AI winters were less about broken algorithms than collapsed legitimacy across technical, institutional and commercial links; the current generative-AI boom risks regulatory and economic cooling unless developers and policymakers shore up evidence, documentation, restraint and plural infrastructure.

Main Finding

Mazurek & Gurczyński (2026) argue that historical “AI winters” are best understood not as pure technical failures but as legitimacy crises: breakdowns in the system of justifications that connect technical promise, funding, commercialization, and political/social authorization. They propose a hierarchical model of four interacting legitimacy gaps (capability, institutional assessment, commercialization, governance) and use it to reinterpret past AI downturns and to diagnose likely instability drivers for contemporary generative AI—principally a potential regulatory‑economic cooling rather than a simple repeat of prior technical winters.

Key Points

• Four legitimacy gaps (hierarchical, interacting):
  • Capability gap: technical approaches fail to scale to promised goals (technical substrate).
  • Institutional assessment gap: sponsors’ expectations/maturity horizons outpace technology maturation (funding/assessment).
  • Commercialization gap: prototypes are overgeneralized into products beyond their operating conditions (product/market mismatch).
  • Governance gap: lack of legal, moral, political authorization (transparency, accountability, rights, public trust).
• Historical reinterpretation:
  • First AI winter (late 1960s–1970s): primarily capability + institutional assessment crisis (ALPAC, Lighthill; funders withdrew patience).
  • Second AI winter (late 1980s): crisis of product, brand, and commercial ecosystem (commercial overreach and loss of market credibility).
• Contemporary generative AI differs because it is deeply infrastructural (models function as epistemic infrastructure: retrieving, ranking, stabilizing public truth). Therefore failure modes include regulatory and institutional legitimacy issues as much as technical limitations.
• Likely contemporary risks: regulatory‑economic cooling driven by compliance costs, disputes over training data, concentration of compute/infrastructure, information manipulation, the rise of algorithmic verification, and growing documentation/verification demands.
• Governance recommendations offered by the authors: promise restraint, robust evidence of deployment, documentation obligations, and infrastructural pluralization to reduce single‑point legitimacy risk.

Data & Methods

• Approach: historical‑conceptual, interpretive (not archival or bibliometric). The paper reconstructs selected turning points and organizes them via legitimacy theory and the sociology of expectations.
• Source selection (purposive):
  • Canonical technical/programmatic texts (e.g., Turing, perceptron critiques, neural network milestones, transformer/LLM papers).
  • Institutional documents shaping funding/assessment (ALPAC 1966, Lighthill report, EU AI Act, NIST guidance).
  • Historical syntheses of AI development (Crevier, Nilsson, Haigh).
  • Recent normative and governance literature (algorithmic verification, epistemic infrastructure, trustworthy AI).
• Analytical framing:
  • Combines Suchman’s legitimacy (pragmatic, moral, cognitive) with sociology of expectations (promises shape investment and coordination).
  • Distinguishes descriptive (historical facts), analytical (framework linking gaps), and normative/prognostic claims (policy implications).
• Limitations acknowledged by authors:
  • Interpretive reconstruction rather than exhaustive causal history.
  • Not a technical evaluation of modern models’ proximity to “general intelligence.”
  • Uses “first/second AI winter” as retrospective labels for clusters of delegitimization rather than globally uniform events.

Implications for AI Economics

• Investment and valuation
  • Legitimacy risk should be treated as a central systematic risk factor. Investor patience (time horizon) and sponsors’ assessment criteria materially affect funding flows; fast promise cycles can create “legitimacy debt” and abrupt re‑pricing.
  • Regulatory‑economic cooling (compliance costs, litigation/data rights disputes) can reduce expected returns and increase capital costs for model providers and startups.
  • Infrastructure assets (data, compute, orchestration, provenance/verification systems) gain relative value compared with transient model/model‑only plays.
• Market structure and competition
  • Compute and data concentration heighten systemic legitimacy risk; dominant infrastructure providers increase political/regulatory exposure and the likelihood of intervention that reshapes market dynamics.
  • Documentation and verification obligations create entry frictions and favor firms that can absorb compliance cost—potentially increasing winner‑take‑most dynamics unless governance fosters infrastructural pluralization.
  • A growing market for algorithmic verification, provenance services, and audit tools: new industrial niches and firms offering third‑party verification could emerge as valuable intermediaries.
• Commercialization strategy and product risk
  • Firms should avoid overgeneralizing prototypes; economic models that assume rapid productization of capabilities are vulnerable to commercialization gaps (narrow operating envelopes, fragility under distributional shift).
  • Business models must price in costs of documentation, human oversight, contestability mechanisms, and potential mitigation of misinformation harms.
  • Demand for demonstrable deployment evidence will increase: randomized trials, field evidence, and sector‑specific validation become economic assets.
• Labor and complementarities
  • If generative models become epistemic infrastructure, their economic role shifts toward mediating knowledge markets (search, curation, verification). This changes where value accrues (platforms, verifiers, content producers) and alters labor complementarities (verification, prompt‑engineering, oversight).
• Policy and macroeconomic implications
  • Policymakers’ governance choices (e.g., strict documentation, data‑use rules, verification standards) materially reshape sector returns and incentive structures; scenario analyses should include regulatory‑economic cooling cases.
  • Public‑good investments in pluralized infrastructure (open compute, federated datasets, public verification services) can reduce concentration risk and lower systemic legitimacy exposure—these have long‑run economic benefits by preserving competition and innovation diversity.
• Practical recommendations for economists, investors, and policymakers
  • Incorporate legitimacy gap analysis into risk models: map where capability, assessment, commercialization, or governance gaps are most salient for a given project.
  • Stress‑test business models for compliance/verification costs and for disputes over data rights; include possible delays from evidence requirements.
  • Value infrastructure and provenance/verification capabilities explicitly when valuing firms; consider cost of shifting away from dominant compute/data providers.
  • Support/incentivize infrastructural pluralization (public compute, interoperable datasets, verification standards) to reduce systemic concentration and legitimacy fragility.
  • Encourage staged promise‑making and phased evidence disclosure to align investor expectations with maturation horizons (promise restraint + evidence of deployment).
• New economic sectors and measurement needs
  • Growth of markets for AI auditing, verification, and documentation services—expect new firms and regulatory demand.
  • Need for new metrics capturing epistemic externalities (misinformation amplification, trust erosion) and for measuring “legitimacy debt” exposure in portfolios.

Limitations to apply cautiously: the paper is interpretive and purposive in source selection; it diagnoses mechanisms and plausible scenarios rather than offering precise probability forecasts. Nonetheless, its legitimacy‑gap framework is directly actionable for economic modeling of AI investment, market design, and regulatory impact assessment.