AI that optimizes on historical signals can make organizations more efficient but less innovative by rendering atypical ideas institutionally invisible; the paper proposes an 'institutional interface' framework and design features—such as non-optimization zones and outlier pathways—to govern this trade-off, while offering no empirical tests.

Main Finding

AI systems that greatly increase organizational efficiency can systematically erode the drivers of innovation. Yang (2026) argues this happens not primarily through technical failure but via meso-level “institutional interfaces” — configured combinations of AI design choices, human–AI interaction protocols, and organizational norms — that render outlier, hard-to-quantify, or low-confidence ideas (termed “residuals”) invisible or excluded. Under specific design conditions, this produces an unconscious, self-reinforcing efficiency lock-in that reduces cognitive diversity and long‑term innovative capacity. The paradox cannot be “solved” once and for all; it must be continuously governed through institutional design (e.g., non-optimization zones, outlier pathways, paradox‑balancing roles), each of which brings persistent dilemmas (isolation, authority, evaluation).

Key Points

• Institutional interface (new meso-level lens)

  • Focuses on how algorithmic operations + human–AI protocols + norms jointly define what counts as legitimate cognition, action, and value at decision junctures.
  • Operationalized at the practice/decision point level and nested across three interfaces: cognitive, action, and value.

• Residuals

  • Elements (ideas, behaviors, values) that deviate from statistical mainstreams or resist quantification.
  • Systematically deprioritized or excluded by AI‑driven institutional execution.

• Unconscious automatic execution (feature of AI-mediated institutionalization)

  • Decomposed into three dimensions: opacity of execution, loss of human agency, and diminished reflexivity (reduced questioning of the decision process).

• Mechanism of erosion

  • Optimization-oriented/discriminative AI + defaulting to AI outputs + absent outlier pathways → pre‑evaluative structural filtering of residuals (they never enter the decision stream).
  • This is distinct from, but compatible with, market‑selection, attention scarcity, or incentive/self‑censorship explanations; it predicts exclusion even when attention or incentives are adjusted.

• Design conditions and countermeasures

  • Erosion is most likely under certain design choices (AI outputs as defaults, no explicit outlier-handling, tight KPI optimization).
  • Institutional features to counteract erosion: non-optimization zones (deliberate spaces not subject to optimization), explicit outlier pathways, and paradox‑balancing roles/personnel.
  • Each countermeasure entails dilemmas (e.g., isolating innovation may reduce influence; who has authority to override AI; how to evaluate outliers fairly).

• Contribution

  • Provides a theoretically grounded, testable framework and propositions for when and how AI adoption reshapes the efficiency–innovation relationship.

Data & Methods

• Paper type: Hypothesis & Theory (no empirical dataset).
• Methods used:
  • Conceptual/theoretical development grounded in literature synthesis (organizational learning, paradox theory, algorithmic management), illustrative composite vignettes, and formulation of conditional, testable propositions.
  • Comparative argumentation distinguishing the institutional interface account from alternative explanations (market selection, attention limits, incentive effects).
• Empirical implications and suggested approaches (from the paper):
  • Observable predictions that distinguish pre‑evaluative structural filtering from post‑evaluative rejection (e.g., ideas never entering pipelines vs. assessed and rejected).
  • Recommended empirical methods for future work include longitudinal firm studies, field experiments on decision defaults, archival analyses of project pipelines, interviews tracing when ideas disappear from processes, and agent‑based or simulation models to study lock‑in dynamics.

Implications for AI Economics

• Firm-level innovation dynamics

  • Short-term productivity gains from AI may coincide with reduced rate of breakthrough innovation, shifting firms toward incrementalism.
  • Risk of “efficiency lock-in”: firms that optimize aggressively via AI may lose long-run competitive dynamism.

• Mis-measurement and growth accounting

  • Standard productivity measures may overstate welfare gains if innovation inputs and future growth potential are eroded and not captured in near-term outputs.
  • Economists should account for omitted long-term innovation externalities when evaluating AI investments.

• Capital allocation and R&D investment

  • Investors and managers optimizing for near-term KPIs may favor AI systems that increase efficiency but under-invest in high‑uncertainty R&D; capital markets could amplify the lock-in.

• Market structure and competition

  • Homogenizing decision rules across firms (e.g., common training data, shared optimization objectives) could reduce inter‑firm heterogeneity, potentially lowering the pace of industry‑level innovation.
  • Potential new forms of superstar/stagnation dynamics: firms that maintain outlier‑friendly institutions may gain long‑run advantage, but such firms are fragile and costly to sustain.

• Labor and organizational policy

  • Role redefinition: managerial and knowledge-worker roles may shift toward monitoring and creating institutional safeguards (outlier pathways, non-optimization zones).
  • Labor-market impacts extend beyond displacement: changes in task portfolios and incentives to propose high-uncertainty ideas.

• Regulation and governance

  • Policy interventions could focus on mandating or incentivizing institutional features that preserve innovation inputs (e.g., audit trails for filtered proposals, transparency/rights to appeal AI defaults, requirements for outlier handling).
  • Antitrust and industrial policy considerations: regulators should consider not only prices and short-run output but also institutional capacity for experimentation and innovation.

• Empirical research agenda for AI economics

  • Measure residuals: develop metrics capturing ideas/proposals filtered before evaluation, diversity of cognitive inputs, and “invisibility” events in decision pipelines.
  • Causal tests: randomized field experiments manipulating default reliance on AI, or presence/absence of outlier pathways, and tracking subsequent innovation outcomes.
  • Longitudinal studies: link firm AI adoption/design choices to long‑run patent quality, product breakthrough rates, and productivity growth.
  • Structural/agent-based models: simulate how institutional interfaces create path dependence and efficiency lock-in across industries.

Overall, Yang (2026) reframes AI’s efficiency gains as an institutional phenomenon with long‑run economic consequences: economists and policymakers should broaden analysis beyond short‑term output metrics to include institutional designs that preserve uncertainty, heterogeneity, and the capacity to generate breakthrough innovation.