Authentic Intelligence: a systems architecture identifies twelve failure modes and governance levers that determine whether organisations keep humans accountable in algorithmic decision-making; cross-validation with 685 EU cases and 37 discontinuations finds six failure modes are highly prevalent and six others moderately common, offering a practical diagnostic for Article 14 compliance.

Main Finding

The paper defines "Authentic Intelligence" as a systems-level construct describing how human judgement is deliberately embedded and maintained within algorithmic decision systems. It delivers a reusable socio-technical architecture and diagnostic framework that specifies the components, decision loci, governance mechanisms, and empirically derived failure modes that determine whether human accountability is preserved in digital strategy implementation. The framework is validated against a large EU implementation catalogue and documented discontinuations and is positioned to support compliance with EU AI Act Article 14.

Key Points

• Novel construct: Introduces "Authentic Intelligence" to capture system-level conditions that keep human judgement meaningfully present in algorithmic decisions.
• Socio-technical architecture: Specifies four structural layers of the problem space:
  • Six primary system functions (core capabilities a system must support to sustain human-accountable decision making).
  • Three decision loci (points in the system where human judgement can be exercised or required).
  • Four governance mechanisms (institutional and technical levers that sustain accountability).
  • Twelve empirically derived failure modes (how embedding human judgement breaks down in practice).
• Empirical validation:
  • 50 semi-structured interviews with senior managers across multinationals produced mechanism-revealing qualitative data.
  • Cross-case mapping against 685 entries from the European Commission’s JRC AI implementation catalogue to assess prevalence and correspondence.
  • Triangulation with 37 documented implementation discontinuations to surface failure patterns.
• Triangulation results: Six of the twelve failure modes show high correspondence in the JRC catalogue (≥ 20% citation rate); the remaining six show moderate correspondence—indicating both common and less frequent practical breakdowns.
• Practical deliverable: A diagnostic framework and architecture that practitioners and policymakers can apply to assess and design systems that preserve human accountability, with direct relevance for meeting Article 14 requirements under the EU AI Act.

Data & Methods

• Stage 1 — Mechanism-revealing thematic analysis:
  • Data: 50 semi-structured interviews with senior managers in multinational organisations.
  • Purpose: Elicit how organisations attempt to keep human judgement embedded in algorithmic systems and identify reported barriers/failures.
• Stage 2 — Configurational cross-case mapping:
  • Data: 685 cases from the European Commission’s JRC AI implementation catalogue.
  • Purpose: Map interview-derived mechanisms and configurations to a broad population of AI implementation cases to assess generalisability and prevalence.
• Stage 3 — Failure mode triangulation:
  • Data: 37 documented implementation discontinuations.
  • Purpose: Compare interview-reported barriers and the architecture’s failure modes against real-world stoppages to validate failure-mode salience.
• Analysis approach: System-oriented, multi-method triangulation combining qualitative thematic extraction, configurational mapping across cases, and cross-validation with discontinuation cases to strengthen external validity and diagnostic utility.

Implications for AI Economics

• Adoption and diffusion:
  • The architecture clarifies costs and organizational capabilities required to preserve human accountability, helping explain heterogeneity in AI adoption across firms and sectors.
  • Failure modes illuminate frictions (coordination, monitoring, skill gaps) that can slow or halt implementations, affecting diffusion dynamics and technology-led productivity gains.
• Regulatory compliance and cost structures:
  • Direct relevance to EU AI Act Article 14 means firms will face measurable compliance tasks; the framework can help estimate compliance costs and necessary governance investments.
  • Governance mechanisms identified by the study can be modeled as organizational capital that affects marginal costs of deploying high-risk AI systems.
• Strategic and competitive effects:
  • Firms that successfully embed Authentic Intelligence may obtain reputational and regulatory advantages (lower enforcement risk, better stakeholder trust), creating potential first-mover benefits and barriers to entry.
  • Conversely, persistent failure modes can lead to stranded investments and market exits, reshaping industry concentration.
• Policy design:
  • Policymakers can use the diagnostic architecture to target interventions (standards, certification, guidance) that reduce systemic failure rates and lower aggregate compliance costs.
  • The evidence-based mapping to JRC cases supports calibrating regulation to realistic implementation constraints across sectors.
• Research opportunities:
  • Quantify the economic impact of each failure mode on implementation cost, time-to-deploy, and realized value.
  • Empirical estimation of the returns to investment in specific governance mechanisms (e.g., audit capability, human oversight interfaces).
  • Modeling how regulation (Article 14 and similar) changes firm incentives, adoption thresholds, and market structure given the costs of maintaining Authentic Intelligence.

If you want, I can (a) extract specific managerial checklists from the architecture for operational use, (b) sketch an economic model incorporating governance costs and failure-mode probabilities, or (c) draft policy recommendations tied to the EU AI Act. Which would be most useful?