There is something deeply human about wanting to recreate ourselves.
For thousands of years, we have made statues, dolls and mechanical automata in human form. Our stories are full of artificial beings brought to life: Pygmalion fell in love with the statue he had created; Mary Shelley imagined Frankenstein assembling a human-like creature from dead matter. Today, our robots increasingly have heads, faces, hands and voices, while artificial intelligences are judged by whether they can talk, write, draw and reason like us. We are interested not merely in creating intelligence, but maybe more so in seeing ourselves reflected back at us?
The history of artificial intelligence contains this fascination from the beginning. Russell and Norvig distinguish between approaches to AI that aim to think or act humanly and those that aim to think or act rationally. The Turing test belongs firmly to the former tradition: success means producing behaviour so human-like that an interrogator cannot reliably tell a machine from a person. Early systems such as Newell and Simon’s General Problem Solver went further, deliberately comparing the program’s reasoning processes with those used by human subjects (Russell & Norvig, 2021). But why should resemblance to humans be a measure of technological success?
Some of humanity’s most transformative inventions became useful precisely because they did not imitate humans, but did the jobs that were boring or dangerous. A washing machine does not reproduce the movements of someone kneeling over a washboard. A calculator does not imitate mental arithmetic. A telephone did not communicate like a human; it enabled humans to communicate better.
More importantly, these machines did not merely increase productivity. They returned human time. Economists have even described household appliances as “engines of liberation”: technologies that reduced the burden of domestic labour, historically borne disproportionately by women, and contributed to the dramatic increase in women’s participation in paid employment during the twentieth century (Greenwood et al., 2005). More recent evidence from China similarly found that increased ownership of labour-saving household appliances reduced housework and increased women’s market work and labour-force participation (Tewari & Wang, 2021). This seems to me one of the most compelling purposes of technology: not to reproduce human labour more impressively, but to release humans from labour that consumes their finite lives.
That makes some of our current technological priorities seem rather peculiar. RoboCup has spent almost three decades pursuing the deliberately ambitious goal that, by 2050, a team of autonomous humanoid robots should defeat the human football world champions. In 2026, two full teams of humanoid robots played an 11-versus-11 match for the first time (RoboCup Federation, 2026). It is an extraordinary engineering achievement. But where is the robot that reliably cleans the bathroom? Where is the machine that washes the windows, changes the beds, scrubs the shower, cleans the toilet and dusts the skirting boards without requiring a human to spend Saturday morning doing it?
The same argument applies to work that is not merely tedious, but dangerous. Robots could enter burning or collapsed buildings, work in radioactive or chemically contaminated environments, inspect unstable mines, handle hazardous waste or perform other jobs where exposing a human being to the risk adds no value. The point is not to create a mechanical person who performs these jobs exactly as we do. It is to build machines that can absorb risk, repetition and drudgery so that humans do not have to.
The same reasoning applies even more strongly to artificial intelligence.
In my previous essay, I set out 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). If we genuinely want AI to help solve problems of this magnitude, we would be better served by developing it around the things human cognition does not do particularly well.
Human brains are poor at scale: no person can read, remember and simultaneously compare millions of records. We are poor at recognising extremely rare events, because an individual may never acquire enough examples to become expert at recognising them. We struggle with high-dimensional complexity, where hundreds or thousands of interacting variables exceed what we can hold in our minds at once. We have difficulty reasoning intuitively about very large numbers, very small probabilities and exponential change. We are poorly equipped for long-term strategic planning, particularly where effects unfold across decades or generations. And our judgement is affected by systematic cognitive and social biases. Tversky and Kahneman showed, for example, how human reasoning under uncertainty relies on shortcuts that can systematically mislead us, including anchoring on an initial value or assumption (Tversky & Kahneman, 1974). Confirmation bias means that once we have formed an explanation, we are prone to seek or interpret subsequent evidence in ways that support it (Nickerson, 1998). Similarity can also influence social judgement: the person who went to our university, shares our background or simply “feels like one of us” may receive an advantage unrelated to actual merit.
AI is not automatically free from these problems. A system trained on biased human data can reproduce human prejudice at enormous scale. The point is therefore not to declare AI objective. The opportunity is to deliberately design it to complement human cognition rather than reproduce it: searching information spaces too large for us, detecting rare signals, comparing complex alternatives, modelling long-term consequences and exposing patterns in our decisions that we cannot see ourselves.
These capabilities could be applied directly to some of humanity’s largest problems. AI could search enormous bodies of scientific evidence for overlooked connections, detect the early signal of an epidemic among millions of health records, optimise food or energy distribution across complex networks, identify subtle ecological changes across decades of environmental data, uncover fraud across billions of transactions, or compare thousands of possible climate-policy strategies before decisions are made.
My own experience in healthcare has made the argument much less theoretical to me.
I have worked with rare-disease diagnoses, made from millions of data points on a single patient. A disease we encounter may be so uncommon that an excellent physician encounters it only once in her entire career. If she fails to recognise it, that does not represent poor medicine, it is simply a fundamental limitation of individual human experience. AI, however, can ask across millions of previous patients: Who was most similar to this patient, and what was the eventual diagnosis? That lets AI contribute what it does best—scale and pattern matching—while the doctor does what humans do best: interpret the result, make the clinical judgement, and support the patient.
This idea is not entirely new. One of the important early expert systems described by Russell and Norvig was MYCIN, developed to diagnose blood infections. With around 450 rules, it could perform as well as some experts and considerably better than junior doctors within its specialised domain (Russell & Norvig, 2021). The interesting conclusion is not that computers are “better than doctors”. It is that humans and computers can be wrong in different ways, and complement and support each other’s strengths.
The doctor can spend more time interacting with the patients, while an AI can analyse very large numbers of healthcare records to identify possible misdiagnoses, or patterns associated with unusually poor clinical outcomes.
Antimicrobial resistance offers another example. Longitudinal monitoring of critically ill patients may require repeated sampling and analysis over time. At sufficient scale and frequency, this quickly exceeds conventional human laboratory capacity. Automation combined with AI could support monitoring far beyond what any finite number of laboratory workers could achieve, potentially detecting emerging resistance patterns sooner and helping clinicians adapt treatment more rapidly. We create a life-saving system capable of doing something that no realistic number of laboratory technicians could do. In my own venture www.relatixhealth.com, we’ve used AI to create automatic self-testing of psychiatric conditions and neurodiversity, which is easily and affordably scalable to the entire humanity. Training a corresponding number of healthcare workers would be very costly, and very difficult to do.
Russell and Norvig describe another tradition in computing that captures this idea particularly well. Doug Engelbart called it intelligence augmentation: using computers to extend human intellectual abilities rather than simply automate human activity. They describe intelligence augmentation and artificial intelligence as complementary approaches (Russell & Norvig, 2021). We have spent centuries building machines that overcome the limitations of the human body. They let us travel faster than our legs can carry us, lift more than our muscles can bear, see farther than our eyes can reach and communicate beyond the range of our voices. Artificial intelligence gives us the opportunity to do something similar for the human mind. There will probably always be something irresistible about building a machine that looks at us, speaks to us and seems somehow familiar. That fascination may be deeply human. But familiarity is not necessarily usefulness.
We already have humans. What I believe we need from artificial intelligence and robotics is what humans cannot be.
References
Citi (2024). Closing the racial wealth gap. Citi Global Perspectives & Solutions.
Epley, N., Waytz, A. & Cacioppo, J.T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. doi:10.1037/0033-295X.114.4.864.
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.
Greenwood, J., Seshadri, A. & Yorukoglu, M. (2005). Engines of liberation. The Review of Economic Studies, 72(1), 109–133. doi:10.1111/0034-6527.00326.
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. Contribution of Working Group II to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press.
Montoya, R.M. & Horton, R.S. (2013). A meta-analytic investigation of the processes underlying the similarity-attraction effect. Journal of Social and Personal Relationships, 30(1), 64–94. doi:10.1177/0265407512452989.
Nickerson, R.S. (1998). Confirmation bias: A ubiquitous phenomenon in many guises. Review of General Psychology, 2(2), 175–220. doi:10.1037/1089-2680.2.2.175.
RoboCup Federation (2026). While the World Watches Soccer, Robots Made History in Incheon. RoboCup Federation.
Russell, S.J. & Norvig, P. (2021). Artificial Intelligence: A Modern Approach, 4th ed. Pearson.
Tewari, I. & Wang, Y. (2021). Durable ownership, time allocation, and female labor force participation: Evidence from China’s “Home Appliances to the Countryside” rebate. Economic Development and Cultural Change, 70(1), 87–127. doi:10.1086/706824.
Tversky, A. & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. doi:10.1126/science.185.4157.1124.
UN IGME (2026). Levels & Trends in Child Mortality: Report 2025. United Nations Inter-agency Group for Child Mortality Estimation.
WHO (2025). Global report on neglected tropical diseases 2025. Geneva: World Health Organization.
WMO (2025). State of the Global Climate 2024. Geneva: World Meteorological Organization.
World Bank (2020). World could achieve gender dividend of $172 trillion from closing lifetime earnings gaps. World Bank.