Anticipating and navigating strategic technological surprise
Technologies do not have to be secret to be surprising. And across sixteen of the cases the literature is built from, how well a development was forecast turns out to tell you almost nothing about whether anyone was ready for it. Fifteen of the sixteen were not.
Maria Langan Riekhof
Herb Lin
The many technological advances underway in artificial intelligence, bioengineering, quantum technologies, energy innovations, and space capabilities are accelerating the pace and scale of change — introducing an era of outsized possibilities and uncharted risks.
New discoveries have always been disruptive, from the wheel and the printing press to the automobile and the internet. What is different now is convergence: advancing technologies plus strong economic incentives are producing new applications and disruptions at breathtaking speed, forcing individuals, societies, institutions, and governments to adapt.
China’s unveiling of DeepSeek was a strategic shock to the U.S. policy community — not because AI was unexpected, but because it overturned the assumption that China could not produce advanced, globally competitive AI at scale.
Once new technologies are unleashed, it may be too late to shape their applications or prepare for their implications. Strategic warning of their emergence, unintended applications, and unanticipated consequences is increasingly essential.
01 / The premise
A disruption becomes strategically significant when it overturns prevailing assumptions, alters the course of events, shifts advantage between actors, and demands a major response. Watch what happens to a planning assumption when the world stops agreeing with it.
Institutions plan against an extrapolation: capability grows at roughly the rate it grew last year, and the actors who hold it stay the same. The line is rarely written down as an assumption. It simply becomes the floor everyone stands on.
Human nature is to assume that the future will resemble the past
The break can come from a discovery, from an unexpected actor, from an old tool used in a new way, or from adoption moving faster than norms. The event itself is often visible in advance. What is missing is the expectation that it could happen now.
Soviet atomic test, August 1949 — at least five years ahead of estimates
The cost of surprise is not the event. It is the distance between where your institutions were prepared to operate and where they now must operate — paid in reallocated resources, new institutions, and policy written under pressure.
Sputnik, 1957 → ARPA founded 1958
02 / The test
From a review of historical cases, four conditions separate a development that is merely startling from one that is strategic. A signal has to clear all four gates. Scroll to send one through.
A rival, a partner, a non-state actor — or no actor at all. In most historical cases the perpetrator intended to catch someone off guard. But a technology can be so transformational that the actors matter less than the impact, and some disruptions arrive with no human intention behind them.
Japan · USSR · al-Qa’ida · HAMAS · China
The surprised party must be a decision maker responsible for shaping policy — not an observer who missed an interesting fact. Warnings may have arrived and still failed: too early, deemed low probability, crowded out by immediate threats, or dismissed because they did not fit the agenda.
Warned is not the same as prepared
Deep and widespread disruptive capacity in at least one domain, overturning the existing equilibrium, standard practice, or norms. Falling costs and open-knowledge ecosystems mean smaller states, non-state actors, and individuals can now reach for effects once reserved for great powers.
September 11th — modest means, decades of consequence
A surprise becomes strategic when it warrants a major response: significant reallocation of resources, institutional change, or policy reform. Which means the threshold is partly discretionary. Surprise is controlled not only by the development, but by the reaction of policymakers to it.
The response is part of the definition
“Developments that significantly and unexpectedly invalidate core planning assumptions, alter a strategic domain, and thereby require a major response.”
Strategic technological surprise, defined
03 / The framework
Technology can produce strategic surprise anywhere along its life cycle — from development, to application, to effects that appear decades later. Each category traces a different curve. Scroll to draw them onto the same axes.
Rare, but real: a breakthrough that fundamentally alters accepted scientific understanding and opens new lines of inquiry. At the moment of discovery, reports are hard to distinguish from error — and a nontrivial share of claims advertised as revolutionary turn out to be hype that does not survive scrutiny.
Penicillin 1928 · Prions 1982 · Cold fusion, had it worked
The capability was on the list. The date was wrong. A breakthrough rises to strategic surprise when its timing is not anticipated or appreciated — the estimate and the event part company, and the plan is calibrated to the estimate.
Soviet atomic bomb, Aug 1949 — 5+ years early
Surprise from the emergence of a technology is less likely in today’s open information environment. Surprise from its use is not. Technology designed for one purpose gets applied in non-obvious ways, in other places and contexts, sometimes transforming a domain it was never aimed at.
Consumer quadcopters → reconnaissance and munitions delivery
The same S-curve, compressed. Early transformational technologies spread over generations; recent ones cross the world in a few years. Even when a technology and its applications are anticipated, the speed and scale of adoption produce outsized disruption before standards can catch up.
Electric light: ~70 years to most U.S. households. Smartphones and LLMs: not that.
An initial disruption can be small enough to go unnoticed, then catalyze a much larger reaction years later. Or an invention solves its problem well and creates a different one somewhere else — in a sector it was never deployed in, forcing new regulation and new norms.
Printing press 1430s → Europe-wide ~65 years later · Asbestos · Plastic
The hardest surprises to forecast are contingent on other technologies and social dynamics arriving at the same time. The internet was designed to network timesharing mainframes; reaching ordinary citizens required personal computers, smartphones, and wireless data networks to show up too.
Network + device + spectrum = the backbone of the modern world
04 / A case in full
Take one signature from the framework — unintended application — and follow a single technology all the way through. A consumer quadcopter was designed to do a handful of things. Here is the job it actually holds.
Consumer quadcopters were developed as inexpensive platforms for aerial photography and hobby use. That is the whole design brief. Everything on the specification sheet, everything in the marketing, everything the engineers optimised for lives in this first group.
Six tasks · all of them anticipated
Off-the-shelf models turned out to do real-time reconnaissance, artillery spotting, and grenade or small-munition delivery. None of that required a new machine. It gave small units and non-state actors precise, ubiquitous airpower that no procurement process had authorised or predicted.
Four more tasks · none on the spec sheet
The last group is the part that is easiest to miss when you count the cost of a surprise. Militaries had to rethink tactics, force protection, and electronic-warfare priorities — work that exists only because a hobby aircraft was pointed somewhere new. The surprise is not the drone. It is this column.
Three more · created by the surprise itself
05 / The horizon
Looking ahead, acceleration is likely to produce surprising strategic effects across multiple domains at once — more consequential to stability, security, and prosperity than traditional surprises confined to the military domain.
Advances increasingly catalyze other advances, creating an accelerating cycle of innovation. CRISPR-Cas9 made programmable gene editing accessible to far more researchers, accelerating biotechnology, agriculture, and therapeutics at once. Computation has run the same loop for decades: better chips fund better chips.
Compound microscope → mass spectrometry → CRISPR
Medical breakthroughs, social media, and AI have catalyzed political debate and deepened societal division. Content — real and fake — can motivate political behavior, erode trust in institutions, polarize societies, and weaken public support for international action. Without domestic support, even a technologically sophisticated military struggles to sustain operations.
Facebook: a college messaging system → a primary source of news
Innovation is disrupting supply chains, labor forces, and consumer markets, creating new winners and losers. Historically, displaced work was replaced: automobiles ended the work of stable hands and farriers and spawned whole complementary industries. Future disruptions are likely to be faster, more uneven, and may not spawn the replacement.
Faster · more uneven · not necessarily replaced
Large-scale kinetic surprise remains possible but probably infrequent. The growth is in civilian technology applied by state and non-state actors to produce frequent, smaller national security surprises that cumulatively carry strategic weight — and to widen the range of gray-zone tactics available below the threshold of open war.
Deep fakes · AI-enabled disinformation · gray-zone coercion
Technology has become a central element of strategic competition, and many civilian technologies carry unexpected geopolitical significance. Sometimes first-mover advantage decides it. Sometimes the advantage goes to whoever scales production and sells to everyone else.
China, EVs and renewables → market share and soft power
06 / The candidate set
A surprise is easier to absorb when the possibility was at least on a list. So here is a list: across six domains, three developments that look probable inside a decade and three that are longshots. These are not forecasts and the odds are not the point. The point is that every one of them is sayable now, which means none of them can later be called unforeseeable.
The probable set is unglamorous and close: fully unmanned factories, supervised autonomy on borders, robots inspecting the pipes and reactors nobody can safely enter. The longshots are what people picture first — humanoids that work, robots an amateur can program, soft robotics that survive contact with the world.
Constraint named here: systems engineering, not intelligence
Reusable heavy launch at a fraction of today’s cost per kilogram, habitation on the moon, manoeuvring without paying for it in fuel. This is also the one domain on this board where the United States is not playing catch-up — and where the longshots turn less on physics than on whether anyone can break a single company’s monopoly.
Orbital data centres · moon business models · life on exoplanets
A fully autonomous lab — hypothesis, synthesis, test, repeat — sits in the probable column, alongside brain-computer interfaces and gene therapy that is actually accessible. The longshots move the hospital into the house and the organ supply into production. Each one relocates medicine away from the institution that holds it now.
Home pharmacy · organisms engineered for climate · organs on demand
The probable calls here all describe one shift: world models that make general-purpose robots work, intelligence available without a network, capability folded into every object until it disappears into infrastructure the way electricity did. And one more that belongs on the list and rarely makes it — that people reject it outright.
Longshots: consciousness · machine-brain fusion · AI that regulates itself
A cryptographically relevant quantum computer, and a collapse in the cost of security vetting, sit beside a third entry that is not a capability at all: people losing trust in security, in truth, in digital information as such. The counter to that erosion belongs in the longshot column. The damage is probable; the repair is not.
Trust listed as probable. Its remedy listed as a longshot
Some entries carry enormous demand and poor prospects — fusion, geoengineering, militarised AGI. Worth asking whether chasing those crowds out duller options that would actually work. And worth asking what counts: does a breakthrough have to arrive at scale, or is being credibly on the path enough to change how everyone else behaves?
Their sharpest question: is a longshot a different kind of change, not a slower one?
“Technical feasibility is rarely the binding constraint. What binds is trust, energy, compute, raw materials, talent, and public willingness to let the technology in at all.”
What actually gates a breakthrough
07 / The record
The framework above asks whether decision makers were unprepared. That is two questions, not one: was the possibility forecast, and was anything done about it? Scoring this paper’s own cases on both axes separates them — and the two turn out to be close to unrelated.
Down the side: was the possibility inside the range analysts were assigning real probability at the time? Across the bottom: were institutions positioned to act without a scramble? Every case in the source paper gets a score on both. The codings are judgments, and they are all listed underneath this chart so you can disagree with any of them.
16 cases · 1941 to 2025 · all drawn from the paper
At one end sit the cases where the warning was explicit and repeated: a global pandemic before COVID-19, war with Japan before Pearl Harbor, the fragments that existed before September 11th. At the other sit developments genuinely outside the distribution — prions overturning the requirement for DNA or RNA, or DeepSeek overturning an assumption about industrial capacity.
Spread on this axis: 0.06 to 0.88
This is the finding. Fifteen of sixteen cases sit in the unprepared half, and most are bunched near the floor. Whatever the forecast said, institutions were positioned about equally badly. The distribution along the bottom of this chart is the argument for the resilience chapter, not the warning chapter.
15 of 16 below the midpoint on readiness
Ukraine in 2022 is the only case in the prepared half, and it got there through eight years of institutional work after 2014 — not through a better forecast. Which is the uncomfortable implication: on this record, improving the forecast is not what moves an institution along the axis that matters. Doing the preparation is.
Highest-forecast cases score lowest on readiness
Every case here is a case where the past was prologue — where a pattern existed to be recognised. If the next disruption genuinely has no historical analog, this chart cannot anticipate it, and neither can a framework assembled from it. That objection lands. But notice which finding survives it: recognising the pattern is what fails, and readiness never depended on recognising the pattern.
The horizontal axis survives the objection. The vertical one does not.
Think of a technology you expect to surprise us — cryptographically relevant quantum computing, engineered pathogens, autonomous weapons, whatever you actually worry about. Score it on the same two questions the sixteen historical cases were scored on, and see where your reading lands among them.
These scores are ordinal judgments, not measurements. There is no accepted scale for “was it forecast” or “was it prepared for,” so the numbers below encode a reading of the historical accounts, coded to two rules. Forecast: was the specific development inside the range relevant analysts were assigning meaningful probability at the time it occurred — not whether it was imaginable in principle. Prepared: were the institutions that had to respond positioned to do so without major reallocation, new bodies, or policy written under pressure. Cases marked contested are ones where a careful reader would plausibly code differently, and the chart’s conclusion should not rest on them. Every case is drawn from the source paper; the sources noted as added in the second table are not, and are flagged as such in the figure.
The second table is the counter-list for figure 06 — warnings that did not land and claims that did not survive. It is illustrative, not exhaustive. No one maintains a register of forecasts that failed in this domain, which is precisely why a base rate cannot be computed. The source paper concedes the same point about breakthrough claims, noting that “a nontrivial number” turn out to be hype, without quantifying it.
08 / The denominator
A framework built only on surprises that arrived cannot tell you how often warning works, because the failures of the other kind — the alarms that never rang true — are missing from the sample. The paper invites exactly this sort of extension, describing itself as a foundation rather than a complete solution.
These are the cases the framework is built from, and each earned its place by having happened. That is a reasonable way to derive a taxonomy of what surprise looks like. It is not a way to find out how well anyone detects surprise, because nothing that was detected in time, or warned about and never arrived, is eligible for the list.
Sampling on the dependent variable
Cold fusion, which the paper names itself. The 2002 estimate of an active Iraqi weapons programme — a warning that was simply wrong. Grey goo, walked back by the person who raised it. Strong AI, forecast at roughly twenty years away in every decade since 1960. These are not embarrassments to be avoided; they are the other half of the evidence.
Illustrative only — nobody keeps this register
Without both rows you cannot say whether scanning more widely improves anything, because you cannot see what it costs in false alarms. “Scan widely” and “red-team the future” are stated here as unambiguous goods with no price. They may well be good. As written, they are not falsifiable — and the next figure shows why that matters.
Ratio of the two rows: unknown
09 / Three that don’t hold
The hardest assumptions to catch are the ones that do not feel like assumptions. Three are load-bearing in a great deal of current thinking about technology and power, and none of the three survives being stated plainly and then examined. Each fails a different way, which is the useful part.
The case for it rests on compounding: get there first, form an ecosystem, and nobody catches you. But primacy and advantage are different quantities, and history keeps separating them — the inventor is regularly not the beneficiary. If first does not mean dominant, then a strategy organised around being first is optimising the wrong variable at enormous cost.
Fails on the gap between inventing and benefiting
Diffusion and deployment may matter more than invention. The country that invents a technology need not be the one that manufactures it, scales it, or sells it to everyone else — and adoption capacity is built from a whole ecosystem of research, manufacturing, talent, capital, regulation and public trust in combination, not from any single breakthrough. Which means the thing we count is not the thing that decides.
Fails on where advantage actually accrues
This one cannot be settled because it was never specified. Whether a tool is cheating depends entirely on the activity, the goal, and what the people involved understood themselves to be agreeing to — and those change case by case. An assumption with no definition cannot be tested, which makes it the most dangerous of the three: it will be argued indefinitely and decided by whoever is loudest.
Fails on terms nobody has fixed
10 / Mitigation
Eliminating surprise is impossible. Academic papers, patents, and commercial activity already provide ample early warning; the challenge is sorting the signal that matters. Each foresight practice widens what you can see — and each widening pulls in more noise than signal.
Lessons-learned exercises — after-action reports, post-mortems, the military’s “hot wash” — let organizations see what has and has not worked. The name matters less than the effect. The Intelligence Community used them after September 11th and after the search for Iraqi WMD to revamp training, build better methodologies, and increase analytic rigor.
Experts want to demonstrate expertise; preparing for shocks requires the opposite. Known unknowns are possibilities we fail to keep scanning for. Unknown knowns are things we know and have chosen to forget. And the hardest assumptions to catch are the most deeply embedded, because they do not feel like assumptions — surfacing them means asking not what we know, but what we have stopped asking about.
Scenario exercises let planners engage with surprises before they arrive: identifying bureaucratic weaknesses, establishing protocols, acquiring capabilities. Eisenhower’s formulation still holds — plans are worthless, but planning is everything. Pre-familiarization with resources, constraints, and decision points is what accelerates replanning later.
Red teaming expands the aperture to how innovations diffuse through markets and industry ecosystems. It asks scientists and engineers to work under the competitor’s constraints and incentives, and to prototype how others might adapt a technology. It fails when participants lack real knowledge of the other side’s mindset and simply play them as themselves.
Systematic monitoring across publications, patents, and commercial activity does surface more real signal, and AI can run it at a speed humans cannot. But the field being scanned is mostly noise, so opening the aperture adds false positives faster than true ones. That is the cost the recommendation does not price, and the reason accountability has to reach the decision not to act — not only the warning.
Widen it and you catch more of what matters. You also drag in far more of what doesn’t, because the field is mostly noise and the middle is where you were already looking. There is no setting that gives you everything, which is the part “scan widely” leaves out.
11 / The dependence
Every chapter so far has treated the technology as the thing that surprises you. There is a harder case. The more a society comes to depend on something that works, the more its failure costs and the more attractive a target it becomes — and that exposure accumulates without any event to mark it. Nothing happens on the day you become dependent.
The first tranche of adoption delivers most of the value. Going from no satellite timing to some satellite timing changes everything; going from heavy reliance to total reliance adds very little. This is the ordinary shape of a good technology, and it is the curve every business case and every procurement decision is written against.
Diminishing returns — the familiar half
The second curve is the one nobody plots. As dependence deepens, the cost of the thing failing does not level off — it accelerates, because the alternatives have been dismantled, the skills have not been kept up, and everything downstream now assumes the service is there. And the same concentration that makes failure expensive makes it worth attacking.
Accelerating, convex — and unfunded
Somewhere the second curve rises above the first, and past that point additional dependence is buying less than it risks. There is no event on that day. No announcement, no test, no headline — which is exactly why it qualifies as a strategic surprise, and why no amount of scanning for new technologies will find it. GPS and the internet are the standing examples. The surprise was never the invention.
GPS · the internet · and no event to mark either
12 / Resilience
Forecasting is only half of foresight. The other half is concrete steps taken beforehand, so that when a surprise lands the recovery is fast. The shock does not get smaller. The window of vulnerability does.
Responding to rapidly evolving developments requires a diverse, expert team already in place. Surprises present opportunities as well as threats, and taking advantage of an opportunity is usually time sensitive. A trusted team has to be ready to move quickly to weigh risks and advantages as the situation evolves.
Intelligence organizations have long kept contrarian units to argue the opposing analytic position; the Catholic Church maintained an advocatus diaboli for centuries. Decision makers need an institutional contrarian voice to force consideration of low-probability events — at minimum, to force contingency plans into existence.
Response plans that can be deployed quickly help manage a surprise when it lands. They need to be domain-specific. Several Nordic countries have published crisis handbooks telling citizens exactly what to do. For quantum decryption — widely treated as a matter of when, not if — agencies could run the pre-mortem now.
Developing governance frameworks, regulatory approaches, and response protocols before a technology reaches maturity shortens the window of vulnerability when the surprise arrives. That requires sustained engagement with scientific communities and iterative policy development that anticipates multiple technological trajectories at once.
If a surprise triggers a crisis, people will want information immediately and will struggle to know whom to believe. Pre-established channels, trusted messengers, and clear procedures for dissemination reduce confusion, prevent panic, and make coordination possible while the situation is still moving.
After a surprise we audit the warning. We ask who saw it coming, when they said so, and who ignored them. We almost never audit the other half — and the other half is the half that did not vary. Warning and readiness are separate institutions, funded separately, and only one of them is ever held to account.
So audit the second one, before the next disruption rather than after it. Not what do we predict, but: what have we actually rehearsed, who is authorised to move without asking, how long does it take us to reallocate, and what have we written down that a stranger could execute. Publish the answers. The forecast will be wrong in ways nobody can fix. The window is the part you control.
A provisional framework
