The Future Does Not Move at One Speed: A Review of Martin Rees’s On the Future

Books about the future have an unusual relationship with time. Ordinary nonfiction becomes outdated around the edges; predictions acquire a second life as evidence of what once seemed plausible. Martin Rees’s On the Future: Prospects for Humanity, first published in 2018, is now nearly eight years old—long enough for that process to have begun.

What is striking is not how much of the book has dated, but how selectively. Its discussions of climate change, energy, biotechnology, scientific research and long-term risk remain largely recognisable. Artificial intelligence is the obvious place to look for failure: almost everything about the public experience of AI in 2026 feels different from 2018. Yet rereading Rees more carefully complicates that judgement. He anticipated professional and cognitive work being automated while occupations such as plumbing and gardening remained difficult. What he did not anticipate was the speed, breadth and accessibility with which generative AI would make that transition visible.

That makes On the Future more interesting to read now than a simple tally of correct and incorrect predictions would suggest. Climate systems, power stations, demographic shifts and software do not develop on the same timetable. A book that places them all under the heading of “the future” inevitably contains futures moving at different speeds.

Cover image: On the Future: Prospects for Humanity by Martin Rees, published by Princeton University Press. Used here for purposes of review and commentary.

Technical Optimism, Political Pessimism

Rees has described himself as a “technical optimist” and a “political pessimist”, and the combination gives the book much of its character. He neither assumes that technological progress will automatically solve our problems nor treats new technologies as inherently suspect. Greater scientific capability expands what humanity can do in both directions: it offers cleaner energy, better medicine and more productive agriculture while also increasing the damage that mistakes, conflict or malicious actors can cause.

His response is generally to favour more science rather than less of it: more research, stronger scientific institutions, international cooperation and a willingness to investigate difficult technologies before circumstances force decisions on us. I find that outlook persuasive. Scientific uncertainty is often used politically as an argument for delay, when uncertainty can just as easily be an argument for learning more.

Energy is a good example. In a published keynote developing many of the book’s arguments, Rees describes wind and solar as front-runners but argues for research across a much wider range of low-carbon technologies. He remains ambivalent about nuclear fission while supporting research into newer reactor concepts, and considers fusion’s potential payoff large enough to justify continued experiments and prototypes. I liked the balance. Neither nuclear power nor fusion has to become a complete answer before research into it becomes worthwhile.

The point is not that one favoured technology will rescue us. Electricity systems have to combine generation, storage, transmission, demand and reliability under real physical and institutional constraints. I have made a similar argument in looking at the energy transition as a system problem: deliberately shrinking the technological menu because particular technologies acquire political identities makes a difficult problem harder.

Rees approaches climate policy in much the same way. Science can estimate warming, model consequences and compare technologies, but it cannot determine how strongly the interests of future people should weigh against present costs. In a Harvard interview around the book’s publication, he focused on discount rates: the economic mechanism by which future costs and benefits are valued less than present ones.

Discount rates sound like technical machinery until one asks what they encode. How much less should damage suffered by someone in 2100 matter because that person happens to live later than we do?

How Much Is the Future Worth?

Suppose preventing serious environmental damage late this century requires substantial expenditure today. Part of the problem is scientific: how much warming will particular emissions produce, how will ecosystems respond and which technologies can reduce the damage? Another part is ethical. How much should someone alive now sacrifice for someone who may be born fifty or eighty years from now?

I find the case for taking the longer view compelling. I have never understood the attitude that what happens after one’s own death is somehow irrelevant. Children and grandchildren make the inadequacy of that position obvious, but I do not think moral concern should stop when we reach descendants we will never personally meet. The future of humanity matters even if none of the people experiencing it will know our names.

Rees does not carry this into the strongest versions of contemporary long-term moral reasoning, but his argument invites the extrapolation. A civilisation surviving for thousands or millions of years could contain vastly more lives than exist today. If intelligent life eventually spreads beyond Earth, the possible numbers become larger still. Give those hypothetical people exactly the same moral weight as people alive now and small changes in the probability of a vast future can begin to dominate almost any present-day calculation.

That is where long-term thinking becomes uncomfortable. A principle meant to correct our obvious bias toward the present can, taken far enough, allow uncertain future populations to overwhelm the needs of actual people. I do not think the answer is to assign the distant future zero value; the uncertainty makes exact moral arithmetic less convincing, not the continued existence of civilisation less important.

Space makes the tension concrete. Rees rejects the idea that Mars provides an escape from environmental problems on Earth. In a 2018 interview discussing the book, he calls mass emigration unrealistic and argues that even hostile environments on Earth are incomparably easier to inhabit than anywhere else in the Solar System. On that point I agree completely. Terraforming Mars is not an alternative to keeping Earth habitable.

His position on settlement is more interesting than a simple opposition to human spaceflight, however. Rees is happy to see privately funded adventurers establish small settlements on Mars or asteroids, and speculates that their descendants might deliberately alter themselves through genetics and cybernetics before eventually becoming electronic or otherwise post-human intelligences capable of spreading beyond the Solar System. He is sceptical of spending large amounts of public money to send ordinary humans into space; he is not sceptical about intelligence eventually expanding far beyond Earth.

My emphasis differs slightly. I place more value on human settlement itself as part of that progression, rather than seeing it mainly as a transitional stage toward beings better adapted to space. But the difference is much narrower than it first appears. Earth can be irreplaceable without being the permanent boundary of civilisation.

Technology Without Technological Tribalism

The same refusal to sort technologies into moral categories appears in Rees’s treatment of biotechnology. He worries considerably about biological risk, particularly as powerful techniques become cheaper and more widely accessible. Around the time of the book he regarded misuse of biotechnology as a more immediate danger than AI, in part because relatively small facilities or groups could cause consequences that spread far beyond them.

Yet this concern does not become hostility to biotechnology itself. Rees explicitly includes genetically modified crops among the technologies that may be needed to feed a larger and more demanding population. I appreciated that stance because GMO debates have often treated the method as though it settles the argument. It does not. A crop engineered for drought tolerance, disease resistance, altered nutrition or some other trait should be judged according to what the modification does, the risks it creates and the alternatives available. “Genetically modified” tells us how an organism was changed; it does not tell us whether growing it is wise.

Rees applies essentially the same standard elsewhere. Nuclear technology can provide useful energy and create serious risks. Biotechnology can improve medicine and agriculture while making new forms of misuse possible. Artificial intelligence can extend human capability while redistributing wealth, authority and vulnerability. The sensible unit of judgement is usually the application and the system around it, not the technological label.

This helps explain why research occupies such an important place in the book. A society faced with uncertain technologies benefits from understanding them better. The complication is that some technologies give institutions far less time to learn than others.

AI Aged Differently

My first reaction was that artificial intelligence was the part of On the Future that had aged badly. Almost everything about AI in 2026 makes 2018 feel distant: conversational systems, generated images and video, coding assistants, large-scale use inside organisations, and software capable of carrying out increasingly complicated sequences of cognitive tasks.

On closer inspection, though, Rees was more prescient about work than that impression gives him credit for. He expected machines to move beyond manufacturing into routine legal work, accountancy, computer coding, medical diagnostics and possibly surgery. By contrast, he identified plumbing and gardening as among the harder jobs to automate because they require non-routine interaction with an unpredictable physical environment.

That is remarkably close to one of the surprises of the generative-AI era. The conspicuous technological advance has not been humanoid robots becoming excellent plumbers. It has been software becoming increasingly competent at prose, programming, image manipulation, document analysis and other activities once treated as evidence of sophisticated cognition. Meanwhile, manipulating an unfamiliar physical environment remains a stubbornly different engineering problem.

There is still an irony here. Many of us intuitively assumed that writing an acceptable essay required a more generally intelligent machine than cleaning an unfamiliar kitchen. The development path suggests that our intuitive hierarchy of difficult tasks was unreliable. Rees, to his credit, was already warning that apparently prestigious professional work might be easier to automate than messy physical competence.

What his discussion does not capture is the form in which that automation arrived. The systems of 2018 were generally discussed as specialised algorithms, robots or decision tools. Generative AI made a wide range of capabilities accessible through the same ordinary interface: language. A user no longer needs a separate program designed specifically to summarise a report, draft an email, explain code, analyse an image and brainstorm a proposal. Increasingly, the same system can move among all of them.

That changes the discussion of work even when it does not immediately eliminate jobs. Occupations are bundles of tasks, organisations adopt technology unevenly, and the presence of an automatable activity does not mean an entire profession disappears. The 2026 Stanford AI Index describes labour-market effects as uneven, with early pressure concentrated in some hiring pipelines and among younger workers in exposed occupations, while large-scale employment losses have not appeared across the economy as a whole.

The speed of diffusion is nevertheless striking. Stanford estimates that generative AI reached 53 per cent population-level adoption within three years and that 88 per cent of surveyed organisations were using AI by 2025. Whatever the final employment equilibrium, a discussion of education, professional work and labour-market adjustment written today would have to devote much more attention to AI than seemed necessary in 2018.

So I would no longer say that Rees simply got AI wrong. The task hierarchy aged surprisingly well. The scale, generality and speed of the systems implementing it did not.

Why Some Futures Move Faster Than Others

That difference points toward a larger difficulty with forecasting. “The future” is not one category of problem.

Climate change is governed by physical systems with enormous inertia. Carbon dioxide accumulates, oceans absorb heat and buildings, vehicles and power infrastructure turn over slowly. Demographic changes unfold across generations. New energy systems require mines, factories, transmission lines, planning approvals, specialised workers and large amounts of capital. Even a spectacular fusion breakthrough would not supply a continent with electricity the following morning; an energy industry would still have to be constructed around it.

Software can move under different constraints. AI is not immaterial—frontier systems depend on specialised chips, data centres, electricity and enormous capital investment—but once a useful capability exists, distributing access to millions of users can happen far faster than replacing a national vehicle fleet or constructing a generation of nuclear reactors. The physical infrastructure may take years to build while the capability running on it changes several times during the same period.

This is one reason I remain cautious about treating technological acceleration as a single universal curve. In my review of Ray Kurzweil’s The Singularity Is Nearer, I found the evidence for accelerating computational capability much stronger than the implication that medicine, energy, institutions and human welfare must inherit the same rate of change. Technologies interact, but they do not all run on the same clock.

Rees himself effectively demonstrates the point. His climate discussion from 2018 still feels contemporary because many of the relevant physical and political constraints remain. His AI chapter belongs visibly to an earlier technological moment even where its underlying judgement about susceptible occupations was good. The forecast did not fail uniformly because the systems being forecast were never moving uniformly.

That distinction matters for politics as well. Slow problems at least give institutions time to adapt, even when they waste it. Rapid technologies can change the environment in which policy is made before legislation, education systems and professional norms have finished responding to the previous version. The difficult task is not merely predicting the correct endpoint. It is identifying which domains can move quickly enough that waiting for certainty guarantees that the response arrives late.

A Broad Book About a Narrow Problem

On the Future covers an enormous territory in a relatively short book: climate, energy, nuclear weapons, biotechnology, artificial intelligence, inequality, scientific research, space and the very long-term prospects of intelligent life. That breadth necessarily limits depth. A reader already immersed in climate economics, AI, energy systems or biotechnology will repeatedly reach the point where another fifty pages would have been welcome.

I do not think that is a serious weakness. The value of the book lies partly in putting subjects beside one another that specialists normally separate. Many of them were already familiar to me individually; seeing them on the same timeline makes their common problem clearer.

Humanity’s technical capacity is increasing faster than its ability to coordinate the consequences.

That is where Rees’s technical optimism and political pessimism meet. We possess, or can plausibly develop, technologies capable of supplying cleaner energy, improving agriculture, extending healthy lives and increasing human productive capacity. The same civilisation possesses nuclear weapons, increasingly powerful biological techniques, cyber vulnerabilities and AI systems whose social consequences are difficult to anticipate. Science expands the range of futures available to us without choosing among them.

Politics then adds a mismatch of time horizons. Research programmes may require decades of funding before they produce anything useful. Climate policies impose present costs partly for the benefit of people who have not yet been born. Preventing a low-probability catastrophe produces little visible political reward when the prevention succeeds. Global problems require cooperation among governments whose incentives remain overwhelmingly national and immediate.

Fusion is almost a caricature of the problem. Its reputation for remaining perpetually some decades away makes an easy joke, but abandoning long-horizon research because success cannot be scheduled would guarantee that difficult technologies never arrive. Rees’s willingness to keep such options open is one of the things I liked most about the book. Research does not need to come with a promise of deployment on an electoral timetable before it has value.

The sections that have dated therefore do not weaken On the Future as much as one might expect from a book carrying such a dangerous title. If anything, reading it nearly eight years later makes its central problem more concrete. Even a cautious and scientifically informed forecaster cannot know which field will suddenly compress decades of expected change into a few years.

That leaves institutions with a different objective from perfect prediction. They need enough scientific knowledge to recognise changing conditions, enough flexibility to alter course, enough technological breadth to preserve options, and a time horizon extending beyond the people currently making the decisions.

On the Future remains worth reading because most of the future it discusses has not disappeared. Some of it has simply moved much farther than the rest.

The future does not advance at one speed. The harder question is whether our institutions can recognise when one part begins to accelerate away from everything around it.