David Collingridge stated the dilemma in a single sentence in his 1980 book The Social Control of Technology: attempts to control a technology come either too early, when its consequences cannot be predicted, or too late, when it has become too entrenched to redirect. The argument is structural, not technological, and it was built from a series of mid-twentieth-century case studies before it was ever extended to anything that came afterward.
The Original Cases
The book is organized as one principle per chapter, each grounded in a worked example. The case studies that carried most of the argument were drawn from the energy and weapons programs of the postwar West.
Lead in petrol was the canonical case of entrenchment. Tetraethyl lead was added to automotive gasoline in the 1920s to suppress engine knock. By the time epidemiological evidence linking environmental lead exposure to childhood neurological damage had become unambiguous in the 1960s and 1970s, the entire installed base of automobiles, refineries, and distribution networks had been built around leaded fuel. Removing the additive required two decades of coordinated regulatory action across automakers, oil companies, and emissions agencies. The information arrived; the cost of acting on it had compounded.
The breeder reactor illustrated lead time. Fast-breeder fission programs in the United Kingdom, France, the United States, and Japan required decades of capital commitment before the underlying economics — fuel cycle, waste handling, decommissioning — could be evaluated against operating data. By the time the evidence on cost and safety stabilized, several billion pounds and a generation of nuclear engineers were committed to a technology whose case had weakened. Cancellation was politically and industrially expensive in proportion to the depth of prior commitment.
MIRV — the multiple independently targetable reentry vehicle — illustrated competition. The decision to place multiple warheads on a single missile was taken on the United States side in the late 1960s on the assumption that Soviet defenses would advance to the point of requiring it. The Soviet Union followed. Once both arsenals had been MIRVed, the strategic-stability arguments against the move became visible, and undoing the deployment proved harder than authorizing it had been. The information problem closed only after the power problem had opened.
Electricity generation at large unit scale, the Manhattan Project, and the broader trajectory of energy R&D rounded out the catalogue. The pattern was consistent across cases: long lead times, large unit sizes, capital intensity, and dependence on specialized infrastructure produced economies of scale during deployment and diseconomies of control afterward.
The Structural Claim
What Collingridge extracted from these cases was not a complaint about any individual technology. It was a claim about the joint distribution of two variables over a technology’s lifecycle. Early in development, information about consequences is low and the flexibility to change course is high. Late in development, information is high and flexibility is low. The two curves cross somewhere in the middle, and that crossing point — wherever it falls for a given technology — is the only window in which intelligent intervention is feasible. The dilemma is that the window is narrow, hard to identify in real time, and easily missed in either direction. Premature regulation looks foolish in retrospect; delayed regulation looks negligent.
Collingridge’s own answer was not to escape the dilemma but to design around it. His prescription, set out in the chapter on “Intelligent Trial and Error,” was to favor technologies that could be deployed reversibly, to keep decision-making decentralized and close to the locus of consequences, to design for modular correction, and to scale incrementally rather than in single large commitments. The argument was not that prediction could be improved enough to close the information gap, but that the cost of being wrong could be lowered enough to make iteration affordable.
The Contemporary Instance: AI
Frontier artificial intelligence is the cleanest contemporary case of the structure Collingridge described. Capability is discovered after training rather than before it. The impacts of a model on labor markets, information ecosystems, security postures, and scientific research can be characterized only by deploying it and observing what happens. This is the information half of the bind in its purest twenty-first-century form.
The power half has arrived faster than most analysts expected. Nvidia’s market capitalization, TSMC’s leading-edge fab buildout, hyperscaler capex from Microsoft, Google, Amazon, and Meta, and the proliferation of open-weight releases from Meta, Mistral, Alibaba, and DeepSeek have together produced a constituency, a supply chain, and an installed base in roughly five years. None of this constitutes lock-in on the scale of leaded gasoline or the breeder reactor, but the trajectory is unmistakable, and each additional quarter of capital commitment narrows the window for intervention.
Three features make AI a harder case than Collingridge’s originals. The development cycle is shorter than any legislative cycle, so the information gap does not close before the power gap opens. The technology is diffusive in a way nuclear was not — model weights move on hard drives, not in fissile-material logistics — so unilateral regulatory action is structurally weaker. And the dual-use surface is total: the same model that drafts a contract drafts a phishing email, with no clean separation between regulated and unregulated applications.
Intelligent Trial and Error, Modernized
The instruments emerging around frontier AI are recognizably Collingridgean. Pre-deployment evaluations, red-teaming, model cards, responsible scaling policies with explicit capability thresholds, staged deployment with post-release monitoring, voluntary commitments at the Bletchley, Seoul, and Paris summits, and the slow construction of a regulatory science of AI evaluations all concede that the information problem cannot be solved in advance and substitute iteration and reversibility for foresight. Whether this is enough depends on a question the framework itself cannot answer: how much lock-in has already occurred, and whether the iterative regime is governing a technology that has already passed the point where iteration can change its trajectory.
The standard reading of the Collingridge dilemma treats it as a paradox to be lamented. In the cases that produced it — leaded gasoline, the breeder reactor, MIRV — it was not a paradox; it was a working condition that decisions had to be made inside. The same is true now. The substantive question is not whether to escape the bind but which side of it to err on, and the honest answer is that nobody yet knows.