The Collingridge dilemma is a problem of timing. David Collingridge, who worked at the Technology Policy Unit of the University of Aston in Birmingham, set it out in his 1980 book The Social Control of Technology. His point was blunt. When a technology is young, it’s cheap to steer, but nobody knows enough to steer it well. By the time its effects are plain, it has wound itself into jobs, budgets, laws and daily habits, and steering it gets slow and costly.
Collingridge compressed it into one line that still gets quoted: when change is easy, the need for it cannot be foreseen. The second half writes itself. Once the need is obvious, change has become expensive.
Later scholars split the dilemma in two. There’s an information problem: you can’t judge a technology’s impact until it has been built out and widely used. Then there’s a power problem. Once a technology is entrenched, the firms, workers and users who depend on it have every reason to resist change, and they usually have the money and the votes to do it. Collingridge’s own phrasing was that early control lacks knowledge and late control lacks leverage. Either way the regulator loses something.
The idea became a fixture of technology assessment, and later of the “responsible innovation” literature that grew up around nanotech, synthetic biology and gene editing. It also turns up in arguments over crypto, facial recognition and driverless cars. For anyone who reads markets, it’s a useful lens on regulatory risk. It says a lot about when that risk tends to show up, and who ends up paying for it.
Why Waiting Feels Safe and Usually Isn’t
Social media is the textbook modern case. In the early 2000s the platforms looked harmless. They were ways to keep up with friends, share photos and find people with odd hobbies. US law had already given online services broad immunity for what users posted, through Section 230 of the Communications Decency Act in 1996, and few people saw a reason to revisit it. The business model that grew on top of that choice was free service paid for by targeted ads, tuned by engagement-ranking algorithms.
Twenty years on, the worries are familiar: misinformation, compulsive use, political polarisation, teenage mental health, surveillance advertising. Governments did eventually move. The EU passed the Digital Services Act, Britain passed the Online Safety Act, and Australia went furthest with a ban on social media accounts for under-16s that took effect in December 2025. Look at what those laws actually do, though. They add duties, audits, age checks and fines around a business model that’s already built. None of them redesign it. That’s the power problem at work. By the time lawmakers knew what they were dealing with, billions of users, a huge ad industry and much of political communication ran through these platforms.
Leaded petrol is an older and harsher example. Tetraethyl lead went into fuel in the 1920s to stop engine knock. Workers at the plants making it were dying within a couple of years, so the hazard wasn’t a secret. Yet the additive spread worldwide because it was cheap and engines were built around it. The US only finished its phase-out in 1996. The last country still selling leaded petrol for cars, Algeria, stopped in 2021, almost a century after the first warnings. Refineries, carmakers and fuel distributors had all organised around the stuff, and each step of removal meant someone’s capital write-off.
The pattern is the same in finance, urban planning and energy. Heading off a problem early is structurally easy but politically hard, because the danger still feels abstract. After the damage is obvious, reform runs into the people whose income depends on the status quo.
When Entrenched Technology Still Gets Changed
The dilemma isn’t a law of nature, and the exceptions are instructive. Chlorofluorocarbons were everywhere by the 1980s: fridges, air conditioners, aerosol cans, foam, electronics cleaning. Even so, the Montreal Protocol of 1987 led to a near-total global phase-out within roughly a decade. Two things made that possible. Substitutes existed or were close, and the few big chemical producers that made CFCs could sell the replacements. Once the incumbents had a way to profit from the switch, their resistance faded.
Leaded petrol shows the opposite case. Removing lead meant new engines, catalytic converters and refinery upgrades, and the costs landed on firms that had nothing to sell in return. That’s why it dragged on for decades.
So the power problem has a weak spot. Entrenchment is beatable when there’s a ready substitute and the incumbents can own it. It’s close to unbeatable when change means pure loss for the people who’d have to carry it out.
Collingridge’s Answer Was Flexibility
Collingridge didn’t think the answer was better forecasting. He doubted anyone could predict social effects well enough, and he was right about that. His answer was to make decisions easy to correct. He called the quality corrigibility. In practice it means favouring options that can be reversed, keeping several alternatives alive, preferring small units with short lead times over giant ones that take a decade to build, and watching closely for early signs of trouble.
He pushed the argument further in a 1992 book, The Management of Scale, which looked at big, inflexible projects such as large nuclear plants and big state-planned programmes. The trouble with these projects, in his view, was less that their planners were foolish. They locked in so much capital and time that mistakes couldn’t be fixed once they showed up.
Flexibility isn’t free, which is the catch. Keeping options open costs money, the same way insurance does. And the early days of a technology, when flexibility is cheapest to build in, are exactly when it’s hardest to say how much it’s worth.
Modern policy borrows heavily from this thinking, usually without crediting him. Regulatory sandboxes, which Britain’s Financial Conduct Authority pioneered for fintech in 2016, let firms test products under supervision before full rules apply. Sunset clauses force lawmakers to revisit a rule after a set period. Staged rollouts, mandatory incident reporting, interoperability and data portability rules all aim at the same target. They try to keep a system changeable after it’s grown large. None of them solves the dilemma. They soften it.
There’s a camp that reads the dilemma the other way. Advocates of “permissionless innovation” argue that early control mostly kills useful things before they prove themselves, so it’s better to let a technology run and fix harms case by case. Close cousins of the idea include the “pacing problem,” the observation that technology moves faster than law can follow. The precautionary principle sits at the opposite pole, and much of the policy debate amounts to choosing a spot between the two.
AI Is the Test Case Running Now
Artificial intelligence is where the dilemma is being argued in real time. Regulate too early and you write rules for systems that will look nothing like the ones in use three years later. Wait too long and AI ends up embedded in lending, hiring, medicine, defence and education before any safeguards exist, with vendors and buyers who’ll fight any rule that forces a redesign.
The EU chose to move early. Its AI Act was adopted in 2024 with obligations phased in from 2025, and the heaviest duties, those covering high-risk systems used in areas like hiring, credit and education, were due to apply from August 2, 2026. They didn’t. A follow-up package known as the Digital Omnibus entered into force on July 27, 2026, days before the deadline, and pushed the high-risk obligations for stand-alone systems to December 2, 2027 and those for AI embedded in regulated products to August 2, 2028.
The stated reason is pure Collingridge. The technical standards needed to test compliance weren’t ready, so regulators didn’t yet know how to measure what they had written into law. That’s the information problem in legal form. Meanwhile the systems the rules were meant to cover keep spreading through European HR departments, banks and schools, building up exactly the dependency that makes later changes harder. Civil-society groups made this point loudly when the delay went through.
The US has gone the other way, relying mostly on existing agencies, state laws and voluntary commitments. That keeps options open in one sense and closes them in another. Lighter rules early mean faster adoption, and faster adoption means the power problem arrives sooner.
What the Dilemma Means for Investors
For markets, the dilemma is a rough guide to how regulatory risk behaves over a technology’s life. A few things follow from it.
Serious regulation usually lands late, after revenue is established, and it rarely takes the form of a ban. It shows up as compliance costs, disclosure duties, conduct remedies and fines. Those weigh on margins but seldom end a business. Big incumbents often absorb them better than challengers, and heavy rulebooks can end up raising the barriers to entry that protect the leaders. Investors who price a looming regulation as an existential threat to a dominant platform have mostly been wrong.
Early movers know this too. When a leading company asks to be regulated, it’s worth asking whose costs the proposed rules would raise. Licensing regimes and expensive certification can turn an early lead into a lasting moat.
The CFC lesson matters for sector calls. Watch for substitutes. Regulators act fastest and hardest when a replacement is ready and someone stands to profit from selling it. That’s when an entrenched product can be phased out quickly, and when the owners of the alternative are worth a look.
Finally, Collingridge’s case for flexibility has a balance-sheet version. Assets that are modular, short-lived and easy to repurpose carry less policy risk than giant, single-purpose ones with forty-year payback periods. Stranded assets in coal and other fossil infrastructure are the dilemma playing out on an income statement. Firms built for a world that regulation later decided to change found their capital couldn’t be moved.
There’s a human side to all of this, too. Societies tend to call technological change inevitable once momentum builds. Collingridge’s work helps explain why. Change isn’t impossible. Dependency piles up quietly, as people retrain, companies build business models and institutions reorganise themselves around a technology long before anyone fully understands what it does.
By the time the bill arrives, everyone’s already invested.
