In 1985, the economic historian Paul David published Clio and the Economics of QWERTY in the American Economic Review. The keyboard layout chosen in the 1870s — partly to slow typists down — had locked in despite alternatives like the 1936 Dvorak layout. The case became the canonical example of path dependence: small early events with large permanent effects. Four years later, Brian Arthur at the Santa Fe Institute formalised it: under increasing returns to adoption, competition between technologies resolves through small early random events, and the winner need not be the best. Stan Liebowitz and Stephen Margolis later challenged the QWERTY-as-clean-lock-in narrative; the framework survived in modified form.
Path dependence is the claim that a system's current state depends on the historical sequence by which it got there, not just on current incentives. It comes in three nested strengths: weak (history matters but optimal outcomes still reachable), semi-strong (history makes the system predictably fail to reach optima but the gap is small), and strong, often called lock-in (history produces significantly suboptimal outcomes that are costly to escape). Strong-form lock-in is the most contested. The mechanisms are well-understood: increasing returns to scale drop per-unit cost as adoption rises; network effects raise the value of a technology with its user base; switching costs make replacement expensive; coordination requirements turn switching into the problem of moving many parties simultaneously when a technology depends on complementary investments — drivers and roads, applications and operating systems, fuel and vehicles; learning effects further reduce the incumbent's cost over time. The classic candidates have aged unevenly. QWERTY vs. Dvorak — the original example — looks weaker than its 1985 reputation: the empirical evidence for Dvorak's significant superiority is thin. VHS vs. Betamax holds up better — Sony's Betamax was technically superior but lost to JVC's VHS on longer recording times, lower licensing costs, and earlier rental-market penetration. Microsoft Windows's dominance over Mac, OS/2, and Linux is a clean network-effects case. The x86 architecture's long dominance over RISC is being broken in real time by Apple Silicon, showing that lock-in can be undone with sufficient resources; the combustion-engine-to-EV transition is the contemporary instance. The quantitative significance of path-dependent inefficiency often dissolves under careful scrutiny.
The global energy system's lock-in to fossil fuels — infrastructure, investment, supply chains, expertise — is the canonical negative path-dependence case of the present century, with the cost of breaking it enormous and the cost of not breaking it enormously larger. The American healthcare system's employer-provided-insurance structure was locked in by World War II wage controls; decades of failed reform reflect the strength of the lock-in. JavaScript — designed in ten days in 1995 — became the only language that runs natively in browsers; its ten-day design has shaped trillions of dollars of subsequent technology. The transformer architecture and scaling-laws-driven AI is arguably an early-stage path dependence whose lock-in status is still being decided. Path dependence is diagnostic, not defeatist.