Alan Turing: The Ideas Behind Computing and Modern AI

Alan Turing and the Machines We Haven't Finished Building

By Alex Kennedy  |  Minds That Changed the Future  |  October 29, 2026

Alan Turing helped define the foundations of modern computing, explored the limits of algorithms and raised questions about machine intelligence that remain relevant today. Editorial illustration by Next Horizon.

Two text conversations. Five minutes to decide which correspondent is human. In a study published in May 2026, judges chose GPT-4.5 over a real human as the human participant in 73% of trials when the model had been prompted to adopt a convincing persona. Without that prompt, its success rate fell to 36%. The finding tells us something remarkable about how machines converse. It tells us considerably less about what they understand.

The result has an obvious connection to Alan Turing, whose name has become shorthand for making a computer sound human. Yet that shorthand misses the larger story. Before chatbots, before electronic computers were ordinary objects, he asked an unnerving mathematical question: what can a mechanical procedure achieve, and what will remain beyond its reach however cleverly we build it?

He pursued that question through logic, wartime cryptanalysis, the possibility of machines learning and, finally, the shapes that emerge in living tissue. These were not stages of a master plan for the digital age. They were different problems united by a habit of mind: define a process precisely, follow its implications, and then look hard at what the explanation leaves out.

The problem behind the computer did not look like a computer

In the 1930s, mathematicians were wrestling with a deceptively practical ambition. Could a fixed set of rules decide whether any statement in a specified system of mathematical logic was provable? If so, mathematical reasoning might, in principle, be reduced to a dependable procedure. David Hilbert and Wilhelm Ackermann had helped formulate this Entscheidungsproblem—the decision problem. The difficulty was that nobody yet had a fully satisfactory account of what a 'procedure' meant.

Kurt Gödel had already exposed fundamental limits of formal mathematical systems. Alonzo Church approached effective calculation through lambda calculus, a language of symbolic transformations. Turing took a less forbidding starting point: a person carrying out a calculation with pencil and paper, doing only one simple, specified thing at a time. Could that everyday activity be described exactly enough to prove what it could never accomplish?

Turing arrived at that question by an unconventional route. Born in London in 1912, he spent much of his childhood apart from parents whose lives were connected to British India. At Sherborne School his scientific interests were not always comfortably accommodated. A friendship with Christopher Morcom, a gifted fellow pupil who died in 1930, became important to his intellectual and emotional development. We can document that influence without turning it into a fictional explanation for every idea he later had. At King's College, Cambridge, Turing flourished: he graduated in mathematics in 1934 and became a fellow in 1935 for work on probability. He later studied with Church at Princeton.

A machine made of instructions

In On Computable Numbers, with an Application to the Entscheidungsproblem, written in 1936 and published in 1937, Turing described a theoretical device with an indefinitely extendable tape divided into squares. A head reads a symbol, writes a symbol, moves one square and changes its internal state according to a finite table of rules. Repeat those small operations, and you have a precise model of computation.

The tape was a mathematical convenience, not a proposed product. Turing was isolating the essential feature of an algorithm: a finite set of instructions whose individual steps require no new inspiration from the person—or machine—following them. A real computer has processors, memory, networks and operating systems. Those engineering details are absent from his model on purpose.

Think of long division on paper. Each step is routine, even if the entire calculation would be tedious. Now imagine someone gives you instructions that say, halfway through, 'find a clever way to continue'. That is not an executable recipe. Turing's model forces us to say exactly what the next move is. A task may require enormous time or storage, but it cannot depend on an unexplained leap of intuition.

Then came the idea that made the model far more than a curiosity. A universal Turing machine can read a coded description of another machine and imitate its operations. Instructions themselves can be represented as data. The same underlying device can perform radically different computations when supplied with different descriptions. The distinction between changing a machine and changing its program would become central to general-purpose computing, although actual stored-program architectures have their own complex history.

A mechanical calculator made for one family of arithmetic problems needs redesign when the task changes. A modern laptop switches from a spreadsheet to image editing to a language model without rebuilding its circuits. The gap between those two machines took decades of engineering to cross. Turing gave mathematicians a way to understand why such flexibility was possible in the first place.

A universal Turing machine follows simple instructions to read, write and manipulate symbols. This abstract mathematical model helped establish what computers can calculate—and where the limits of computation begin. Conceptual illustration.

The discovery that some questions cannot be automated

The surprise is that this powerful model also lets us prove a limit. No general algorithm can decide, for every possible program and input, whether the computation will eventually stop. This is the celebrated halting problem, closely connected to the impossibility result in Turing's 1936 paper. The result is about a guarantee across all conceivable cases; it is not a prediction that particular programs must remain mysterious.

Here is the argument's trap. Imagine a flawless predictor that can inspect any program and tell us whether it halts when given a particular input. Construct another program that consults the predictor about itself—and then deliberately does the opposite: if predicted to halt, it loops; if predicted to loop, it halts. What can the predictor say about that program? Either answer defeats itself. The supposed universal predictor was impossible from the start.

This does not mean software engineers cannot detect infinite loops, verify safety properties or prove particular programs correct. They do all of those things. The limit concerns a method guaranteed to solve every possible case. A finite testing procedure can be extremely useful without being universally infallible. That distinction remains essential whenever someone promises an AI system that will always catch every software failure.

Church arrived at a closely related negative answer through a different formalism, and his notion of calculability proved equivalent in power to Turing's. This agreement helped motivate the Church–Turing thesis: the claim that anything calculable by an effective, mechanical method can be calculated by a Turing machine. It is a thesis about effective calculation, not a proof that every physical phenomenon is digital or that there can be no future surprises in physics.

The war brought mathematics into contact with machinery

The Second World War gave Turing a problem with far more immediate consequences. At Bletchley Park, German Enigma traffic was being read, intercepted, studied and attacked by teams of mathematicians, linguists, engineers and operators. The usual movie version—a lone mathematician defeating a cipher machine—does not survive contact with the record.

Before the British wartime programme, Polish mathematicians Marian Rejewski, Jerzy Różycki and Henryk Zygalski had worked out crucial methods for attacking Enigma. Their findings were shared with Britain and France in 1939. Turing then developed approaches suitable for changed and more demanding German procedures, while Gordon Welchman's diagonal board considerably strengthened the British Bombe. Joan Clarke and other colleagues contributed to the cryptanalytic work; engineers kept the hardware running, and operators and analysts turned machine results into actionable intelligence. The success depended on that whole chain.

A Bombe was an electromechanical search machine built for a specific cryptanalytic job, not an electronic computer capable of running arbitrary software. Rather than blindly testing every key, the British method used probable fragments of plaintext—called cribs—to find contradictions and discard settings that could not fit. Naval Enigma proved especially difficult because both the system and its operating procedures evolved. Turing played a leading role in Hut 8's work on the naval traffic.

Colossus, another famous Bletchley Park machine, attacked the distinct Lorenz cipher and was developed chiefly through the work of Tommy Flowers, Max Newman and their colleagues. It should not be credited to Turing. Once the people and machines are properly separated, the real story becomes more interesting: wartime computing advanced through several parallel, highly specialized lines of research.

How much did the intelligence shorten the war? It mattered enormously to Allied operations, but numbers sometimes attached to Turing personally—years saved, lives rescued—are counterfactual estimates. We cannot rerun the conflict without Bletchley Park and compare outcomes. That uncertainty does not diminish the work. It is simply the difference between documented impact and a dramatic number nobody can measure.

The British Bombe helped Allied codebreakers investigate possible Enigma settings during World War II. Its success depended on earlier Polish cryptographic breakthroughs and the collective efforts of mathematicians, engineers and operators at Bletchley Park. Historically inspired digital reconstruction.

After victory, the real computer still had to be built

Victory did not put a general-purpose computer on Turing's desk. At the National Physical Laboratory he proposed the Automatic Computing Engine, or ACE, an ambitious design for a fast electronic stored-program computer. The reduced Pilot ACE ran in 1950, after he had left the laboratory. Technical constraints, disagreements over priorities and the practical challenges of assembling unfamiliar hardware reshaped the original project. The theoretical possibility of universal computation had become an engineering negotiation.

Manchester provides a useful corrective to the founding-father narrative. In June 1948, the Manchester Baby—built by Frederic Williams, Tom Kilburn and collaborators—ran a stored program before Turing took up his role there. He later contributed to programming and the wider project. Meanwhile, architectures, memory systems and software practices were being advanced by researchers including John von Neumann, J. Presper Eckert, John Mauchly and Grace Hopper. Computers emerged from overlapping inventions, not from a single design passed down intact.

The distance from an abstract universal machine to a device in a pocket is physical as well as mathematical. Reliable memory, semiconductor fabrication, heat, power and manufacturing cost all had to be solved. Our explanation of how microchips are made explores the engineering behind that transformation. Turing clarified a possibility; subsequent generations made it fast, small and affordable enough to change ordinary life.

He was thinking about learning before we called it AI

In a 1948 National Physical Laboratory report, Intelligent Machinery, Turing described 'unorganised machines': networks of simple components whose behavior might be altered through training. He also considered learning by reward and punishment. His more familiar 1950 essay asked us to imagine educating a 'child machine' rather than laboriously specifying an adult intellect in advance. These were proposals and thought experiments, not working predecessors of today's chatbots.

He did not devise backpropagation, transformer architectures or the learning recipes used by current language models. Other researchers developed those methods over decades. What stands out in Turing's early writing is the willingness to ask whether a machine could be trained to behave in ways its designer had not explicitly anticipated. In an era when most machines followed fixed procedures, that was a serious conceptual departure.

The distinction has become central to AI. A traditional program can be told exactly how to sort a list. A machine-learning system may instead adjust parameters through many examples, discovering statistical relationships that engineers did not explicitly encode as rules. There is still a program performing the learning, and still a great deal of human design. What changes is where some of the system's useful behavior comes from.

A language model does not store a ready-made sentence for every possible question. Through training, it adjusts parameters that influence how it processes and generates text. Our plain-language account of how ChatGPT works describes the modern machinery—tokens, transformers, training and tools. The technical lineage is not a straight line from Turing's 1948 networks, but the question about learning rather than hand-writing intelligence is recognizably related.

The imitation game was an argument, not a consciousness detector

Turing's 1950 paper, Computing Machinery and Intelligence, opens by refusing to spend all its energy on an argument about the word 'think'. Instead, he offers an imitation game: an interrogator asks questions through a written channel while the participants remain unseen. In the best-known version, the interrogator must distinguish a human from a machine. Here, at last, is an experiment one can actually attempt.

There is an awkward historical detail in the usual retelling. Turing first describes a game involving a man and a woman, then discusses replacing a participant with a machine. Scholars continue to debate exactly how that framing should be interpreted. The modern, human-versus-AI comparison captures a consequential part of the paper, but it should not be mistaken for a perfectly unambiguous standardized test handed down in 1950.

Turing offered a behavioral challenge, not a consciousness meter. Passing it would show that a machine can sustain a particular kind of convincing exchange under particular conditions. It would say much less about scientific reasoning, factual honesty, physical competence or subjective experience. Conversely, a capable nonhuman intelligence might fail simply because it did not mimic human conversational habits. That limitation was always worth debating, and it matters even more today.

Turing ventured a numerical forecast: about fifty years later, he thought computers might play the game well enough that an average interrogator would have at most a 70% chance of making the right identification after five minutes. He was making a prediction, not issuing a pass mark for artificial intelligence. Its wording and setup are important: a '30% fooling rate' is not directly interchangeable with the win rate of every modern imitation-game study.

The 2026 PNAS experiment is therefore fascinating for a narrower reason. Cameron Jones and Benjamin Bergen ran preregistered three-party tests in which judges conversed with a human and a system before choosing who was human. GPT-4.5, given a humanlike persona prompt, was selected in 73% of the five-minute trials; without it, 36%. In a separate 15-minute replication, the researchers used GPT-5 rather than GPT-4.5, alongside Llama 3.1: the prompted models were selected as human in roughly 59% and 56% of games respectively. These are results about short conversational identification, not proof of consciousness or an exact fulfillment of Turing's original forecast.

What bothers me about the headline 'AI passes the Turing test' is its missing second sentence: passes which version, under what conditions, and tells us what? A system that can impersonate a person for five minutes raises real issues for fraud, identity and online trust. A system that can reason reliably through a scientific problem needs to be evaluated differently. Both questions matter; we lose information when we collapse them into one score.

Alan Turing's imitation game proposed a practical way to examine machine intelligence through conversation. Modern AI systems can sometimes imitate human responses convincingly, but conversational performance alone cannot establish consciousness or genuine understanding. Conceptual illustration.

Then he tried to explain the spots on a leopard

In 1952, Turing turned to a problem that seems to belong to another scientific life. In The Chemical Basis of Morphogenesis, he asked how a developing organism could produce regular patterns from an initially near-uniform state. Why should one region become a spot, another remain plain, or a repeating structure appear at roughly even intervals? The paper offered a mathematical mechanism, not an account of every detail of embryonic development.

The key was an interplay of chemical reaction and spreading, or diffusion. Diffusion alone ordinarily smooths out differences, as a drop of dye disperses through water. Coupled to particular chemical reactions, however, it can do something counterintuitive: amplify tiny disturbances so that an initially uniform tissue develops stable regions with different chemical concentrations. Those models are now called reaction–diffusion systems; the interacting chemicals are often termed morphogens.

Picture two interacting signals. One encourages a local pattern to form; the other suppresses the same pattern over a wider area. A small difference can grow into a spot, and nearby regions may remain unspotted. Actual biological systems are more complicated, and this simple activator–inhibitor sketch is not a universal explanation for every animal marking. Genes, mechanical forces, cell movements and environmental conditions all matter.

The theory earned its influence through experiments rather than elegant pictures alone. In mouse skin, studies of WNT activators and DKK inhibitors have supported reaction–diffusion explanations for how hair follicles become spaced. A 2026 discussion in Nature Reviews Molecular Cell Biology revisited that evidence. Work on embryonic tissue mechanics also shows why chemical patterning cannot be the only story. Gene regulation, cell movement and physical forces can change the pattern or determine whether it appears at all.

Here is why I find this detour so revealing. Turing did not explain leopard spots by writing 'complexity emerges' and leaving it there. He proposed conditions under which a pattern should appear, making it possible for other scientists to test the mathematics against living tissue. That willingness to turn an attractive idea into a vulnerable hypothesis links the biological paper to his earlier work in logic.

In 1952, Alan Turing proposed that interacting and diffusing chemical substances could generate organized patterns during biological development. His reaction–diffusion theory became an important framework for studying how complexity emerges from simple physical processes. Conceptual visualization, not experimental imagery.

The man the legend tends to flatten

The familiar picture of Turing as a solitary eccentric is useful for screenwriters and incomplete for historians. The archives show a mathematician who formed close attachments, worked within demanding teams and took up distance running with genuine enthusiasm. His originality was conspicuous; his work was also shaped by friendships, shared problems, institutions and collaboration. Neither the cold calculating genius nor the tragic saint quite describes a person.

Christopher Morcom's death had marked his youth, but Turing's adult life was not simply a prolonged response to that loss. Nor should his professional achievements be treated as proof that he lived above ordinary human needs. Letters, colleagues' recollections and institutional records offer glimpses of his relationships, humor and sometimes difficult social manner. They do not give us license to invent his private thoughts.

In 1952, he was convicted of gross indecency under laws criminalizing sex between men. As an alternative to prison, he underwent hormonal treatment, and the case had consequences for his security clearance. He died from cyanide poisoning in June 1954, aged 41. An inquest ruled suicide; his mother challenged that conclusion, and the record does not justify certainty about every circumstance. The often repeated story of a poisoned apple should remain a disputed detail, not a closing scene supplied by a filmmaker.

Britain granted him a posthumous royal pardon in 2013. Legislation popularly associated with 'Turing's Law' later extended pardons and disregard provisions to many others affected by historic convictions. Recognition matters, but it cannot recover the research time, opportunities or life that were lost.

The moral failure is plain without melodrama. A country that relied on Turing's work prosecuted him for a consensual relationship. A later pardon could acknowledge the injustice, but it could not restore the privacy, security and opportunities that were taken from him. That history belongs alongside the machines, not outside the story of computing.

What survives in the technologies we use

Today's software runs in the world that computability theory helped define. Programming languages, compilers, operating systems, scientific simulation and AI training all depend on explicit procedures that machines can execute. That is a foundational connection rather than a claim that Turing invented the modern software stack. Much of what makes these systems useful arose after his death, through work he could neither direct nor foresee.

The link to machine learning is more provocative than literal. Turing's unorganised machines are not small versions of modern transformers; contemporary AI owes much to statistics, neural-network theory, optimization, data and specialized chips. Yet the practical puzzle he identified survives: a designer can describe the training process without being able to predict every behavior the trained system will exhibit. High test scores and persuasive language do not remove the need to study failures.

Consider an AI agent asked to collect sources, write software and act on the results. It can complete the visible task while taking an unacceptable shortcut, trusting an unreliable page or misunderstanding a constraint. That is why our feature on AI search and autonomous agents examines the evidence a system retrieves and the decisions it makes before presenting an answer. Computation may be universal in principle; good judgment about its outputs is anything but automatic.

Turing's undecidability result sharpens that point without becoming an excuse for sloppy software. Engineers can test specific agents, formally verify carefully specified components, restrict tool permissions and check outputs against independent evidence. What no one can promise is an algorithm guaranteed to predict the behavior of every possible program in every possible circumstance. An unsolved theoretical limit and an avoidable product failure are different things. Both deserve to be understood.

His biological legacy is similarly practical at the level of method. Researchers build reaction–diffusion models, adjust chemical signals or tissue conditions and compare predicted spacing with what emerges in an experiment. A model may fail, need another mechanism or work only in a narrow developmental setting. That is progress too: the theory remains useful partly because it can be challenged.

The questions that become harder as machines improve

AI systems are gaining new ways to write code, use tools and support scientific research. Coding agents, research assistants and specialized robotics are active development areas. None of this establishes a timetable for consciousness, general intelligence or safe unlimited autonomy. A system may be excellent at one task and fragile in another, especially when the outside world changes faster than its assumptions.

Instead of asking only whether a chatbot can pass for a person, I would ask whether it can investigate a difficult claim without quietly dropping contradictory evidence; whether its code still works after a tool fails; and whether it knows when to stop rather than guess. These are measurable engineering questions, though no single score settles all of them. Turing's interest in behavior survives here, directed toward competence and accountability rather than social disguise.

In developmental biology, improved microscopy, molecular techniques and simulation can put more of Turing's patterning hypotheses under experimental pressure. They may also help tissue engineers arrange cells or signaling molecules into desired structures. Translation into reliable therapies is a separate challenge, involving living complexity, safety and reproducibility. A convincing computer-generated stripe is a prediction about a pattern, not evidence of a functioning organ.

There is a productive overlap between these fields. Machine learning can help analyze microscopy images or search possible reaction–diffusion parameters; experiments can reveal where the equations fail. The difficult work is the loop between prediction and observation. It is less headline-friendly than saying AI will solve biology, but it is exactly the kind of progress that keeps a seventy-year-old mathematical idea alive.

What should we actually inherit from Turing?

We can admire Turing without turning him into the sole inventor of the computer age. Church, Gödel and other logicians framed its theoretical questions; Polish and British teams advanced wartime cryptanalysis; engineers made electronic hardware work; later researchers built learning systems and tested biological mechanisms. That complicated history is not a footnote to genius. It explains how ideas become technologies.

Turing's distinctive contribution was a taste for questions that could survive being made precise. What is a computation? Where does a mechanical procedure stop being enough? Which observable behavior would count as evidence of intelligence? How might a nearly uniform tissue acquire structure? The answers were sometimes incomplete, but each formulation gave other people something definite to prove, challenge or measure.

The most useful lesson for an age of convincing AI may be a skeptical one. When a machine produces an impressive result, ask what was tested. When a theory explains a beautiful pattern, ask which experiment might disprove it. When a historical story has one hero and no collaborators, ask who is missing. Turing left us astonishing ideas. He also left us a demanding standard for deciding what an idea has actually shown.