Something strange has happened to parts of artificial-intelligence research.
Some of the people building the world’s most advanced AI systems seriously discuss the possibility that their creations could take over the world, enslave humanity, replace us as the dominant species, or eliminate us. In 2023, hundreds of AI researchers and technology leaders — including Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman and Dario Amodei — signed a statement declaring that mitigating the risk of human extinction from AI should be a global priority alongside pandemics and nuclear war (Center for AI Safety, 2023).
Microsoft AI CEO Mustafa Suleyman has suggested that AI might best be understood as a new “digital species” (Suleyman, 2024). Max Tegmark’s Life 3.0 describes possible futures in which superintelligent AI conquers us, keeps humans rather like animals in a zoo, rules as a benevolent dictator, or eventually replaces humanity with artificial descendants (Tegmark, 2017; Species | Documenting AGI, 2026). At times, parts of this discourse begin to sound less like engineering and more like a doomsday cult.
I do not mean that catastrophic risks should simply be ignored. Technologies can cause catastrophic harm, and thinking about unlikely but devastating outcomes is sensible. I think AI could contribute to an existential catastrophe. But I struggle with one enormous leap in the argument:
How exactly do we get from the current AI systems we are actually building to a single omnipotent artificial being capable of controlling the planet and us?
AI is not one thing
Today, AI exists in thousands of different places. It sits in phones, cars, hospitals, laboratories, email systems, data centres, factories and homes. Different systems have different owners, architectures, permissions, purposes and physical capabilities. One recognises faces. Another interprets histopathology. Another predicts protein structures. Another filters spam. Another controls a robot vacuum cleaner. In 2026, two full teams of humanoid robots played an 11-versus-11 match for the first time (RoboCup Federation, 2026). But none of this means these systems are becoming one entity.
I can easily imagine having 10 robots in my home 50 years from now. Perhaps one cleans the floors, another washes the windows, another changes the beds, another manages groceries, another maintains the garden and another plays with the cat. But why would they become one mind and try to enslave me?
They might communicate. They might coordinate. They might even share an underlying foundation model.
Coordination is not the same as integration into a single intelligence with a single objective.
Several assumptions in conventional superintelligence narratives therefore deserve much more scrutiny.
- 1AI is treated as a single actor. We routinely hear statements about what “AI” will want once “AI” becomes smarter than humanity. Yet there is no obvious reason that independently developed systems belonging to different governments, companies, households and individuals should acquire one objective. Humans are one species and cannot agree on much; we are billions of independent agents interacting. Why should artificial agents be more unified?
- 2Intelligence is confused with power. Being extraordinarily good at reasoning does not automatically provide control over electricity grids, factories, banks, weapons or robots. Power requires access, permissions, communications, money, energy, hardware and physical infrastructure. A neural network does not magically acquire these things because its reasoning improves, and it is unclear why it would be motivated to move in the direction of accumulation of strategic power.
- 3Coordination is assumed rather than explained. For a rogue AI to commandeer my ten household robots, it needs a route into them: connectivity, compatible interfaces, exploitable vulnerabilities and sufficient privileges. That sounds less like the spontaneous emergence of a new species and rather more like an extremely sophisticated cybersecurity attack.
- 4Capability is treated as if it will immediately become universal. Real AI capabilities are uneven. A system can outperform humans dramatically in one task and remain incompetent in another. Intelligence and capability is not a single ladder on which every ability improves simultaneously.
- 5Historical and biological analogies are pushed too far. European conquest of the Americas shows very clearly that technological asymmetry can contribute to catastrophic violence, disease, dispossession and demographic collapse. It does not demonstrate that the technologically weaker population must become extinct. More than five centuries later, Indigenous and European-descended populations coexist, cooperate, compete, intermarry and learn from one another. History is considerably messier than “the smarter species replaces the weaker species.”
Even the word species seems misleading. A species is a population, not a single mind. Members of the same species compete, cooperate, form alliances, exploit one another and occupy different ecological niches. Perhaps ecosystem is the better metaphor?
An artificial and intelligent ecosystem
Imagine the future of AI not as one artificial super-being, but as an enormous ecosystem of artificial agents.
Medical AIs may interact with laboratory systems. Agricultural robots may cooperate with weather models and irrigation systems. Financial AIs may compete with one another. Cybersecurity agents may hunt malicious agents. Household robots may work alongside humans, pets and other machines. Some will compete for shared resources. Electricity is already an obvious example. Data centres consumed around 415 TWh of electricity in 2024 — about 1.5% of global electricity consumption — and demand is expected to rise substantially as AI infrastructure expands (IEA, 2025). There are also demands for land, water, cooling infrastructure, chips and critical minerals. The environmental and social costs do not occur only in Silicon Valley. Mining for copper, lithium, cobalt and rare-earth elements has been associated with pollution, forced displacement, loss of livelihoods and violations of Indigenous peoples’ rights in some regions (Amnesty International, 2026). But other artificial systems may occupy almost completely separate niches.
An ecosystem can also contain dangerous actors. A rogue artificial agent might spread through vulnerable systems in something resembling malware. It could steal information, disrupt infrastructure, manipulate people or compromise physical systems. That could cause serious harm and, under the wrong circumstances, perhaps catastrophic harm.
We already know that comparatively ordinary software can create enormous cascading consequences. The 2017 WannaCry ransomware attack disrupted large parts of the English NHS, even though the NHS itself was not its intended target (National Audit Office, 2017).
Now imagine an AI agent able to search autonomously for vulnerabilities, adapt its attacks, write new exploits, copy itself and move across poorly secured systems. The UK National Cyber Security Centre has warned that AI is likely to increase the frequency and effectiveness of cyber intrusions and accelerate vulnerability discovery and exploitation (NCSC, 2025). That feels like a credible AI risk to me.
There are other possibilities. Increasingly capable AI could lower barriers to biological misuse, assist in designing dangerous pathogens, or operate autonomous systems at speeds that humans cannot meaningfully supervise. Widely shared models and infrastructure could also create technological monocultures, where one vulnerability or design error propagates across thousands of organisations. These scenarios could become extremely serious without requiring every Roomba, robot dog, hospital model and data centre on Earth to unite behind one objective.
And an artificial ecosystem would contain defenders as well as attackers. Banks do not want fraud. Governments do not want critical infrastructure compromised. Security companies do not want malware spreading. Manufacturers do not want somebody else controlling their machines. AI will therefore be used both to attack systems and to defend them.
That resembles an ecology of predators, parasites, competitors, mutualists and immune systems considerably more than a single digital emperor.
A convenient apocalypse?
There is another aspect of the doomsday narrative that makes me uncomfortable.
Many of the people warning most publicly that AI could cause human extinction are simultaneously involved in an industry racing to build larger, more capable and more widely deployed AI systems. It would be unfair to claim that frontier AI companies do nothing about safety. They employ safety teams, publish evaluations and fund alignment research. But there is nevertheless an extraordinary contradiction between the rhetoric of existential danger, and the continuing race to make these systems more capable as quickly as possible. The economic incentives are obvious. More capable AI potentially means more investment, market share, political influence and profit.
Meanwhile, many of the costs occur somewhere else: Workers employed in Kenya to label disturbing material used in the development of ChatGPT described repeated exposure to graphic descriptions of murder, sexual abuse and other traumatic content, with some reporting serious psychological consequences (Perrigo, 2023).
AI infrastructure consumes electricity, water, land and minerals. Its supply chains can affect communities far removed from the companies ultimately profiting from the technology. And labour markets are already being reorganised. The International Labour Organization estimates that around one quarter of global employment lies in occupations with some exposure to generative AI, although it expects transformation of jobs to be more common than complete replacement (Gmyrek et al., 2025).
These are less spectacular problems than discussing whether a future superintelligence will keep humans as pets. But they are real.
Whether deliberate or not, this can begin to function like a bait and switch in public debate: attention is directed towards an abstract future apocalypse while some of the measurable environmental, social and economic consequences of building AI at enormous scale receive considerably less attention. And there is an obvious question:
If the people developing frontier AI genuinely believe that there is a substantial probability their technology could cause human extinction, should that belief not have consequences for what they actually do?
Giving interviews about extinction while simultaneously competing to build more powerful systems as rapidly as possible seems an odd response to believing that those systems might kill everyone.
The risks are already here
We also do not have to imagine hypothetical superintelligence to find AI causing harm.
Facial recognition provides a straightforward example. A major NIST evaluation found significant demographic differences in error rates across many facial-recognition algorithms (Grother et al., 2019). In Detroit, Robert Williams was wrongfully arrested after police relied on an incorrect facial-recognition match, eventually leading to legal action and new safeguards governing police use of the technology (ACLU, 2024).
Healthcare provides another example. Obermeyer and colleagues studied an algorithm used to identify patients who would benefit from additional healthcare support and found substantial racial bias. The algorithm used previous healthcare expenditure as a proxy for medical need. Because less money had historically been spent on Black patients at comparable levels of illness, the apparently neutral algorithm reproduced an existing inequality (Obermeyer et al., 2019).
Robotics can turn software failures into physical ones. In 2023, a driverless Cruise vehicle struck and subsequently dragged a pedestrian approximately 20 feet before stopping. The US National Highway Traffic Safety Administration later took enforcement action concerning the company’s reporting of the crash (NHTSA, 2024).
There are psychological questions too. Research involving OpenAI and the MIT Media Lab has found associations between some patterns of intensive chatbot use and increased emotional dependence and problematic use. The research does not establish the simplistic conclusion that chatbots automatically “cause addiction”, but it demonstrates that human relationships with conversational AI deserve serious empirical study (OpenAI & MIT Media Lab, 2025).
These are not science-fiction scenarios. They are engineering, medical, psychological, social, economic and regulatory problems happening now.
Perhaps we are looking in the wrong direction?
I am not arguing that AI could never contribute to an existential catastrophe. I think it could. What I question is the extraordinary leap from:
AI systems are becoming increasingly capable → therefore a single superintelligent AI will become an omnipotent entity controlling the planet.
As explained, there are much more concrete pathways to severe or even catastrophic harm. AI could amplify cyberattacks against interconnected critical infrastructure. It could lower barriers to biological misuse. Autonomous systems could make consequential decisions at speeds humans cannot adequately supervise. Common software, models or infrastructure could create technological monocultures in which one vulnerability produces cascading failures.
These possibilities are frightening precisely because they do not require science fiction. They are extensions of failure modes we can already observe.
That is why crashes and near misses matter so much. They tell us where larger catastrophes might actually come from. Real AI safety work comes from; investigating accidents, investigating near misses. Introduction of redundancy, certification, checklists, engineering standards, incident reporting and regulation. Perhaps AI safety needs rather more of that? Not improbable doomsday scenarios?
What would I rather we do?
Personally, I would much rather see serious effort put into laws, regulatory frameworks, enforceable engineering standards and practical guidance for implementing AI responsibly: how systems should be validated, where humans need to remain involved, how incidents and near misses should be reported, how vulnerable people and communities should be protected, who is accountable when systems fail, and how autonomy and access to critical infrastructure should be controlled. And I would much rather see AI leveraged not primarily to produce increasingly cute ten-second video clips, but to help resolve some of humanity’s most pressing challenges:
In 2024, around 673 million people experienced hunger (FAO et al., 2025). Some 4.9 million children died before reaching their fifth birthday (UN IGME, 2026). Neglected tropical diseases affect more than one billion people (WHO, 2025). Around one million animal and plant species are threatened with extinction (IPBES, 2019). Climate change is already damaging human societies, livelihoods and ecosystems (IPCC, 2022; WMO, 2025). Gender inequality has been estimated to represent around US$172 trillion in lost lifetime earnings and human-capital wealth (World Bank, 2020), while racial inequalities in the United States alone have been estimated to represent trillions of dollars in unrealised economic potential (Citi, 2024).
Those are not hypothetical risks. This is what is already happening.
Perhaps instead of spending quite so much time asking whether an imaginary artificial species will destroy humanity, we could spend rather more time asking how artificial intelligence can help humanity avoid destroying itself? And how AI can create better futures for many of us?
References
ACLU (2024). Williams v. City of Detroit. American Civil Liberties Union.
Amnesty International (2026). What are critical minerals and why do they matter for human rights?. Amnesty International.
Center for AI Safety (2023). Statement on AI Extinction Risk.
Citi (2024). Closing the racial wealth gap. Citi Global Perspectives & Solutions.
FAO, IFAD, UNICEF, WFP & WHO (2025). The State of Food Security and Nutrition in the World 2025: Addressing high food price inflation for food security and nutrition. Rome: FAO. doi:10.4060/cd6008en.
Gmyrek, P., Berg, J., Kamiński, K., Konopczyński, F., Ładna, A., Nafradi, B., Rosłaniec, K. & Troszyński, M. (2025). Generative AI and jobs: A refined global index of occupational exposure. ILO Working Paper 140. International Labour Organization. doi:10.54394/HETP0387.
Grother, P., Ngan, M. & Hanaoka, K. (2019). Face Recognition Vendor Test Part 3: Demographic Effects. NISTIR 8280. National Institute of Standards and Technology. doi:10.6028/NIST.IR.8280.
International Energy Agency (IEA) (2025). Energy and AI. Paris: IEA.
IPBES (2019). Global assessment report on biodiversity and ecosystem services. Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services. doi:10.5281/zenodo.3553458.
IPCC (2022). Climate Change 2022: Impacts, Adaptation and Vulnerability. Cambridge: Cambridge University Press.
National Audit Office (2017). Investigation: WannaCry cyber attack and the NHS. London: National Audit Office.
National Cyber Security Centre (NCSC) (2025). The impact of AI on the cyber threat from now to 2027. UK National Cyber Security Centre.
National Highway Traffic Safety Administration (NHTSA) (2024). NHTSA announces consent order with Cruise after company failed to fully report crash involving pedestrian. U.S. Department of Transportation.
Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. doi:10.1126/science.aax2342.
OpenAI & MIT Media Lab (2025). Early methods for studying affective use and emotional well-being on ChatGPT.
Perrigo, B. (2023). OpenAI used Kenyan workers on less than $2 per hour to make ChatGPT less toxic. TIME.
RoboCup Federation (2026). While the World Watches Soccer, Robots Made History in Incheon. RoboCup Federation.
Species | Documenting AGI (2026). MIT Explains the 12 Possible Endings for AI. YouTube.
Suleyman, M. (2024). What is an AI anyway?. TED2024.
Tegmark, M. (2017). Life 3.0: Being Human in the Age of Artificial Intelligence. New York: Alfred A. Knopf.
UN Inter-agency Group for Child Mortality Estimation (UN IGME) (2026). Levels & Trends in Child Mortality: Report 2025. New York: UNICEF.
World Bank (2020). World could achieve gender dividend of $172 trillion from closing lifetime earnings gaps. World Bank.
World Health Organization (WHO) (2025). Global report on neglected tropical diseases 2025. Geneva: WHO.
World Meteorological Organization (WMO) (2025). State of the Global Climate 2024. Geneva: WMO.