The partial adoption trap: Coordination failure, trust, and
cultural lock-in in health AI adoption · 2026
Read The partial adoption trap: Coordination failure, trust, and by cultural lock-in in health AI adoption (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-17388.txt.
Introduction
Ercole models health-AI uptake as an evolutionary game in which doctors choose genuine use, partial use or rejection, with systemic gains needing a population threshold. The system is generically bistable: a partial-use trap coexists with full uptake, widened by threshold coordination failure, trust failure and cultural lock-in among doctors. The trap bites hardest for the most valuable tools, and standard incentive-only policy, he predicts, systematically reproduces it.
Catalog record
| CREATOR | cultural lock-in in health AI adoption |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36058 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.17388
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The partial adoption trap: Coordination failure, trust, and cultural lock-in in health AI adoption Abstract Abstract Health artificial intelligence (AI) adoption presents a paradox: point-solution tools diffuse readily through clinical populations, yet system-change AI, which carries the greatest potential for pathway-level transformation, consistently stalls at partial adoption. An evolutionary game theoretic model is developed to explain this pattern. Doctors choose among three strategies: genuine adoption, partial adoption, and rejection, where genuine adoption is required for systemic benefits to materialise above a population threshold. The system is shown to be generically bistable, with a stable partial adoption equilibrium coexisting alongside full genuine adoption. The basin of attraction of the partial adoption trap is enlarged by three compounding failure modes: a threshold coordination failure arising from the non-appropriable nature of systemic benefits; a trust failure arising from the organisation’s inability to credibly commit to sharing productivity gains; and a cultural failure arising from negative coordination norms among doctors. These failure modes are shown to be most severe precisely for the technologies with the greatest systemic value: the Value-Adoption Paradox. A cost ratchet dynamic implies that failed adoption attempts permanently lower barriers even when embedding fails, but this benefit is offset when trust erosion is rapid. Conditions are derived under which sustained but imperfect adoption pressure is welfare-improving, and the policy architecture required to escape the trap (targeting trust, sequencing, and team-level adoption) is characterised. Standard health system digital transformation policy, which typically addresses only the threshold failure through individual incentives, is predicted to systematically produce the partial adoption trap. Introduction The deployment of artificial intelligence in health systems has accelerated markedly over the past decade. Diagnostic support tools, administrative automation, and clinical decision aids have proliferated, and adoption rates by conventional measures appear substantial. Yet a striking divergence has emerged between adoption of point-solution AI (tools that improve performance on a specific clinical task for the adopting clinician) and system-change AI, which is intended to restructure care pathways, reduce transaction costs, and generate productivity gains at the system level. Point solutions diffuse readily. System-change AI is widely installed but rarely embedded: the tool is present, often formally adopted, but the pathway transformation it is intended to enable does not occur. Ambient voice technology (AVT), AI scribes that transcribe consultations and automate clinical documentation in real time, illustrates the problem acutely. AVT is explicitly identified as a core enabler of the analogue-to-digital shift in the NHS Fit for the Future 10-Year Health Plan [ 1 ] , which commits to national rollout across all care settings and targets the productivity gains needed to meet a 2% annual efficiency improvement. The individual benefit is real and well-documented: documented time savings allow each doctor to do more with existing capacity [ 2 ] . Yet widespread partial adoption of AVT, using the tool to ease documentation without restructuring clinic flow or throughput, will not deliver the capacity gains the Plan anticipates. That requires enough of a clinic population to genuinely restructure appointment patterns that the booking system, administrative staff, and patient pathways all change coherently: a collective threshold that no individual doctor can cross alone. AVT is used throughout this paper as a running example because its structure maps cleanly onto the general model, but the argument applies wherever system-change AI requires collective workflow restructuring to generate its intended benefit. This divergence between individual and collective adoption incentives is not explained by standard technology adoption models. Diffusion-of-innovations theory [ 12 ] treats adoption as binary and individual, driven by perceived usefulness and ease of use. Technology acceptance models [ 7 ] similarly focus on individual-level attitudes. Neither framework captures the collective action structure that characterises system-change AI, where the benefits to any individual clinician depend on the adoption decisions of colleagues, and where systemic benefits only materialise when adoption exceeds a population threshold. The economics literature offers a closer analogue. Canton et al. [ 6 ] develop a model in which workers resist technology adoption when costs fall disproportionately on those with least time to recoup benefits, and where the share of productivity gains accruing to workers depends on market competition. Their overlapping-generations framework with majority voting captures the distributional politics of adoption resistance. However, that model does not capture three features distinctive to clinical AI adoption in a public health system: the threshold structure of systemic benefits, the trust problem arising from the organisation’s inability to credibly commit to sharing gains, and the social dynamics by which partial adoption norms become self-reinforcing within clinical teams. This paper contributes to the literature by developing an evolutionary game theoretic (EGT) model of clinical AI adoption that nests all three failure modes in a single framework and characterises their interactions. EGT is appropriate here for two reasons. First, adoption decisions in large clinical populations are made by boundedly rational agents who observe and imitate peers rather than solving complex optimisation problems, a pattern well-documented in the clinical practice variation literature [ 16 ] and consistent with the imitation dynamics that underpin replicator equations. Second, EGT naturally accommodates a population-level strategy distribution as the state variable, allowing characterisation not just of equilibria but of basins of attraction and the dynamics of convergence, which are essential for policy design. The model has three strategies. Genuine adoption (G): the doctor integrates the AI system in a way that restructures fundamental working patterns, contributing to the population threshold required for systemic benefit. Partial adoption (P): the doctor uses the tool for tasks where it generates immediate private benefit (time savings, decision support) but does not restructure working patterns, capturing private gains without contributing to systemic transformation. Rejection (R): no adoption. This strategy space is richer than existing models and captures a distinction that is both empirically important and analytically consequential: partial adoption is not non-adoption, but it is not the adoption that generates systemic value. The main results are as follows. First, under plausible parameter conditions the replicator dynamics on the strategy simplex are bistable, with two stable equilibria: full genuine adoption and the partial adoption trap in which the entire population partially adopts. The partial adoption trap is the generic attractor for most initial conditions. Second, the cost of genuine adoption decays over time conditional on the system embedding (a cost ratchet)which creates path dependence: failed adoption attempts that push effective adoption above the systemic threshold, even temporarily, permanently lower future barriers. Third, when the organisation’s sharing of productivity gains is modelled as a trust game, rational anticipation of reneging endogenises the appropriability parameter and generates a self-confirming trust trap that compounds the coordination failure. Fourth, negative coordination norms among doctors create cultural lock-in that deepens the partial adoption trap independently of individual incentives. Fifth, these failure modes compound each other most severely for technologies with the greatest systemic value, the Value-Adoption Paradox, implying that standard adoption policy fails most where it matters most. The policy implications are concrete. Escaping the partial adoption trap requires addressing all four failure modes in the correct sequence. Trust architecture must precede adoption requirements. Cultural preparation must precede individual incentives. Threshold subsidies must be concentrated and targeted at seeding rather than universal adoption. Implementation support must be concentrated in the embedding window when cost