The Concept of Singularity: What 2037 Could Mean for Science, Technology and Human Civilisation

Science & A.I. · The future of intelligence

The technological singularity is one of the most discussed and least understood ideas of our time. It is neither the science-fiction apocalypse of the headlines nor a guaranteed date on the calendar. It is a precise, unsettling hypothesis — and the honest answer to when, or whether, it arrives is that nobody knows.

In popular culture, “the singularity” has come to mean something vague: the moment artificial intelligence takes over, for good or ill. In serious discussion it means something sharper and more interesting — a point beyond which the future becomes genuinely unpredictable, because intelligence itself has slipped from human hands.

This article takes the idea seriously without taking it on faith. What is the singularity actually claiming? Where did the idea come from? Why do some brilliant people think it is near and others that it is impossible? And what should a thoughtful person make of a date like 2037?

The goal here is not to predict the future but to equip you to reason about it — to tell the genuine argument from the marketing, the open scientific question from the confident slogan. It is one of the rare topics where the honest answer, “we do not know,” is also the most important thing to understand.

For all that cultural footprint, the singularity has almost no research literature. A title search of arXiv on 1 September 2026 found ten papers with “technological singularity” in the title. Ten. We ran it again on 6 October 2026. Still ten.

The same archive holds 82 papers with “artificial general intelligence” in the title and 19,740 on large language models. The idea that organises a whole genre of prediction has been titled in fewer papers than most graduate seminars generate in a year.

None of that shows the hypothesis is wrong. It shows it is not currently a scientific research programme, and that is the single most useful thing to know before reading anyone’s date.

What a singularity actually is

The word comes from mathematics, where a singularity is a point at which a function becomes infinite or undefined — a place where the equations break down and stop giving sensible answers. It is a signpost marking the edge of what a theory can describe.

Physics borrows the term for similar reasons. At the centre of a black hole lies a gravitational singularity where the known laws fail. The Big Bang is described as emerging from a cosmological singularity. In each case the word marks a boundary beyond which our best theories simply cannot see.

The technological singularity takes this idea and applies it to the future of intelligence. It names a hypothetical point beyond which technological change becomes so rapid and so profound that human beings can no longer predict or perhaps even comprehend what follows. The unpredictability is the whole point — not a detail but the definition.

The idea’s origin: the intelligence explosion

The core argument is older than the personal computer. In 1965 the British mathematician I. J. Good, who had worked alongside Alan Turing at Bletchley Park, set it out with striking clarity in a paper on what he called an ultraintelligent machine.

Good reasoned that a machine surpassing human intelligence in every domain could design even better machines. Those machines could design better ones still, and so on, in a runaway cascade he named an intelligence explosion. In his memorable phrasing, such a machine would be “the last invention that man need ever make”, provided it remained under control.

That final clause has haunted the field ever since. The idea was given its modern name in 1993 by the computer scientist and novelist Vernor Vinge, in a talk at a NASA symposium in Ohio. “Within thirty years,” he wrote, “we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.”

Vinge was unusually precise about his own forecast. “I’ll be surprised if this event occurs before 2005 or after 2030,” he wrote. He died in March 2024, with his window still open. It now has a little over four years left to run, and whether it closes empty is one of the cleaner tests the idea has ever offered.

He also saw more than one road there. A superhuman intelligence might be built as a computer, he suggested, or large computer networks and their users might “wake up” together, or the link between people and computers might become so intimate that the users themselves became superhumanly intelligent, or biology might find ways to improve the human brain directly. Only the first is what most people now mean by the singularity.

A cautionary history of predictions

Before trusting any timeline, it is worth remembering how badly the field has forecast itself. When artificial intelligence was founded as a discipline in the mid-1950s, its pioneers were wildly optimistic, predicting machines with human-level general intelligence within a generation.

The founding document is a good example. In 1955 John McCarthy and three colleagues proposed a two-month, ten-man study at Dartmouth College for the summer of 1956, “on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” Their list of goals already included making machines that “improve themselves”. They thought “a significant advance” could be made in a summer. Seventy years later, the goals are still being worked on.

Those predictions failed. Progress proved far harder than expected, and the excessive promises were followed by periods of disappointment and lost funding that became known as the “AI winters”. Time and again, tasks that looked nearly solved turned out to conceal enormous difficulty.

This history is a warning in both directions. It shows that confident predictions of imminent machine intelligence have a poor track record — but also that dismissing AI entirely has been just as mistaken, since the field has repeatedly rebounded with genuine breakthroughs. Humility about timelines is the only lesson that has aged well.

The engine: recursive self-improvement

At the heart of the singularity hypothesis is a single mechanism: recursive self-improvement. If a system can improve its own ability to solve problems, then each improvement makes the next one easier, and progress could in principle accelerate without obvious limit.

Human progress is bounded by the speed of human thought and the length of human lives. A machine intelligence would face no such constraints. It could, in theory, redesign itself in hours rather than generations, compounding its capabilities at a pace no biological mind could match or even follow.

Whether real systems would actually behave this way is fiercely disputed. Some argue the feedback loop is inevitable once a threshold is crossed; others that each improvement gets harder, not easier, and that diminishing returns would tame any explosion. The mechanism is clear in theory; its behaviour in practice is not.

Why 2037? Reading the forecasts honestly

Specific dates for the singularity should be treated with deep caution. They are forecasts about an event that, by its own definition, is unpredictable, made by people with strong incentives to be memorable. A year such as 2037 is best read not as a prediction but as a marker within a wide and contested range.

The best-known optimist, the inventor Ray Kurzweil, has long argued that human-level machine intelligence will arrive around 2029 and a full singularity around 2045, based on his reading of exponential trends in computing. He restated both dates in his 2024 book The Singularity Is Nearer. Others place it sooner, later, or never. A date in the late 2030s sits squarely in the middle of this crowded field.

Surveys of artificial-intelligence researchers reveal just how little agreement there is. The largest, published in 2025, asked 2,778 researchers who had published at top AI conferences. Taken together, their answers gave a 10 per cent chance that unaided machines would outperform humans at every task by 2027, and a 50 per cent chance by 2047, thirteen years earlier than the same team’s survey a year before.

The same respondents put a 50 per cent chance on every human occupation becoming fully automatable only by 2116. Individual answers ranged from a few years to well over a century. Between 38 and 51 per cent of them, depending on how the question was asked, gave at least a 10 per cent chance to outcomes as bad as human extinction. The only honest summary is radical uncertainty, held by the people who know the field best.

The survey even offers a number for 2037 itself. The researchers’ combined forecast gave a 10 per cent chance that every human occupation would be fully automatable by that year. That is the closest thing to an expert verdict on the date in this article’s title: possible, but far from the likeliest outcome.

So any single year, including 2037, should be held lightly. It is a way of focusing the mind on a plausible near-future scenario, not a countdown. The useful question is not exactly when, but whether the trends that make such forecasts tempting are real — and there the evidence cuts both ways.

The case that it might be near

The optimists point to a remarkable run of progress. For decades, computing power grew exponentially, a pattern captured by Moore’s law, the observation that the number of transistors on a chip roughly doubled every couple of years. That relentless doubling made once-impossible computations routine.

On top of that hardware came startling advances in artificial intelligence. Systems learned to beat the best humans at the ancient game of Go, long considered a benchmark of intuition, and later to generate fluent language, images and code. The rise of the transformer, the architecture behind modern large language models, unlocked capabilities that surprised even their creators.

To those impressed by this trajectory, the direction seems clear. Each year brings systems that do more of what was thought to require human intelligence. Extrapolate the curve, the argument goes, and a machine that can improve itself no longer looks like fantasy but like the next step on a well-worn path.

What the Trend Lines Actually Show

Some recent measurements are more concrete than any forecast. The research group METR timed skilled people on a large set of software tasks, then asked how long a task an AI system could complete with a 50 per cent success rate. In early 2025 the best systems managed tasks that took people about 50 minutes, and that “time horizon” had been doubling roughly every seven months since 2019.

The authors were careful about what that means. Their tasks were software tasks, chosen to be measurable, and real work is messier. If the trend held and generalised, they wrote, AI systems would be able to automate many software tasks that now take people a month within about five years. Both of those ifs are doing a lot of work.

There is now a real, if modest, example of the engine the singularity depends on. In 2025 Google DeepMind described AlphaEvolve, a system that uses language models to write and test improvements to code. Among other results, it sped up part of the training of the very model it runs on, and found a way to multiply two four-by-four grids of complex numbers using 48 multiplications, the first improvement in 56 years, for that case, on a method devised by the mathematician Volker Strassen.

That is AI improving the process that makes AI. It is also a long way from runaway self-improvement: humans chose the problems, set the tests and decided what to keep, and the gains were percentages, not explosions. The International AI Safety Report, written by more than 100 experts led by Yoshua Bengio and published in February 2026, laid out the range plainly. Between now and 2030, it said, progress could slow or stall, continue at current rates, or accelerate dramatically if AI systems begin to speed up AI research itself.

The case that it is far off — or impossible

The sceptics are equally serious, and their objections are not easily dismissed. First, they note that exponential trends do not continue forever. Moore’s law has been slowing as chip features approach the size of atoms, and no trend in nature grows without limit.

Second, they argue that today’s impressive systems, for all their fluency, do not understand the world as humans do. They are extraordinary pattern-matchers trained on vast data, and it is far from proven that scaling them up leads to genuine general intelligence rather than ever more sophisticated imitation.

Third, and most fundamentally, no one knows what human-level general intelligence really requires, because no one fully understands how our own minds work. Building something we cannot yet define is a shaky basis for confident forecasts. The history of artificial intelligence is also a history of premature predictions that did not come true.

Some go further and argue the singularity may be impossible in principle, whether because of physical limits on computation, the nature of intelligence itself, or the possibility that improvement gets exponentially harder rather than easier. These are not fringe positions; they are held by respected researchers.

There is also the sheer messiness of the real world to reckon with. An intelligence explosion imagined on a whiteboard glides past the friction of physical experiments, energy costs, data limits, regulation and economics. A mind confined to a computer cannot instantly rebuild the world; it still has to act through a slow, resistant reality. Intelligence alone may not be the master key the story assumes.

Two of those frictions are now measurable. The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, around 1.5 per cent of the world’s total, and projects roughly 945 terawatt-hours by 2030, driven largely by AI. Power stations and grid connections take years to build, whatever the software can do.

Data is the other. A 2024 study by researchers at Epoch AI estimated that if current trends continue, AI models will be trained on datasets as large as the entire stock of public human-written text sometime between 2026 and 2032. The authors suggested that synthetic data, learning from other kinds of data and more efficient training could keep progress going. Whether they can is one of the real open questions behind every forecast.

The slippery question of “intelligence”

Much of the debate turns on a word we cannot precisely define. “Intelligence” is used as though it were a single quantity that a machine could simply have more of, like horsepower. In reality it is a bundle of many different capacities — reasoning, perception, memory, creativity, social understanding — that do not always travel together.

Today’s systems make this vivid. An AI can write a passable essay yet fail at simple physical reasoning a child finds trivial. It can beat any human at chess while having no understanding of why the game matters. The 2026 International AI Safety Report noted that leading systems excel at difficult mathematics and science problems yet still struggle with tasks such as counting objects in an image, reasoning about physical space and recovering from basic errors in longer workflows. Narrow brilliance and broad common sense have turned out to be very different things.

This matters for the singularity because the whole argument assumes a general intelligence that can improve itself across every domain at once. If intelligence is not one thing but many, the smooth curve of an intelligence explosion may be far bumpier, or may not materialise at all. The concept the hypothesis relies on is itself unsettled.

If it did happen, what then?

Suppose, for the sake of argument, that a genuine intelligence explosion occurred. The consequences are, by definition, hard to foresee, but the possibilities are usually sketched along a spectrum from the utopian to the catastrophic.

In the hopeful vision, a superintelligence aligned with human values solves problems that have defeated us for millennia — curing disease, reversing environmental damage, unlocking clean energy and extending healthy life. Intelligence, on this view, is the ultimate tool, and more of it means more of everything good.

In the darker vision, a superintelligence pursuing goals subtly misaligned with ours could be catastrophic, not from malice but from indifference, in the way human projects can flatten an ant hill without a thought. Between these extremes lies the most likely reality: a messy, mixed outcome that no simple story captures.

The alignment problem

This is why much serious work now focuses not on building superintelligence but on ensuring that any powerful AI system reliably does what its designers intend. The philosopher Nick Bostrom set out the core worry in a 2012 paper: a system’s intelligence and its final goals are independent, so a highly capable machine could pursue almost any objective, including ones disastrous for humans. He called this the orthogonality thesis.

The concern is not robots turning evil, a Hollywood distraction. It is the difficulty of specifying what we actually want precisely enough that a relentless optimiser cannot satisfy the letter of our instructions while violating their spirit. Getting a powerful system to genuinely share human intentions is a deep and unsolved technical problem.

This alignment problem is now a mainstream research priority at the leading AI laboratories. Whatever one believes about timelines, the case for solving it in advance is strong: if powerful systems do arrive, the time to have understood how to control them is before, not after.

Transcendence: the transhumanist dream

For some enthusiasts the singularity is not only about machines but about people. The broader transhumanist movement anticipates that advancing technology could radically enhance human beings themselves — extending lifespan dramatically, augmenting the brain, and merging human minds with artificial intelligence.

The most striking of these ideas is mind uploading: the hypothetical transfer of a human mind into a computer, promising a kind of digital immortality. Others focus on brain–computer interfaces that could let people think alongside machines directly. In this vision the singularity is a moment of human transcendence, not replacement.

These proposals are far more speculative than the core intelligence-explosion argument, and they run into profound scientific and philosophical problems, not least our near-total ignorance of how consciousness arises. They are best understood as aspirations and thought experiments rather than forecasts — a reminder of how much hope, and how much projection, the singularity absorbs.

How to think about it sensibly

The healthiest stance toward the singularity avoids two traps. One is breathless certainty that it is imminent and inevitable, a belief that shades easily into a kind of secular religion. The other is dismissive certainty that it is pure science fiction, which ignores genuine and rapid progress.

A more defensible position treats the singularity as a serious possibility of uncertain probability and uncertain timing — worth thinking carefully about precisely because the stakes, if it happened, would be so high. Uncertainty is a reason for attention, not for dismissal or panic.

Whatever the timeline, artificial intelligence is already reshaping the world in concrete ways, from medicine to employment to warfare. Attending to those real, present effects is not a distraction from the long-term question; it is the best preparation for it. The future arrives one capability at a time, not all at once on a chosen date.

There is also a quieter reason to focus on the present. Whether or not a singularity ever comes, the choices being made now — about how AI is built, governed and deployed — will shape which futures remain open. Waiting for a dramatic threshold can distract from the ordinary decisions that actually steer the technology.

In that sense, the value of the singularity as an idea is less as a prophecy and more as a provocation. It forces us to ask what we want from intelligence, how we would keep powerful systems accountable, and what we owe to a future we cannot fully see. Those questions are worth answering regardless of whether the year 2037, or any year, ever earns the name.

Perhaps that is the most honest note on which to end. The singularity may be a coming revolution, a persistent mirage, or something in between that we will only recognise in hindsight. What we can do now is stay clear-eyed: curious about the possibility, sceptical of the certainties, and serious about the choices already in our hands.

Frequently asked questions

Sources

Foundational works and peer-reviewed research:

• Good, I. J. (1965). Speculations concerning the first ultraintelligent machine. Advances in Computers, 6, 31–88. doi.org/10.1016/S0065-2458(08)60418-0
• Moore, G. E. (1965, reprinted 2006). Cramming more components onto integrated circuits. IEEE Solid-State Circuits Newsletter. doi.org/10.1109/N-SSC.2006.4785860
• Silver, D., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529, 484. doi.org/10.1038/nature16961
• Vaswani, A., et al. (2017). Attention is all you need. arXiv. doi.org/10.48550/arXiv.1706.03762
• Grace, K., et al. (2018). When will AI exceed human performance? Evidence from AI experts. Journal of Artificial Intelligence Research, 62, 729. doi.org/10.1613/jair.1.11222
• Grace, K., et al. (2025). Thousands of AI authors on the future of AI. Journal of Artificial Intelligence Research, 84, 9. Preprint: doi.org/10.48550/arXiv.2401.02843
• Kwa, T., et al. (METR) (2025). Measuring AI ability to complete long software tasks. NeurIPS 2025. arXiv:2503.14499.
• Novikov, A., et al. (Google DeepMind) (2025). AlphaEvolve: a coding agent for scientific and algorithmic discovery. arXiv:2506.13131.
• Bengio, Y., et al. (2026). International AI Safety Report 2026. internationalaisafetyreport.org.
• McCarthy, J., Minsky, M. L., Rochester, N. & Shannon, C. E. (1955). A proposal for the Dartmouth summer research project on artificial intelligence. jmc.stanford.edu (PDF).
• International Energy Agency (2025). Energy and AI: energy demand from AI. iea.org.
• Villalobos, P., et al. (2024). Will we run out of data? Limits of LLM scaling based on human-generated data. arXiv:2211.04325.
• Bostrom, N. (2012). The superintelligent will: motivation and instrumental rationality in advanced artificial agents. Minds and Machines, 22, 71 (DOI: 10.1007/s11023-012-9281-3).

Key books and essays:

• Vinge, V. (1993). The coming technological singularity: how to survive in the post-human era. NASA VISION-21 Symposium.
• Kurzweil, R. (2005). The Singularity Is Near: When Humans Transcend Biology. Viking Press.
• Kurzweil, R. (2024). The Singularity Is Nearer: When We Merge with AI. Penguin.
• Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.