In Artificial Intelligence: A Modern Approach, Russell and Norvig describe humanity’s attempt to create AI. Should a machine think like a human, act like a human, think rationally, or act rationally? Their preferred framework is the rational agent: an agent that chooses the action expected to produce the best outcome. But they are careful about the limitations. Perfect rationality is often impossible, and an agent can optimize the wrong objective (Russell & Norvig, 2021). A few milestones show how we got here:
- 1943: McCulloch and Pitts combined knowledge of brain neurons with mathematical logic to describe artificial neural networks (McCulloch & Pitts, 1943).
- 1950: Alan Turing sidestepped the philosophical question “Can machines think?” and instead asked whether a machine could behave convincingly enough to be mistaken for a human (Turing, 1950).
- 1956: the Dartmouth workshop helped establish artificial intelligence as a field (Russell & Norvig, 2021).
- 1960s–1980s: symbolic reasoning, search and expert systems such as DENDRAL and MYCIN showed that machines could perform sophisticated tasks in restricted domains. DENDRAL inferred molecular structures from mass-spectrometry data using expert chemical rules, while MYCIN used expert rules and uncertainty estimates to diagnose blood infections (Russell & Norvig, 2021).
- 1986: back-propagation helped revive multilayer neural networks (Rumelhart, Hinton & Williams, 1986).
- 1988 onward: probabilistic reasoning, Bayesian networks and machine learning increasingly allowed AI to work with uncertainty rather than relying only on rigid logical rules (Pearl, 1988).
- 2000s–2010s: big data, greater computing power and deep learning produced the current explosion in AI capability (Russell & Norvig, 2021).
AI's history is not simply a history of computers becoming faster. It is a history of changing ideas about what intelligence is, and which parts of it can be reconstructed mathematically. I would argue that in order to develop AI responsibly – we need to critically examine the history of AI development, and learn a few more insights from biology.
I came to AI through biology
My own route into this subject was biology and statistics, not computer science, and I’ve found AI very useful for tackling the multifactorial and complex problems found in biology. Working on reconstructing population history and gene flow in mosquitoes that spread malaria, I used Bayesian and machine-learning models (Zarowiecki et al., 2014). I have also used HMM-based protein-domain annotation and comparative genomics to understand protein function and parasite evolution, including the comparative analysis of four tapeworm genomes (Tsai, Zarowiecki et al., 2013) and a broader comparison across sequenced helminth genomes (Zarowiecki & Berriman, 2015). I have also implemented neural-network approaches to classify acute myeloid leukaemia patients according to their predicted cell of origin, with classifications associated with survival and treatment resistance (Zeisig et al., 2021). More recently, I worked on an NLP pipeline using BERT-based named-entity recognition, regular expressions and text mining to extract genomic variants and gene–mutation associations from the scientific literature for WormBase (Mallick et al., 2022). So I am deeply grateful these methods exist. I have spent much of my career putting them to very good use.
What I am sceptical of is a lot of the discourse around AI, including that presented by Russell and Norvig. AI is a great tool, but we humans need to understand ourselves and our intelligence better to see the massive biases in AI research.
Which problems do we choose to give AI?
This is what worries me most about the development of AI. It is visible throughout its history: some of the most celebrated AI successes have been – defeating humans at the games chess (Campbell, Hoane & Hsu, 2002) and Go (Silver et al., 2016), winning Jeopardy! (Ferrucci et al., 2010), and achieving spectacular performance in image recognition (Krizhevsky, Sutskever & Hinton, 2012). Why were these chosen to demonstrate the peak of human intelligence and capability?
It is tempting to dismiss these as benchmarks: convenient problems on which to test algorithms before applying them somewhere useful. But I am a scientist too. I have spent years benchmarking algorithms, including algorithms for cancer classification. Benchmarks are not neutral. What you choose to measure reflects what you have decided counts as success. If somebody had handed me Deep Blue and asked me to demonstrate to the world that it was an important intelligent system, I would not have chosen chess.
I notice that we now spend extraordinary amounts of scientific talent, money, electricity and computing power using AI to generate pictures, essays and videos, automate office tasks, target advertising and recognize objects in photographs. We can make an AI distinguish a blueberry muffin from a dog. That is technically impressive. But is it really one of humanity's most urgent problems – one worthy of the attention of the most powerful tools humans have invented?
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). The economic costs of inequality are similarly vast. The World Bank estimates that closing gaps in lifetime labour earnings between women and men could add US$172 trillion to global human-capital wealth (World Bank, 2020). Racism is harder to quantify globally; in the United States alone, Citi estimates that gaps between Black and white Americans in income, business ownership and homeownership could have cost the economy up to US$3.9 trillion between 2020 and 2024 (Citi, 2024).
Of course AI is being used in medicine, biology, agriculture, climate research and conservation. My own work is an example. My question is about focus, priority and scale. In 2025, AI firms attracted around US$258.7 billion in global venture-capital investment, with AI infrastructure and hosting alone attracting US$109.3 billion (OECD, 2026). More than 90% of notable AI models produced in 2025 came from industry (Stanford HAI, 2026). How much of the world's AI effort and compute is directed towards what is actually humanity’s and the planets largest and existential challenges, like eliminating neglected diseases, reducing child mortality, ending hunger, providing clean energy or saving biodiversity? I cannot find a meaningful global accounting. Perhaps we should have one?
The most powerful person in an optimization problem is not the optimizer. It is the person who chooses the objective and decides what is the correct solution. In AI currently, humanity is using the most powerful tool we’ve invented to make funny social media memes (I do it too ;-p ). Very human – not very smart!
Models need boundaries. Reality doesn't have them.
Even choosing a good objective is not enough. Statistical models have to simplify reality. We choose variables, make assumptions and draw boundaries around the problem. Real life has no such obligation. Optimize renewable-energy production and there are still consequences for mines, materials, manufacturing, land use and ecosystems. Optimize crop production and you affect soils, water, insects and biodiversity.
A particularly neat example comes from healthcare. A widely used algorithm studied by Obermeyer and colleagues was good at predicting what it had been asked to predict: healthcare costs. Unfortunately, costs had been used as a proxy for healthcare need. Because less money had historically been spent on equally sick Black patients, the algorithm systematically underestimated their needs. Changing the objective substantially reduced the bias (Obermeyer et al., 2019). The algorithm was not irrational; it was rational about the wrong thing.
I think of problem and solution space like an ecosystem; the real problem space is effectively infinite. The consequences do not stop where our model stops. If we think of a problem space not as something neatly bounded, but as something more like an ecosystem—where every problem and every solution exists within a larger context of other problems and solutions—we may become better at avoiding solutions that simply create larger problems elsewhere. We should not solve “transport” by creating “global warming,” or solve “cheap storage” by creating “plastic waste covering the planet.”
Humanity has to learn not to solve one problem by creating one or more new problems. Even current AI development is failing in this perspective; exhausting limited resources like clean water, rare metals and energy (IEA, 2025; Lei et al., 2025).
Humans are not failed optimization algorithms
There is another problem with thinking about intelligence mainly as rational optimization. Humans have evolved over millions of years, and part of that evolution involves fear, empathy, exhaustion, empathy, attachment, pain, instinct, experience, and an awareness that we will die. These things have helped keep many of us alive, and regularly prevent us from doing what would be “optimal” under a narrowly defined objective. I think of these side-steps from cold rationality as a main feature, not a bug.
Consider war. If the objective is to defeat an opponent while minimizing your own losses, an automated targeting system that kills enemy soldiers accurately and without hesitation may outperform a human soldier. Humans often do something else. During the Christmas Truce of 1914, unofficial ceasefires appeared along parts of the Western Front. British and German soldiers met in no-man’s-land, exchanged food and souvenirs, buried the dead and, in some places, played football, despite orders against fraternisation (Rix, 2014). This kind of restraint is not unique. Research documents soldiers and other armed personnel refraining from killing despite having the opportunity and legal justification to do so, particularly when they recognize the humanity of the person in front of them (Baggaley, Marques & Shon, 2019). Renic (2019) describes the related phenomenon of battlefield mercy: combatants sparing an enemy whom they would have been legally and morally permitted to kill. If military AI is built around cold optimization of objectives such as identifying and neutralizing targets, without the human capacity for empathy, hesitation or mercy, it may therefore behave very differently from the soldiers it replaces.
The same principle appears in less dramatic situations. A self-driving car may follow traffic rules perfectly and still behave stupidly in context. A human driver may mount a kerb to let an ambulance pass. The technically “incorrect” action may be the intelligent one because the driver understands what the rule is for, rather than merely following it (Zhang et al., 2026).
This makes me wonder whether we are trying to recreate only half of human intelligence. We copy prediction, language, calculation, pattern recognition, planning and optimization, but leave behind the biological framework in which those abilities evolved – and it is to our detriment. Fear discourages reckless behaviour. Pain makes damage difficult to ignore. Empathy makes another person's suffering relevant. Empathy makes lives matter. Exhaustion forces limits. Mortality makes irreversible consequences meaningful. I am not arguing that an AI literally needs to be frightened, tired, or afraid of death. But perhaps it needs functional equivalents of the constraints these traits provide? Biology has spent billions of years building these mechanisms into us, and they help keep us alive. If AI eventually becomes vastly more capable than we are, then copying only half of a human—our ability to calculate and optimize without the biological framework that constrains those abilities—may be particularly dangerous.
So perhaps trying to create an artificial human intelligence is the wrong goal altogether? Perhaps something closer to a benevolent God is a better aspiration: an intelligence far more capable than ourselves, but whose fundamental purpose is to protect and care for beings weaker than itself?
If we create something god-like in capability, perhaps we should work equally hard to make it God-like in benevolence, so that it can properly transcend the limitations of humanity?
The danger of making machines look human
This brings me back to the Turing Test: can a machine act so much like a human that a human thinks it is human? Russell and Norvig distinguish acting humanly from thinking humanly. I would add another distinction: Acting humanly is not the same as being human. Humans may be less good at maintaining that distinction; we anthropomorphize almost anything. We talk to animals and cars as if they are human. We become “friends” with chatbots. We see spirits and gods in trees and rivers.
Recent experimental work found that many people were willing to attribute at least some possibility of phenomenal consciousness to large language models (Colombatto & Fleming, 2024). Research on AI language warns that describing machines as thinking, understanding, wanting and feeling encourages anthropomorphism (Shardlow & Przybyła, 2024). This anthropomorphism can have consequences. A recent review describes harms associated with LLM-based chatbots ranging from emotional dependency and problematic use to inappropriate interactions in severe mental-health situations (Diel et al., 2026). The evidence is still developing, but these risks are no longer merely hypothetical.
Even our ordinary language gives the machine agency: AI decided. AI thinks. AI wants. AI is taking people's jobs. But AI did not decide to replace a worker. People decide to replace a worker using AI, and we should not blame-shift those choices.
We have to stop anthropomorphizing AI and start placing responsibility with the humans to whom it belongs.
The genie is out of the bottle
Powerful AI is widely available now. It is not going back into the bottle. So perhaps the most interesting question is no longer how intelligent AI will become. It is: what do we choose to do with it?
We can use extraordinary amounts of human ingenuity, electricity, money, and computation to automate payroll, optimize another advertisement, and generate another video. Or we can direct much more of it towards solving humanity's and the planet's largest, most complex and existential issues: preserving biodiversity, eliminating neglected tropical diseases, reducing hunger and child mortality, tackling racism and gender inequality, understanding complex biological systems and creating a sustainable future?
It remains our responsibility to give AI problems worth solving, and to define sustainable solutions. It is a tool – let’s not be.
References
Baggaley, K., Marques, O. & Shon, P.C. (2019). An exploratory study of the decision to refrain from killing in the accounts of military and police personnel. Journal of Military Ethics, 18(1), 20–34. doi:10.1080/15027570.2019.1625508.
Campbell, M., Hoane, A.J. Jr. & Hsu, F.-h. (2002). Deep Blue. Artificial Intelligence, 134(1–2), 57–83. doi:10.1016/S0004-3702(01)00129-1.
Citi (2024). Closing the racial wealth gap. Citi Global Perspectives & Solutions.
Colombatto, C. & Fleming, S.M. (2024). Folk psychological attributions of consciousness to large language models. Neuroscience of Consciousness, 2024(1), niae013. doi:10.1093/nc/niae013.
Diel, A., Torous, J., Cuijpers, P., et al. (2026). A scoping review on the mental health harms of LLM-based chatbots. npj Digital Medicine, 9, 644. doi:10.1038/s41746-026-03054-x.
FAO, IFAD, UNICEF, WFP & WHO (2025). The State of Food Security and Nutrition in the World 2025. Rome: Food and Agriculture Organization of the United Nations. doi:10.4060/cd6008en.
Ferrucci, D., Brown, E., Chu-Carroll, J., et al. (2010). Building Watson: An overview of the DeepQA project. AI Magazine, 31(3), 59–79. doi:10.1609/aimag.v31i3.2303.
IEA (2025). Energy and AI. Paris: International Energy Agency.
IPBES (2019). Summary for policymakers of the global assessment report on biodiversity and ecosystem services. Bonn: IPBES Secretariat. 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.
Krizhevsky, A., Sutskever, I. & Hinton, G.E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 1097–1105.
Lei, N., Lu, J., Shehabi, A. & Masanet, E.R. (2025). The water use of data center workloads: A review and assessment of key determinants. Resources, Conservation and Recycling, 219, 108310. doi:10.1016/j.resconrec.2025.108310.
Mallick, R., Arnaboldi, V., Davis, P., Diamantakis, S., Zarowiecki, M. & Howe, K. (2022). Accelerated variant curation from scientific literature using biomedical text mining. microPublication Biology, 2022. doi:10.17912/micropub.biology.000578.
McCulloch, W.S. & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 5, 115–133. doi:10.1007/BF02478259.
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.
OECD (2026). Venture capital investments in artificial intelligence through 2025. OECD Policy Briefs, No. 50. Paris: OECD Publishing. doi:10.1787/a13752f5-en.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann. doi:10.1016/C2009-0-27609-4.
Renic, N.C. (2019). Battlefield mercy: Unpacking the nature and significance of supererogation in war. Ethics & International Affairs, 33(3), 343–362. doi:10.1017/S0892679419000364.
Rix, H. (2014). The Christmas Truce. National Army Museum, London.
Rumelhart, D.E., Hinton, G.E. & Williams, R.J. (1986). Learning representations by back-propagating errors. Nature, 323, 533–536. doi:10.1038/323533a0.
Russell, S.J. & Norvig, P. (2021). Artificial Intelligence: A Modern Approach. 4th ed. Pearson.
Shardlow, M. & Przybyła, P. (2024). Deanthropomorphising NLP: Can a language model be conscious?. PLOS ONE, 19(12), e0307521. doi:10.1371/journal.pone.0307521.
Silver, D., Huang, A., Maddison, C.J., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529, 484–489. doi:10.1038/nature16961.
Stanford Institute for Human-Centered Artificial Intelligence (2026). The 2026 AI Index Report — Research and Development. Stanford University.
Tsai, I.J., Zarowiecki, M., Holroyd, N., et al. (2013). The genomes of four tapeworm species reveal adaptations to parasitism. Nature, 496, 57–63. doi:10.1038/nature12031.
Turing, A.M. (1950). Computing Machinery and Intelligence. Mind, 59(236), 433–460. doi:10.1093/mind/LIX.236.433.
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. Washington, DC: World Bank.
World Health Organization (WHO) (2025). Global report on neglected tropical diseases 2025. Geneva: WHO. ISBN 978-92-4-011404-3.
World Meteorological Organization (WMO) (2025). State of the Global Climate 2024. Geneva: WMO.
Zarowiecki, M. & Berriman, M. (2015). What helminth genomes have taught us about parasite evolution. Parasitology, 142(S1), S85–S97. doi:10.1017/S0031182014001449.
Zarowiecki, M., Linton, Y.-M., Post, R.J., et al. (2014). Repeated landmass reformation limits diversification in the widespread littoral zone mosquito Anopheles sundaicus sensu lato in the Indo-Oriental Region. Molecular Ecology, 23(10), 2573–2589. doi:10.1111/mec.12761.
Zeisig, B.B., Fung, T.K., Zarowiecki, M., et al. (2021). Functional reconstruction of human AML reveals stem cell origin and vulnerability of treatment-resistant MLL-rearranged leukemia. Science Translational Medicine, 13(582), eabc4822. doi:10.1126/scitranslmed.abc4822.
Zhang, Q., Marcinkiewicz, V., Bowen, L., Kozlowski, T., Inatani, T., Ueda, Y., Katsuno, H., Asada, M. & Morgan, P.L. (2026). Should AI be allowed to break the law? Public acceptance and tolerance of discretionary actions performed by autonomous vehicles in response to situations varying by degree of acuteness. Transportation Research Part A: Policy and Practice, 205, 104885. doi:10.1016/j.tra.2026.104885.