Generative AI is not a neutral tool but an active participant in organisational knowledge work, accelerating decisions while amplifying existing habits and risking systemic fragility when it shortcuts foundational sensemaking.

Main Finding

Generative AI (GenAI) is not just a neutral productivity tool but an active participant in organisational collective intelligence (CI). The paper introduces a dependency-structured, five-layer framework showing that GenAI can speed and smooth higher‑level outputs while amplifying upstream weaknesses (in scanning, framing, or evidence) that produce epistemic fragility — coherent-looking decisions built on narrowed or shallow evidential bases.

Key Points

• Five-layer dependency framework for organisational CI:
• Information acquisition & environmental scanning
• Sensemaking & framing
• Shared reasoning & option generation
• Coordinated action & execution
• Organisational memory & learning
• Dependency logic: higher-level performance depends on the integrity of lower-level epistemic processes; failures propagate upward.
• GenAI’s dual effects:
  • Augmentation: extends attention, speeds synthesis, scaffolds deliberation, lowers documentation costs.
  • Distortion: narrows source diversity, creates premature coherence/anchoring, dampens substantive disagreement, diffuses ownership, and produces voluminous but thin organisational memory.
• Epistemic fragility defined: outputs that appear coherent but rest on insufficient or narrowed evidence; not necessarily immediate error but reduced resilience and learning over time.
• GenAI amplifies existing organisational tendencies (e.g., reliance on dominant sources, under-curation of knowledge) rather than reliably improving CI.
• Diagnostic tools provided: a matrix of diagnostic questions to locate bottlenecks across the five layers; a hypothetical case (Arden Finance) illustrates how compression of upstream processes can yield brittle but seemingly efficient outcomes.

Data & Methods

• Approach: conceptual/theoretical synthesis and framework-building.
  • Literature integration across information science, organisational theory (sensemaking, routines), collective intelligence, and human–AI collaboration.
  • Constructs a dependency-structured, layered model (visualised as a pyramid) to map where GenAI participates and failure modes arise.
• Tools in the paper:
  • Figure: conceptual framework showing layers and GenAI activities/failure modes.
  • Table: diagnostic guidance matrix with probing questions by layer.
  • Hypothetical case study (mid-sized financial advisory firm) to illustrate mechanisms in practice.
• Not an empirical causal study:
  • No original quantitative datasets or causal identification.
  • Intended as diagnostic/explanatory lens rather than a predictive model.
  • Limitations: conceptual framing — useful for hypothesis generation, organisational diagnosis, and governance design but not for estimating causal effect sizes.

Implications for AI Economics

• Productivity vs. robustness trade-off
  • Short-run productivity gains (speed, polished outputs) may mask long-run losses in resilience, learning, and quality of judgement.
  • Economists should distinguish measured throughput from genuine improvement in decision quality and long-term organisational capital.
• Complementarity and task reallocation
  • GenAI complements certain cognitive and documentation tasks but can substitute for parts of upstream epistemic work (search, primary-source engagement).
  • This changes returns to skills: value rises for roles that preserve and verify upstream processes (curation, critical sensemaking) and falls for rote synthesis tasks.
• Information capital, governance, and managerial value
  • Returns to investments in information infrastructure, curation, and governance likely increase — firms with superior governance may capture sustained gains.
  • Misgoverned adoption creates negative externalities (propagation of fragile decisions) that managers must internalise.
• Market structure and concentration
  • Data and model advantages can create first-mover benefits and path dependence: organisations that lock in broad, well‑curated corpora and governance may outperform others.
  • Externalities to public knowledge commons (pollution of shared corpora by low-quality AI outputs) can affect market-level information quality.
• Externalities and systemic risk
  • Epistemic fragility inside firms can cascade across markets (e.g., in finance, compliance failures, mispriced risks), implying potential systemic risk and a role for regulation and standards.
• Measurement and empirical research directions for economists
  • Key variables to measure: source diversity, primary-source engagement rates, degree of anchoring to AI outputs, changes in disagreement dynamics, thickness of documented rationales, and reuse quality of organisational memory.
  • Suggested empirical designs:
    • Field experiments randomising GenAI interfaces (e.g., summaries vs. source-linked digests) to measure effects on primary‑source use and decision outcomes.
    • Difference-in-differences / event studies around firm-level GenAI rollouts to track productivity, error rates, compliance incidents, and hiring/composition changes.
    • Audit studies comparing decisions made with vs. without GenAI scaffolds to estimate fragility and error propagation.
    • Instrumental variable or matched designs exploiting staggered adoption driven by exogenous shocks (e.g., vendor availability, regulatory changes).
    • Network-level analyses to detect spillovers and commons degradation from widespread GenAI use.
• Policy and organisational recommendations with economic implications
  • Emphasise governance that preserves upstream processes: provenance, citation to primary sources, curated knowledge bases, and explicit ownership of AI-generated artifacts.
  • Consider regulation/standards for decision‑critical domains to require explainability, traceability, or human verification thresholds.
  • Incentivise investments in information stewardship and training: these are likely high-return in an AI-enabled environment.
• Theoretical implications
  • Models of technological adoption should incorporate not only productivity gains but also changes to the information production function and learning dynamics.
  • Endogenous epistemic narrowing (reduced exploration, path dependence) should be modelled as a channel affecting long-run innovation and firm heterogeneity.

Overall, the paper provides a conceptual framework that is directly relevant to economic questions about productivity measurement, complementarities between AI and human capital, firm strategy around information assets, market-level externalities, and the design of governance or regulation to manage the fragility that GenAI can amplify.