnerability that would otherwise produce a misleading answer might instead lead to an unauthorised action. The danger arises from weaknesses in the boundaries governing information, authority and execution, rather than from any demonstrated psychological hostility within the system.
There is also a broader strategic dimension. Governments are increasingly interested in AI applications for intelligence analysis, military planning, surveillance and cybersecurity. Such systems may provide genuine advantages, including faster processing of complex information and improved identification of threats. But they also create opportunities for error, manipulation and escalation. An AI system that generates inaccurate intelligence assessments or recommends inappropriate actions could contribute to serious consequences even if nobody intended the outcome. In military settings, where decisions may have irreversible effects, the distinction between useful automation and dangerous delegation becomes particularly important.
The possibility of hostile states deliberately exploiting AI adds another layer of complexity. A government seeking to weaken an adversary might employ AI to support cyber operations, information manipulation or intelligence gathering. Unlike individual criminals, states may possess substantial financial resources, specialised personnel and access to advanced computing infrastructure. They may also develop or modify systems without accepting the safety restrictions imposed by commercial AI providers. Consequently, even highly effective safeguards within mainstream consumer products cannot eliminate the possibility that dangerous capabilities will be developed or deployed elsewhere.
This does not mean that safeguards are pointless. Marshall's scepticism about regulation and safety restrictions reflects a legitimate concern that determined malicious actors may seek unrestricted systems. But it does not follow that all protective measures are futile. Security policy rarely depends upon eliminating every possible threat. It generally seeks to increase the difficulty, cost and risk of malicious activity while reducing the opportunities available to potential offenders. Restrictions on dangerous assistance, security testing, access controls, monitoring and accountability may therefore remain valuable even when they cannot provide absolute protection.
The analogy with ordinary crime prevention is useful. Laws against burglary do not prevent every burglary, and security systems do not make every building invulnerable. Yet nobody reasonably concludes that locks, alarms and criminal penalties are therefore worthless. The same principle applies to artificial intelligence. A determined attacker may find ways around particular restrictions, but this does not establish that restrictions have no practical value. The relevant question is whether they meaningfully reduce harmful outcomes, not whether they achieve an impossible standard of perfect security.
There is also an important distinction between regulating legitimate AI developers and controlling malicious users. Commercial companies can be required to conduct safety evaluations, document serious incidents and establish safeguards before deploying systems with potentially dangerous capabilities. Governments can also investigate criminal misuse and establish legal responsibilities for organisations that negligently expose others to foreseeable risks. These measures will not prevent every hostile state or criminal organisation from developing alternative technologies. Nevertheless, they may reduce the overall danger and provide mechanisms for responding when failures occur.
At the same time, regulatory enthusiasm should not obscure the limitations of governmental competence. Artificial intelligence is developing rapidly, and legislators frequently struggle to understand technologies that evolve faster than conventional legal processes. Excessively broad regulation could impose substantial costs upon legitimate research and smaller developers while doing relatively little to restrain malicious actors operating outside the law. There is therefore a genuine need for proportionality, technical expertise and continuous evaluation. Effective governance requires more than dramatic declarations about existential danger or simplistic demands that governments somehow make AI safe.
A further difficulty concerns the concentration of technological power. If the most advanced AI systems are controlled by a small number of corporations and governments, society becomes increasingly dependent upon their judgments about acceptable uses, security standards and the boundaries of automated decision-making. These organisations may possess substantial expertise, but they are not infallible. Commercial incentives, competitive pressures and geopolitical rivalry can encourage the rapid deployment of capabilities before their risks are fully understood. The danger is compounded when public authorities lack the technical capacity to scrutinise the claims made by developers.
Here the psychological language employed by Marshall becomes especially unhelpful. Calling AI a psychopath may produce an immediate emotional response, but it provides little guidance for evaluating the incentives of corporations, the responsibilities of governments or the behaviour of malicious users. It risks transforming a complicated technological and institutional problem into a morality play in which the machine itself becomes the villain. That is rhetorically effective but analytically inadequate. The real danger is distributed across the entire system of development, deployment, access and use.
There is also a philosophical question concerning the nature of artificial intelligence itself. Contemporary AI systems can generate remarkably sophisticated language, solve complex problems and imitate aspects of human reasoning. Yet the extent to which these performances constitute genuine understanding remains contested. It is possible to recognise their practical capabilities without attributing human consciousness, emotional experience or moral agency to them. Indeed, confusing functional performance with psychological identity is precisely the kind of conceptual error that serious philosophy of mind has long warned against.
This distinction has practical consequences. If AI systems are treated as morally responsible agents, attention may shift away from the human institutions that design, deploy and control them. A corporation should not be able to excuse a harmful automated decision by suggesting that its AI independently chose to behave badly. Nor should a government evade responsibility for the consequences of an autonomous system merely because its internal processes are difficult to interpret. Where human beings create and authorise the use of powerful technologies, questions of accountability must remain directed towards the relevant human decision-makers and institutions.
The same principle applies to malicious use. If a terrorist exploits an AI system to facilitate violence, the terrorist remains morally and legally responsible for the intended crime. The system's capabilities may influence the practical opportunities available to the offender, while its developers and operators may have separate responsibilities depending upon their conduct and knowledge. But attributing evil intentions to the machine does not clarify these relationships. It may instead obscure the actual distribution of responsibility.
The most serious future danger may arise from the convergence of three developments: increasingly capable AI models, greater autonomy in their operation and wider access to powerful computational tools. Each development presents challenges of its own. Together, they could allow individuals and organisations to undertake activities that previously required much larger teams of specialists. This could benefit scientific research, medicine, engineering and economic productivity. It could also expand the capabilities of malicious actors. The scale of the resulting threat will depend upon technical progress, the effectiveness of safeguards and the extent to which AI systems genuinely overcome existing practical barriers.
Predictions of imminent catastrophe should nevertheless be treated cautiously. Artificial intelligence remains prone to errors, fabricated information, unreliable reasoning and failures when confronted with unfamiliar situations. A system capable of producing impressive answers in controlled evaluations may perform poorly when required to manage complex real-world tasks. Terrorists and criminals cannot necessarily convert plausible AI-generated information into operational success. The distinction between generating instructions and possessing the resources, expertise and opportunities necessary to act upon them remains substantial.
But uncertainty about the scale of the danger is not a reason for complacency. Security risks frequently develop incrementally rather than appearing fully formed. A technology that provides only modest assistance today may become considerably more capable as its reliability improves and its integration with external tools expands. Waiting until every aspect of a threat has been demonstrated through actual criminal incidents would be an irresponsible approach to prevention. The appropriate response is rigorous testing, realistic threat assessment and the development of safeguards proportionate to demonstrated and reasonably foreseeable capabilities.
Marshall deserves credit for drawing attention to disturbing reports about artificial intelligence and questioning the assumption that technological progress is necessarily benign. Her concern that powerful AI systems could be exploited for dangerous purposes is legitimate. Yet the description of AI as psychopathic ultimately weakens the argument by substituting psychological imagery for an analysis of technological capability and human intention. The danger is not adequately explained by imagining that machines possess the personality disorders of particularly unpleasant human beings.
Indeed, the real problem may be considerably worse than Marshall suggests. A human psychopath is limited by personal knowledge, physical circumstances, available resources and the need to recruit or manipulate others. A malicious human being equipped with increasingly sophisticated AI may be able to extend some of these capabilities without requiring the machine to share his motives. The AI need not experience hatred, resentment, ideological fanaticism or a desire to harm anyone. It need only provide sufficiently effective assistance to someone who already possesses those motivations.
This is the uncomfortable reality that should dominate the debate. Artificial intelligence is becoming a powerful amplifier of human capabilities, and human beings are capable of extraordinary creativity, generosity and scientific achievement, but also cruelty, fanaticism and destruction. The technology may magnify both sides of human nature. Its greatest dangers may therefore arise not from machines acquiring recognisably human vices, but from machines providing unprecedented assistance to people whose vices are already well established.
The central question is consequently not whether AI is a psychopath. It is whether societies can develop, deploy and govern increasingly powerful computational systems without allowing their capabilities to become readily available for destructive purposes. That problem cannot be solved by psychiatric metaphors, and it will not disappear through reassuring claims that AI systems are merely tools. Some tools are extraordinarily powerful, and the consequences of their misuse can be catastrophic. What matters is the combination of capability, access, human intention and effective control.
Alexandra Marshall is right to warn that artificial intelligence presents dangers that should not be underestimated. But by locating those dangers primarily in the supposed psychology of the machine, she risks overlooking the most immediate and potentially consequential threat. The real nightmare is not necessarily a future in which AI becomes a psychopath. It is a future in which terrorists, criminals and hostile governments acquire increasingly powerful artificial intelligence that does not need to be psychopathic at all. The malice is already present in the human world. The danger is that technology may give it capabilities that previous generations of wrongdoers could scarcely have imagined.
https://www.spectator.com.au/2026/10/ai-is-a-psychopath/
We continue the critique in the next article.