Technological Lock-In

• Technological lock-in refers to entrenched technology use due to historical decisions, escalating commitments, and network effects, despite the availability of better options and manifestations in areas like software, supply chains, institutions, and economies.
• Increasing returns, network effects, and path dependency are key drivers of technological lock-in, as exemplified by models like Arthur’s, which show how early adoptions can amplify over time to create self-reinforcing dominant standards.
• Reversing technological lock-in can be challenging but effective intervention points include software architecture, modular design, staged approval, network planning, and preserving fallback capabilities, also factories can enhance flexibility and avoid over-commitment.

Technological lock-in is the persistence of a technology, technical standard, infrastructure, supplier, organizational practice, or sociotechnical trajectory because historical adoption generates feedback, switching costs, institutional commitments, or capability erosion that make alternatives progressively less viable. Lock-in may arise from increasing returns and network externalities, escalating organizational commitment, immutable software or smart-contract architectures, inherited supply-chain dependencies, declining human skills, or the redistribution of fixed infrastructure costs to a shrinking user base. It is therefore not reducible to technical superiority: an inferior or inefficient option may remain dominant, while an improved alternative may fail to diffuse. Lock-in can be technological, institutional, behavioural, infrastructural, economic, political, or epistemic, and these dimensions frequently reinforce one another.

1. Conceptual foundations and forms

Technological lock-in is commonly associated with path dependence: earlier events, investments, agreements, standards, and skills constrain subsequent choices. Four properties characterize this process: non-ergodicity, whereby contingent historical events lead systems toward different outcomes; non-predictability, whereby the eventual outcome could not necessarily have been predicted in advance; inflexibility, whereby later actors have limited ability to redirect development; and path inefficiency, whereby the selected outcome is inferior to an available alternative.

Lock-in should be distinguished from early commitment and escalating commitment. Early commitment can reduce delay and provide direction. Escalating commitment occurs when resources continue to be allocated despite adverse information or poor performance. Lock-in is the negative condition in which escalating commitment produces over-commitment to an inefficient technology, project, supplier, or specification and closes realistic alternatives (Cantarelli et al., 2013).

Several related forms are analytically distinguishable:

• Network lock-in: adoption increases a technology’s attractiveness through installed-base effects, compatibility, or coordination benefits.
• Institutional and behavioural lock-in: decision-makers defend prior choices through sunk costs, self-justification, political commitments, and inflexibility.
• Vendor lock-in: technical, contractual, organizational, or governance costs make replacing a supplier impractical.
• Soft lock-in: alternatives remain legally or technically permissible but are expensive, uncertain, or organizationally unattractive. Open-source licensing does not prevent soft lock-in when documentation, testing, maintainership, procurement, and expertise remain concentrated in one supplier (Persson et al., 2024).
• Infrastructural and economic lock-in: durable assets and fixed costs create pressure to continue operating a system and recover its capital expenditure.
• Cognitive or epistemic lock-in: reliance on external systems erodes independent capabilities or concentrates beliefs and viewpoints.
• Supply-chain lock-in: upstream platform decisions, such as an SoC vendor’s kernel baseline, are inherited by downstream manufacturers and users.

Lock-in is not synonymous with dominance, standardization, or popularity. A dominant technology may provide coordination benefits, and early standardization can sometimes be socially desirable. Conversely, a persistent majority does not necessarily constitute lock-in unless the inferior option has a positive probability of retaining a stable long-run majority. The relevant issue is whether alternatives have become difficult to evaluate, adopt, maintain, or recover.

2. Increasing returns, network effects, and evolutionary trajectories

Arthur’s model of competing technologies formalizes lock-in through two technologies, $A$ and $B$, and two consumer types, $R$ and $S$. $R$-agents naturally prefer $A$, while $S$-agents naturally prefer $B$. Their returns include intrinsic preferences and network effects:

$\begin{array}{c|cc} & A & B\ \hline R & a_R+r n_A & b_R+r n_B\ S & a_S+s n_A & b_S+s n_B \end{array}$

Here, $n_A$ and $B$0 are previous adoption counts, while $B$1 and $B$2 measure network-effect strength. With $B$3 and $B$4, each consumer type has a natural preference, but adoption raises future attractiveness. The resulting feedback is path-dependent: random early differences can be amplified by subsequent choices.

For an $B$5-agent, $B$6 becomes preferable when

$B$7

which is equivalent to

$B$8

Once the installed-base advantage exceeds this threshold, even consumers intrinsically inclined toward $B$9 may adopt $R$0. The process resembles a Pólya urn because previous choices alter the attractions governing later choices. The resulting dominant state is absorbing in practical terms: reversal becomes increasingly difficult as the installed base grows.

Leydesdorff and Van den Besselaar interpret lock-in, lock-out, substitution, and coexistence as alternative evolutionary trajectories rather than movement toward a single globally optimal equilibrium. Their baseline parameters,

$R$1

produced lock-in in all ten reported runs with 10,000 adopters, although lock-in did not always occur within the simulation horizon. Reducing the network effects to $R$2 often prevented lock-in. Consumer uncertainty about market-share differences below 5 percent prevented lock-in in nine of ten cases, while a 5 percent switching threshold for reflexive consumers similarly suppressed lock-in (Leydesdorff et al., 2010).

The model distinguishes changes in intrinsic preference from changes in network parameters. A dramatic improvement in $R$3’s intrinsic attractiveness did not generally dislodge $R$4 after $R$5 had captured approximately two-thirds of the market. By contrast, reducing the incumbent’s technology-specific network parameter from $R$6 to $R$7, or increasing the subordinate technology’s network effect, could induce ordered substitution. This distinction suggests that interventions affecting infrastructure, compatibility, standards, or installed-base effects may be more consequential than improvements in intrinsic technical performance alone.

A related behavioural mechanism is the marginal majority effect: a discontinuous increase in choice probability when an option becomes marginally more popular than its competitor. If $R$8 is the probability of choosing an inferior option with current share $R$9, the effect at the majority boundary is

$S$0

The process is lock-in-prone if there exists $S$1 such that $S$2. A sufficient condition is that the marginal-majority effect exceed the intrinsic-quality difference:

$S$3

Across the experiments analyzed in the paper, lock-in occurred in 27 of 28 cases when $S$4, compared with 5 of 29 cases when $S$5. These experiments concerned sequential binary choices rather than technological markets, so their technological implications are extrapolative. They nevertheless identify a behavioural complement to conventional increasing-returns models: a small popularity advantage can become self-reinforcing when users interpret “currently ahead” as a categorical signal of safety, compatibility, legitimacy, or future adoption (Gelastopoulos et al., 2024).

3. Commitment, institutions, and project-level entrenchment

Lock-in also emerges through organizational decision-making. In large infrastructure projects, it may occur before formal approval, when the project becomes practically irreversible, and after approval, when actors become committed to inefficient designs or implementation arrangements.

The distinction between the real decision to build and the formal decision to build is central. The real decision is the earlier point at which abandoning or fundamentally reconsidering the project ceases to be realistic. The formal decision is the official authorization by Parliament or another authority. In the Betuweroute and HSL-South cases, formal approval occurred in 1996, while the start of the PKB planning procedure in 1992 was identified as the effective point of no return.

Five principal lock-in indicators are identified:

1. Sunk costs in time, money, studies, land acquisition, design, and organizational effort.
2. Need for justification, including self-justification, cognitive dissonance, face-saving, reputational concerns, and political pressure.
3. Escalating commitment, involving continued allocation of resources to justify earlier decisions.
4. Inflexibility, or refusal to revise routes, technologies, financing, or specifications in response to new information.
5. Closure of alternatives, whether through explicit exclusion, solution-driven problem definition, procurement requirements, coalition agreements, or late consideration.

Lock-in affects cost overruns through both methodology and practice. If the cost estimate at the real decision is $S$6, while the estimate at the formal decision is $S$7, then

$S$8

The later formal baseline may already incorporate cost increases incurred after the project became effectively irreversible, thereby understating the increase relative to the actual commitment point.

For the Betuweroute, final 2007 costs of €4.663 billion corresponded to a 64.6% overrun relative to the 1992 real-decision baseline, but only 12.7% relative to the 1996 formal-decision baseline. For HSL-South, final costs of €7.17 billion corresponded to a 403.67% overrun relative to the earlier real-decision baseline and 110% relative to the formal 1996 baseline (Cantarelli et al., 2013).

The cases also show that lock-in is neither uniform nor total. The Betuweroute retained flexibility concerning the Pannerdensch Kanaal crossing and tunnel-boring method, while the HSL-South project replaced an initial national plan and added Antwerp as a stop. Lock-in can therefore coexist with flexibility in specific sub-decisions while constraining the overall trajectory.

4. Lock-in in infrastructures, software, and supply chains

Durable infrastructure creates lock-in by combining physical persistence with financial recovery rules. In natural-gas heating transitions, household electrification reduces gas demand while depreciation, return on utility plant, maintenance, administrative costs, and portions of taxes remain. These fixed and semi-fixed costs are then allocated across fewer gas customers. The resulting mechanism can produce a utility “death spiral”:

$S$9

Natural-gas pipelines commonly have modeled service lives exceeding 50 years; the study uses a 60-year depreciation life. In Massachusetts, electrification alone was projected to raise average per-customer gas costs by approximately 43% by 2039 and approximately 180% by 2050. Combined electrification and pipeline replacement produced increases of approximately 60% over the next 15 years and as much as 225% by 2050. Households that did not electrify were projected to experience an average energy-bill increase of approximately 46% over the next 15 years (Garibay-Rodriguez et al., 4 Mar 2025).

The distribution of electrification matters. If higher-income households electrify first, lower-income households remain concentrated among gas customers and face larger burdens. If lower-income households electrify first, the distributional increase is smaller, although the underlying fixed-cost problem remains. Neighborhood-scale electrification and coordinated gas-network retirement can reduce the persistence of fixed costs more effectively than scattered household departures.

Software systems exhibit comparable dynamics even when source code is open. In the municipal e-service platform studied by Persson and Linåker, more than 190 municipalities used an AGPL 3.0 platform supplied primarily through one vendor. The platform was formally open, but inadequate documentation, unclear dependency management, limited test coverage, ambiguous maintainership, restrictive qualification requirements, fragmented communication, and municipal comfort with the status quo made alternative suppliers difficult to engage. This is soft lock-in: formal rights exist, but effective substitutability is limited (Persson et al., 2024).

Blockchain oracle selection adds immutability and governance to the lock-in structure. Protocols may hard-code oracle addresses and data logic into smart contracts that cannot be modified after deployment. Replacing a provider may require contract changes, re-auditing, governance approval, TVL withdrawal, parameter retuning, or alterations to consensus and block-building. In the reported survey, 93.75% of respondents would not choose another oracle merely because it offered the same service at a lower price. Protocols overwhelmingly preferred outsourcing when viable third-party solutions existed, while internalization was generally driven by specialized requirements such as TWAP, VWAP, illiquid-asset feeds, or custom fallback logic (Caldarelli, 29 Nov 2025).

SOHO devices illustrate supply-chain lock-in. SoC vendors distribute SDKs containing customized Linux kernels, drivers, configurations, and build systems. ODMs and OEMs generally integrate these SDKs rather than migrating to newer upstream kernels. In the analyzed corpus, 306 devices were linked to 102 SoCs, and 94.7% of identified SoCs were supplied by five major vendors. Among 174 cases with both source-derived and binary kernel versions, only 12, or 6.8%, showed a shipped kernel newer than the SDK-linked source kernel. The inherited kernel creates vulnerability debt: known defects remain in a baseline that downstream actors lack the resources, documentation, or proprietary drivers to replace (Badola et al., 9 Jun 2026).

These cases show that lock-in is often located not in a single product but in complementary assets: documentation, interfaces, drivers, standards, audit records, skills, contracts, governance procedures, and supplier relationships. The practical exit option is therefore smaller than the formally available option set.

5. Learning dynamics, catastrophe, and artificial-intelligence dependence

Population games provide a formal model of how a socially costly incumbent can be displaced without directly subsidizing the alternative. In the model, $R$0 is a wasteful incumbent and $R$1 is a superior alternative. With normalized payoffs,

$R$2

where $R$3 is the fraction using $R$4 and $R$5 is its private cost. The pure equilibria $R$6 and $R$7 are both evolutionarily stable, while the unstable threshold

$R$8

separates their basins of attraction.

Boltzmann Q-learning introduces an exploration parameter $R$9. In the normalized game, the dynamics are

$A$0

For $A$1, equilibria are quantal response equilibria rather than Nash equilibria. As $A$2 increases, the upper stable equilibrium and the unstable middle equilibrium collide in a saddle-node bifurcation at a critical value $A$3. Above this value, only a lower stable equilibrium remains. Reducing $A$4 after the state has crossed the original threshold sends the system toward $A$5, the efficient technology. The temporary intervention is symmetric, but its consequences are asymmetric because it changes the stability structure of a history-dependent system. The process exhibits hysteresis: the same control value can produce different outcomes depending on the system’s history (Leonardos et al., 2020).

This mechanism is structurally robust to bounded perturbations and to general network-effect exponents $A$6, although sufficiently large or unbounded disturbances can destroy the guarantee. The model is homogeneous and stylized; real interventions may affect agents unequally, and agents may respond strategically.

Artificial-intelligence dependence extends lock-in from technologies and infrastructures to human capabilities and epistemic systems. AI Lock-In is defined as a condition in which reliance on AI becomes cognitively, culturally, economically, and infrastructurally entrenched, making reversion to a pre-AI state nearly impossible. The proposed progression is:

$A$7

The mechanisms include cognitive offloading, organizational substitution of junior work, erosion of tacit-knowledge pipelines, provider concentration, cloud and API dependence, and tightly coupled software supply chains. Individual reliance can become organizational dependence, while organizational dependence can become national infrastructural dependence. “Sovereign AI” may reduce dependence on foreign providers without eliminating dependence on AI itself (Kim et al., 28 May 2026).

A related feedback-loop hypothesis concerns LLMs. Humans supply beliefs and preferences to models; models generate aggregated or amplified outputs; users update their beliefs; and the resulting beliefs become future training or preference data. The model formalizes this process through a trust matrix $A$8. In the human–LLM specialization, the critical condition is

$A$9

where $S$0 represents the AI’s effective weight on human beliefs and $S$1 represents human trust in the AI. Above the threshold, repeated circulation of beliefs causes aggregate precision to grow exponentially rather than linearly, and convergence to the truth fails in the model. Agent-based simulations produced convergence, increased confidence, and semantic diversity loss. WildChat analyses reported several model-version-associated reductions in conceptual diversity, although the observational evidence is mixed and does not establish irreversible societal belief lock-in (Qiu et al., 6 Jun 2025).

6. Detection, mitigation, and governance

Lock-in should be assessed by reconstructing how alternatives became progressively less viable. Relevant diagnostic questions include:

• When did the real decision or point of no return occur?
• Which alternatives were available and seriously evaluated at that time?
• Which costs are sunk, and which future costs remain avoidable?
• Are technical interfaces, standards, dependencies, or contracts supplier-specific?
• Can an independent actor build, test, operate, modify, and maintain the system?
• Does the system preserve human or institutional fallback capacity?
• Are market-share signals, rankings, or model outputs creating discontinuous positive feedback?
• Who bears the costs of failure, migration, stranded assets, or inherited vulnerability debt?

Mitigation generally operates on network structure, reversibility, information, and governance rather than on technical performance alone. Measures include interoperability and open standards; compatibility layers; modular architectures; staged approval and funding; independent cost and benefit reviews; scheduled reconsideration points; sunset clauses; pilot projects; flexible functional specifications; reproducible builds; explicit dependency declarations; automated testing; neutral maintainership; and procurement rules that preserve supplier contestability.

For open-source public-sector systems, an Open Source Steward can provide neutral maintainership, repository and issue-tracker management, contribution review, documentation, release coordination, pooled procurement, and an interface between municipalities and suppliers. For SOHO devices, mitigation requires supported or long-term-support kernel baselines, open drivers and documentation, software bills of materials supplemented by security-status reports, and SoC-vendor engagement with community projects. For blockchain protocols, upgradeability, provider abstraction, fallback mechanisms, multi-provider interfaces, and explicit migration procedures should be designed before deployment. For gas transitions, integrated gas-electric planning, neighborhood-scale electrification, strategic network retirement, revised depreciation, and non-pipeline alternatives can reduce fixed-cost lock-in.

AI dependence requires both AI-engaged literacy and AI-independent literacy. Organizations can preserve fallback capacity through continued junior hiring, senior review, manual procedures, alternative tools, provider diversification, and periodic “AI-free drills.” Critical infrastructures require tested operation without AI rather than merely documented contingency plans.

The effectiveness of symmetric interventions is context-dependent. In population games, equal exploration can destabilize a wasteful incumbent because the system is nonlinear. In social-choice systems, exposing quantitative rather than merely ordinal popularity information can weaken marginal-majority effects. In LLM ecosystems, provenance controls, independent data sources, multiple models, calibrated uncertainty, and diversity monitoring may reduce recursive amplification, although these interventions have not been conclusively tested.

7. Controversies, limitations, and analytical significance

Lock-in is not inherently inefficient. Coordination around a common standard can generate benefits, and a dominant technology may be socially preferable. The central question is whether dominance reflects beneficial coordination or accidental, costly, or inequitable path dependence.

The empirical status of lock-in claims varies substantially. Arthur-style simulations establish mechanisms under specified parameters but do not predict all real markets. The infrastructure studies reconstruct decision processes and project cost effects but cannot demonstrate that every overrun resulted from lock-in. The marginal-majority experiments directly study binary sequential choices, not technological markets. The natural-gas findings are projections dependent on assumed electrification rates, cost allocation, depreciation, customer-exit patterns, and utility planning. The LLM evidence is observational and vulnerable to temporal confounding. The blockchain-oracle study uses a small expert sample with private data. The SOHO study identifies source patterns and supply-chain associations, not universal exploitability or every internal vendor decision. The AI Lock-In position paper is principally conceptual and does not yet demonstrate irreversible societal dependence.

A further controversy concerns the relationship between technical openness and practical autonomy. Open source, open standards, and multi-sourcing can reduce formal barriers while leaving knowledge, governance, and operational capacity concentrated. Conversely, apparent redundancy may be complementary rather than substitutable: multiple providers can share a common dependency, and multiple technologies can remain unable to displace the incumbent.

The most general conclusion is that technological lock-in is a property of evolving sociotechnical systems. Technologies become entrenched through interactions among adoption, infrastructure, standards, knowledge, institutions, incentives, governance, and human capabilities. Reversal is most feasible before alternatives disappear, skills atrophy, fixed costs become stranded, or commitments become embedded in immutable code and organizational routines. Consequently, technological evaluation must consider not only performance at adoption but also reversibility, maintainability, contestability, distribution of transition costs, and the preservation of independent alternatives.