Collingridge Dilemma

Prime #

710

Origin domain

Ethics Of Technology And Ai Governance

Subdomain

technology assessment and anticipatory governance → Ethics Of Technology And Ai Governance

Aliases

Pacing Problem

Core Idea

The Collingridge dilemma is a temporal asymmetry between two rising curves: the information about a system's consequences and the cost of changing it both rise with time and lock-in, so by the time enough is known to intervene wisely, intervention has become expensive or impossible.

How would you explain it like I'm…

The Drying Cement

Imagine wet cement you're about to step in. Early on it's soft and easy to change, but you don't know yet where your footprints should go. Later you know exactly where they should go, but the cement has hardened and you can't move them. By the time you're sure, it's too late to fix.

Too Late to Fix

When people build something new, two things change over time but pull against each other. First, you slowly learn what the new thing really does and whether it causes problems. Second, the new thing gets harder and more expensive to change as more people rely on it. The trap is that both move the same way: by the time you know enough to fix it well, fixing it has become really hard or impossible. Acting early is cheap but blind; acting late is well-aimed but ruinously costly.

Know-It vs. Change-It

The Collingridge Dilemma describes a window where two trends cross at the worst possible time. The information curve — how much you know about a system's eventual effects — starts low and rises with deployment and experience. The cost-of-change curve also starts low (few commitments made) and rises as the system locks in. The problem is they move in the same direction, so by the time you know enough to intervene wisely, intervening has become expensive or impossible. This isn't anyone's mistake — even a perfect reasoner would face it, because the information only exists after deployment. The useful object it names is the 'window of revisability': the time, possibly empty, when action is both informed enough and cheap enough.

 

The Collingridge Dilemma is the structural pattern in which a system passes through a window where two curves cross with disastrous timing. The information curve — what is known about eventual consequences — starts low and rises with time, deployment, and lived experience. The intervention-cost curve — what it costs to change the system — also starts low (few commitments made) and rises with deployment and lock-in. Because the two curves move in the same direction, by the time enough is known to intervene wisely, intervention has become expensive or impossible; early action is cheap but uninformed, informed action well-targeted but ruinously costly. The load-bearing structure is a temporal asymmetry between learning and locking-in, and it is a property of irreversible systems, not a failing of any actor. The key derived object is the window of revisability: the period, possibly empty, when intervention is both informed and feasible. Systems that learn slowly and lock in fast (large infrastructures, deployed technologies) hit it hard; those that learn fast and lock in slowly (version-controlled software, reversible policy experiments) hit it lightly. The structural response is neither waiting nor aggressive early action but bending both curves — accelerate learning through pilots and monitoring, and slow lock-in through modular, reversibility-preserving design, sunset clauses, and distributed deployment.

Broad Use

• Technology policy: nuclear power, germline editing, geoengineering — deployment-scale consequences become legible only after rollback grows prohibitive.
• Software architecture: schemas and API contracts are cheap to change before dependents exist, expensive once many systems rely on them.
• Urban planning: highway alignments and zoning are cheap before the city organizes around them, near-immovable after.
• Climate policy: emissions choices are cheap before fossil capital is sunk; atmospheric carbon persists for centuries.
• Organizational design: founding-stage choices about equity and charter are easy to set, hard to revise once scaled.
• AI governance: the dominant frame for frontier-model deployment, sharpened because capabilities scale faster than evaluation and weight release is irreversible.

Clarity

Separates information available about a system from its malleability — both functions of time, moving in conflict — and reframes "wait and see" from a free default into a priced choice that silently consumes intervention optionality.

Manages Complexity

Collapses a family of "we should have acted earlier" regrets into one geometry, and a family of design heuristics (fail fast, build in optionality, stage commitments) into one principle: bend the curves.

Abstract Reasoning

Installs a recurring analysis — map the curves, quantify the window of revisability, bend the information curve up (pilots, monitoring) and the lock-in curve down (modularity, sunset clauses) — and treats inaction as a choice with timing costs.

Knowledge Transfer

• AI governance: anticipatory-governance frameworks and staged-deployment protocols explicitly widen the revisability window.
• Public policy: software practices — feature flags, canary deployments — port to vaccine rollouts and drug approvals as revisability preservation.
• Ecology: adaptive management in fisheries is the explicit Collingridge design — monitor, learn, revise, preserve optionality.

Example

A public API contract is free to change on day one but breaks every consumer once thousands of integrations harden around it; the disciplined response is versioned APIs (bending the cost curve down) plus a private beta (bending the information curve up).

Relationships to Other Primes

Foundational — no parent edges in the catalog.

Children (1) — more specific cases that build on this

• Path Dependence is a decomposition of Collingridge Dilemma — path_dependence (the rising lock-in curve alone) is one of the two curves the Collingridge dilemma composes — it adds the co-rising INFORMATION curve and the window between them. Path dependence is broader (it appears outside the dilemma), so this is part-of, not a reparent.

Not to Be Confused With

• The Collingridge Dilemma is not Reversibility and Irreversibility because irreversibility is a static property of a commitment whereas the dilemma is the temporal race between rising information and rising lock-in.
• The Collingridge Dilemma is not Sunk Cost because sunk cost is a bias about past expenditure whereas the dilemma arises even for an unbiased reasoner, since the information is only available after deployment.
• The Collingridge Dilemma is not Optionality because optionality is the value of preserved choices and the remedy whereas the dilemma is the problem that motivates buying it.