What we want is a machine that can learn from experience. The possibility of letting the machine alter its own instructions provides the mechanism for this.
— Alan Turing, lecture to the London Mathematical Society, 1947
Uncertain Intelligence: A Beginning or an End?
Tim Girling-Butcher • November 16, 2025
A few years before his 1947 lecture, Alan Turing had helped the Allies win World War II by developing the Bombe - a sophisticated, logic-based machine that performed computer-like operations to decrypt German military communications. More than a tool, it was a prototype for what would follow: a framework that shaped Turing’s vision of a machine capable of learning from its own experience.
search
With computers capable of 'intelligence' still many decades away, few paid much attention. Yet the idea had been set in motion - sparking years of speculation, optimism, frustration, and renewal that would eventually make artificial intelligence one of the most consequential dilemmas society would ever have to face.
Just shy of ten years later, the first coordinated effort to study Artificial Intelligence gathered momentum at the Dartmouth Workshop of 1956, where a small group of mathematicians, engineers, and psychologists met to explore the possibility of creating “artificial intelligence.” Turing had died in 1954 so wasn’t part of this group but the proposal itself — a summer research project — now reads as a founding document for the field.
What made the workshop so important wasn’t just the optimism, but the diversity of approaches it set in motion. From it emerged entire disciplines: machine learning, natural language processing, computer vision, knowledge representation, and robotics — each attempting to capture some facet of human cognition in code. The participants believed that reasoning, perception, and language could all be modelled computationally if given enough time and data.
Among them was Marvin Minsky, who in 1951 built the SNARC machine, an early electromechanical neural network that used vacuum tubes and randomly connected synapses to simulate how neurons might learn through experience — arguably the first physical model of a learning machine. Minsky championed the idea that intelligence could emerge from networks of simple processing units, echoing early theories of how neurons interact in the brain.
It was clear that if we could build a machine that learns, then perhaps we could understand how the brain does it.
— Marvin Minsky, reflecting on SNARC (quoted in Pamela McCorduck, Machines Who Think, 1979, p. 111)
By the mid-twentieth century, our understanding of the brain had evolved dramatically. The neuron doctrine, first established in the late 1800s by Ramón y Cajal and Golgi, had shown that the brain was built from individual nerve cells connected through vast, branching networks. Later research revealed how these cells transmitted electrical and chemical signals, forming the biological basis of thought and perception. By the 1940s, scientists such as Warren McCulloch and Walter Pitts began modelling these neurons as simple on–off switches, laying the conceptual foundation for artificial neural networks — machines that might, in time, learn and adapt as the brain does.
Early advocates of neural networks were driven by the simple but powerful intuition that intelligence arises not from explicit rules, but from patterns of connection and adaptation. If the human brain could learn by strengthening or weakening links between neurons, then perhaps a machine could do the same. These researchers believed that cognition might emerge from networks of simple units — each doing little on its own, but collectively capable of recognising, generalising, and learning from experience. Unlike symbolic approaches, which required every rule to be hand-coded, neural networks promised a system that could teach itself through feedback — learning to solve problems by exposure rather than instruction.
The promise was immense, but the timing was premature. The computing power of the 1950s simply couldn’t support the scale these ideas required. Early neural networks could recognise basic patterns, but memory, processing, and data storage were too limited to make learning sustainable. The vision of machines that could truly adapt and reason had to wait for technology to catch up.
Minsky would go on to become frustrated by these limitations and the weaknesses they brought about with early neural networks — his later book Perceptrons (1969), co-authored with Seymour Papert, highlighted their mathematical limitations and effectively paused research in the area for almost two decades.
Within popular culture however, writers and filmmakers like Arthur C. Clarke and Stanley Kubrick were captivated by the potential, collaborating on what remains one of the most authoritative depictions of AI ever created. HAL — the calm, calculating intelligence guiding a mission to Jupiter’s moons — is just as relevant today as it was in 1968, when 2001: A Space Odyssey was released. The questions it raised about autonomy, emotion, control and motivation still feel like the ones we’re asking now.
Progression was further impacted in the early 1970s. The British government commissioned the Lighthill Report (1973) to investigate whether tax funded investment was paying off. The report was devastatingly critical. Lighthill concluded that outside of narrowly defined areas like pattern recognition and robotics, AI had failed to produce tangible results (see a 1973 BBC debate on the subject below).
The Lighthill debate on Artificial Intelligence: "The general purpose robot is a mirage", 1973
This ushered in a period of low investment and slow progress. In the UK, funding for AI research was slashed almost entirely. Across the Atlantic, American agencies like ARPA also pulled back support after a series of disappointing results. By the mid-1970s, much of the early momentum had evaporated — AI entered what became known as its first “winter.”
Through the 1980s and 1990s, AI slipped quietly into utility. Ray Kurzweil’s OCR machine gave blind readers access to printed text, while the Prometheus Project (1989) explored integrated cognitive architectures — early efforts to model reasoning itself. IBM’s Deep Blue (1997) demonstrated that machines could plan and adapt strategically, defeating the world chess champion and marking a turning point in how artificial intelligence was publicly perceived.
By the early 2000s, however, the commercial landscape had shifted. Companies like Google began investing heavily in new approaches to machine learning, seeing potential where decades of symbolic AI had stalled. Geoffrey Hinton - now often referred to as the godfather of AI - and others revived the neural-network model, showing that layered architectures — crude approximations of the human brain — could learn to identify, classify, and generate patterns with remarkable accuracy. Their work demonstrated the power of systems that train and adapt through feedback, delivering far more impressive results than the rigid, hand-coded approaches that had dominated the previous era.
Then came the pocket revolution. With the arrival of Siri in 2011, artificial intelligence became something we could all carry — moving from labs and research papers into everyday life. For the first time, millions of people interacted directly with a machine that could listen and respond in small, useful ways. Although it now seems clunky compared to contemporary platforms like Claude or ChatGPT, AI had quietly entered the public sphere — not through breakthroughs in theory, but through accessibility.
Today, in the generative era, AI no longer just recognises but creates. Systems trained on vast data sets can now produce text, images, sound, and video that blur the boundary between synthesis and invention. It can study us, influence us, titillate us, predict us, and coerce us in ways humans would struggle to match. AI is evolving so quickly that policy and oversight can’t keep pace, and its immediate impact remains difficult to fully comprehend.
Offering enormous and almost endless potential, we are now on the precipice of Artificial General Intelligence. With it could come immense benefits for humanity — the end of suffering, the gradual end of disease, fair and balanced leadership, and freedom from work. Yet it also brings great risk. If control rests only with those who seek to exploit and dominate, it could bring about inequality, suffering, and coercion on a scale humanity has never before imagined.
Because this technology is concentrated in the hands of a few corporations, it raises significant questions about public accountability and the alignment of commercial incentives with the common good. These risks, therefore, cannot be easily dismissed.
What comes next?
The coming years could see social and technological shifts as profound as the past two thousand years — compressed into little more than a decade. Many will be deeply challenged by how artificial intelligence reframes our understanding of the human brain, the idea of a soul, and the fragility of religion — as well as any lingering notions of consciousness or sentience as uniquely human.
As AI replaces increasing amounts of human labour, it will test our sense of identity, purpose, and self-worth. If governments respond with universal basic income or similar models, society will face new forms of upheaval as people seek ways — rational or irrational — to defend their patch, property, and perceived value.
When machines are specifically built to discriminate, rank and categorize, how do we expect to teach them to value equality?
— Mo Gawdat
The concentration of this power creates a significant risk of exacerbating inequality or developing systems that prioritise profit and control over human well-being. History has shown that humanity's previous encounters with such rapid, concentrated power shifts have rarely resolved equitably or without significant short-term disruption.
Our greatest challenge in the years ahead is defending the integrity of informed debate. With AI already being used to skew the information ecosystem, the task of separating fact from fiction—and public interest from competing agendas—is more critical and more difficult than ever.
We are, once again, standing in our own version of the Dartmouth moment: imagining what learning machines might become, and how they might reshape us in return. The difference this time is scale—the experiment is no longer confined to a summer workshop. It is unfolding everywhere, all at once, and in ways that will test not only our ingenuity, but our ability to stay informed, purposeful, and self-aware in the process. The core question, first posed by Turing, remains: Can we build a machine that learns from experience, and more critically, can we learn from the machine without losing what makes us human?
Related
Read more →
Shutting the gate on reality: how AI will build the world around you
The web is becoming an invisible, personalised layer around us - and that may quietly seal us inside our own reality.
If the past is just a dataset, it can be re-weighted
How AI's ability to re-weight our collective history and private memory may create a radical, terrifying, and beautiful possibility: the end of fixed history.
The Age of 'Hacking Humans': Harari on AI and the End of Free Will
Throughout history, leaders have sought new ways to exert influence and control. In Nexus, Yuval Noah Harari explores how AI could shift surveillance from passive monitoring to predictive behavioural control, potentially enabling digital dictatorships unless strong democratic safeguards are put in place now.