Soumya Banerjee — University of Cambridge
Few technologies have captured the imagination, or attracted the anxiety, of the early twenty-first century as thoroughly as artificial intelligence. In academic research, in industry, and in public discourse the term appears with a frequency that often outstrips understanding. This post is an attempt to remedy that imbalance for a specific audience: students and early-career researchers who are excited by AI, who want to work with it, and who deserve a clear account of both its genuine promise and its substantial risks.
The central argument is straightforward. AI is a powerful family of methods for finding patterns in data, making predictions, and automating decisions. It is not magic, it is not infallible, and it is not a substitute for human judgment. Understanding what AI actually is — including its historical roots, its current capabilities, and the mechanisms by which it can fail or cause harm — is not merely intellectually satisfying; it is a practical necessity for anyone who will build, deploy, or be governed by AI systems over the coming decades.
Artificial intelligence is most usefully understood as the science and engineering of building systems that can perform tasks which, when performed by humans, would be described as requiring intelligence. These tasks include recognising speech, translating languages, diagnosing diseases from medical images, playing complex games, and generating text that reads as though it were written by a person.
The foundational inspiration for AI is biological: the human brain. Nature has spent hundreds of millions of years producing nervous systems that are extraordinarily good at perceiving the world, learning from experience, and deciding how to act. The central question of AI, posed by its pioneers, is whether the computational principles underlying biological intelligence can be extracted and implemented in machines. Early answers took the form of the perceptron, introduced in the late 1950s, which modelled a single neuron as a linear threshold unit. The perceptron demonstrated that a machine could adjust its own parameters to improve its performance on a task — a property we now recognise as the seed of machine learning.
A question that has accompanied AI from its earliest days is: can machines think? The computer scientist Edsger Dijkstra offered a memorable response:
“The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.”
The point is not dismissive; it is a call for precision. A submarine moves through water by mechanisms utterly unlike those used by a fish; we do not say a submarine swims because swimming implies a biological mode of locomotion. Similarly, when an AI system beats a world chess champion or writes a convincing essay, it does so by mechanisms — principally statistical pattern matching over large datasets — that are quite unlike the mechanisms by which a human plays chess or writes. Whether the word “thinking” applies to such mechanisms is a question for philosophers; the engineering and scientific questions are substantially more tractable and more important.
The history of AI is also, in part, a history of the human tendency to attribute intelligence to mechanisms that are nothing of the kind. This history warns against the uncritical anthropomorphisation of AI systems — a tendency that remains very much alive today.
In 1673, the Jesuit polymath Athanasius Kircher invented the statua citofonica, the “talking statue,” a device that conveyed sound through hidden tubes so that a statue appeared to speak. In 1739, Jacques de Vaucanson unveiled his Digesting Duck, an automaton that appeared to eat grain and digest it. The duck was widely regarded as a marvel of natural philosophy; in fact, the “digestion” was achieved by a hidden compartment. More consequentially, the Mechanical Turk, constructed by Wolfgang von Kempelen in 1770, appeared to play chess autonomously and defeated many opponents over several decades. The Turk was, in fact, operated by a skilled human chess player concealed within its cabinet.
These cases share a common structure: a mechanism produces behaviour that observers interpret as evidence of intelligence, when in fact the behaviour arises from hidden, mundane causes. The lesson for the contemporary moment is that the tendency to treat impressive outputs as evidence of genuine understanding or consciousness is as present today — in discussions of large language models — as it was in the salons of Enlightenment Europe.
Arthur C. Clarke’s third law holds that any sufficiently advanced technology is indistinguishable from magic. The corollary that matters for AI is that magic is performed by people. The scientists and engineers at large technology companies who design, train, and deploy AI systems have an interest in maintaining the mystery. Understanding AI, opening its “black box,” is therefore not merely an intellectual exercise; it is a democratic and ethical imperative.
The current era of AI is dominated by two related developments: deep learning and generative AI.
Deep learning refers to neural networks with many layers of computation, trained on very large datasets. Since the early 2010s, deep learning has produced dramatic improvements in image recognition, speech recognition, natural language understanding, and protein structure prediction. The success of deep learning rests on three convergent factors: the availability of large labelled datasets, hardware accelerators such as graphics processing units, and algorithmic innovations including attention mechanisms and residual connections.
Generative AI is a class of models trained to produce new content — text, images, audio, video, or code — that is statistically consistent with their training data. Large language models have demonstrated striking capabilities in generating coherent prose, answering questions, writing code, and summarising documents. Image generation models can produce photorealistic images from text descriptions.
Applications span virtually every sector. In healthcare, AI is used for medical image analysis, clinical decision support, drug discovery, and patient risk stratification. In finance, AI powers fraud detection, algorithmic trading, and credit scoring. In autonomous systems, deep learning underlies self-driving vehicles and robotic warehouses. In science, AI is accelerating discovery from particle physics to climate modelling.
The same capabilities that make AI useful also make it risky. Here are five categories of risk that every student and researcher should understand.
AI systems learn patterns from data. When those data reflect historical inequities — in lending, hiring, healthcare, or criminal justice — the system will encode and propagate those inequities at scale. A facial recognition system trained predominantly on images of light-skinned individuals will perform less accurately on people with darker skin tones. A language model trained on the internet will absorb and reproduce internet biases. These are not merely theoretical concerns; they have caused measurable harm in real deployed systems. Addressing bias requires deliberate effort at every stage of the AI pipeline: data collection, model training, evaluation, and post-deployment monitoring.
AI systems require vast quantities of data for training, much of it personal in nature: medical records, financial transactions, browsing histories, and private communications. The aggregation and use of such data at scale raises serious questions about consent, surveillance, and the right to privacy. Individuals often have limited knowledge of how their data is used and limited recourse when AI systems make consequential errors about them.
AI and automation have the potential to displace workers across a wide range of occupations, including many previously considered resistant to automation. Routine cognitive tasks such as data entry and document processing are increasingly being automated. More sophisticated AI is beginning to affect knowledge work — legal research, medical diagnosis, software development. The distributional effects depend heavily on policy choices about education, social insurance, and the regulation of AI deployment.
The development of frontier AI systems requires enormous resources: vast training data, expensive computing infrastructure, and large teams of skilled researchers. These requirements mean that the most capable AI systems are being developed by a small number of large technology companies. This concentration raises questions about who controls AI, whose values are encoded in AI systems, and whether the benefits of AI will be distributed broadly or captured by a narrow elite. The historical analogy of the East India Company — and the concentration of economic and technological power that came with it — serves as a cautionary reminder of what can happen when power becomes dangerously concentrated. We should not want a repeat of that.
Perhaps the most immediate and concrete risk is the inappropriate substitution of machine judgment for human judgment in high-stakes decisions. The UK Post Office Horizon scandal is instructive. Between approximately 1999 and 2015, the Post Office prosecuted around 736 postmasters for theft, fraud, and false accounting on the basis of evidence provided by the Horizon IT system, developed by Fujitsu. Many of these individuals were convicted; some served prison sentences. Their concerns that the discrepancies were caused by faults in the Horizon system were dismissed, because the Post Office treated the system’s outputs as authoritative. Subsequent investigations established that Horizon was deeply flawed, and hundreds of wrongful convictions are being overturned.
The takeaway is not that IT systems are inherently untrustworthy. It is that decisions with major consequences for individuals’ lives must be made by accountable human beings who understand the limitations of the systems they use. An AI system is a tool; responsibility for the decisions made with that tool remains with humans.
A subtler but pervasive risk is the tendency to attribute human qualities — understanding, intentionality, consciousness — to AI systems. This tendency is actively encouraged by the way AI products are designed and marketed: systems are given human names, voices, and conversational interfaces that invite users to relate to them as persons rather than programs. Anthropomorphisation leads users to trust AI systems in domains where that trust is not warranted and to abdicate personal responsibility to systems that are not genuinely capable of bearing it.
A few examples from my own work illustrate how AI methods are applied to real scientific problems.
Hierarchical Bayesian models of infection — work published in the Journal of the Royal Society Interface used probabilistic models to characterise the dynamics of infectious disease at individual and population levels. Hierarchical models are particularly suited to problems where observations are grouped into clusters — patients within hospitals, households within communities — because they allow information to be shared across groups in a principled way.
Clinical informatics for serious illnesses — in collaboration with clinicians and statisticians, machine learning was applied to stratify patients with serious illnesses on the basis of clinical and genomic data, with careful attention to data security and ethical governance. Published in Nature Communications. The machine learning is only as useful as the clinical judgment brought to bear on the results.
Predicting disease — work in Gut applied machine learning to predict the onset of gastrointestinal disease from clinical features. Translating predictive performance into clinical benefit requires careful evaluation in real-world conditions, attention to how the model performs across different patient subgroups, and sustained engagement with the clinicians who will use it.
Interpretable machine learning in psychiatry — work in npj Schizophrenia developed class-contrastive, human-interpretable methods for predicting mortality in patients with severe mental illness. The emphasis on interpretability is a practical and ethical imperative: clinicians and patients need to understand not just what a model predicts, but why, in order to evaluate whether to act on it. Explainable AI is a prerequisite for responsible deployment in high-stakes clinical settings.
Machine learning in supply chains — the same underlying methods transfer across domains. Supply chain forecasting is particularly challenging because the relevant causal relationships are complex, data are noisy, and decisions must be made under uncertainty at multiple time horizons.
The common thread across all of this work is that AI succeeds when it is embedded in deep domain knowledge and genuine collaboration — not when it is applied as an off-the-shelf solution to a problem nobody has bothered to understand properly.
Two directions are likely to be particularly important over the coming decade.
Explainable AI (XAI) aims to make AI systems’ reasoning processes more transparent and interpretable. Current deep learning models are, in many cases, highly accurate but opaque — they produce correct answers without being able to articulate the reasons in terms humans can understand or verify. This opacity is a significant obstacle to deployment in medicine, criminal justice, and autonomous vehicles, where accountability requires that decisions can be explained and contested.
Human-AI collaboration focuses on designing systems and workflows that combine the complementary strengths of human and machine intelligence. Humans bring contextual judgment, ethical reasoning, creativity, and accountability; AI brings speed, scale, and the ability to detect subtle statistical patterns in large datasets. The most productive applications of AI in the near term will augment rather than replace human judgment, keeping humans genuinely in the loop.
India’s relationship with technology-driven economic transformation has a clear recent precedent. The information technology boom of the 1990s demonstrated that a large, young, and technically trained population can rapidly seize the opportunities created by a technological transition. The current AI moment offers an analogous opportunity.
The IndiaAI Mission and the Viksit Bharat 2047 vision articulate an ambition for India to be not merely a consumer of AI technologies developed elsewhere, but a global contributor to their creation. Several considerations are particular to the Indian context:
India’s civilisational depth, its mathematical and scientific traditions, and the entrepreneurial spirit of its people are resources that no amount of capital alone can replicate.
As Napoleon once said, forty centuries look down upon us.
Cultivate curiosity. The most important quality a researcher can bring to any field is a genuine desire to understand how things work, not merely to use tools that other people have built. In a field that produces impressive-sounding results at high speed, intellectual honesty requires courage.
Collaborate and work in teams. The most impactful AI research — particularly in applied domains such as healthcare and climate science — is done in teams that combine machine learning expertise with deep domain knowledge. A machine learning researcher who does not understand the clinical context of a medical AI problem cannot build systems that are useful in practice. Collaboration across disciplines is not a sign of weakness; it is a prerequisite for doing consequential work.
Focus on basic concepts, not buzzwords. The vocabulary of AI changes rapidly; the terms fashionable today will be superseded by new vocabulary in a few years. The underlying mathematical and statistical concepts — probability theory, linear algebra, optimisation, information theory — do not become obsolete. Researchers who understand these foundations can adapt to new methods as they emerge; those who have only learned to use specific tools will find their expertise dated quickly.
Artificial intelligence is a genuine scientific and technological achievement whose applications are producing real benefits across a wide range of domains. It is also a source of serious risks that require clear-eyed analysis and deliberate governance. The risks are not, for the most part, the science-fiction risks of superintelligent machines pursuing their own agendas; they are the more mundane but more immediate risks of biased systems, opaque decision-making, concentrations of power, and the substitution of machine outputs for human accountability.
Navigating these risks and realising the genuine benefits requires people who understand AI well: who know how models are built and evaluated, who recognise the ways in which AI systems can fail, and who bring the ethical commitments and domain expertise needed to deploy them responsibly.
The appropriate watchword is not “artificial insanity” or “artificial stupidity,” but what AI at its best actually stands for: a tool that amplifies human capability, built and governed by humans who remain accountable for what it does.
Soumya Banerjee is a researcher at the University of Cambridge working on machine learning and its applications in healthcare and beyond. Contact: sb2333@cam.ac.uk