Cadre théorique pour l’évaluation éthique des IA en entreprise

Thesis Abstract

In the contemporary corporate landscape, the deployment of artificial intelligence (AI) systems intersects with complex ethical, legal, and social implications that influence stakeholder trust, regulatory compliance, and organizational performance. The study addresses the absence of a coherent, operationalized theoretical framework to evaluate the ethics of AI in enterprise settings, with particular attention to governance, accountability, algorithmic fairness, transparency, and impact on employees and customers. The aim is to develop a rigorously specified theoretical framework—comprising constructs, relationships, and measurement—capable of guiding ethical evaluation and decision-making across diverse organizational contexts. Specific objectives are (1) to identify core ethical dimensions salient to enterprise AI through a multidisciplinary synthesis of ethics, information systems, and governance literature; (2) to articulate a parsimonious yet comprehensive model linking ethical constructs to organizational outcomes such as trust, risk, and performance; (3) to operationalize a measurement instrument and validate its reliability across industries; (4) to test the framework empirically using multi-source data; and (5) to generate practical guidance for governance mechanisms, audit processes, and policy design. The methodology adopts a multi-phase mixed-methods design. Phase I consists of a conceptual synthesis and theory-building exercise, drawing on the normative theories of deontological ethics, virtue ethics, and consequentialism, complemented by governance theories (stewardship theory, legitimacy theory) and IT ethics frameworks (availability, privacy, fairness, accountability, transparency—FAccT). Phase II develops a measurement instrument comprising 42 items across seven constructs Fairness, Transparency, Accountability, Privacy, Explainability, Human Oversight, and Organizational Justice. The instrument is validated through content validity with a panel of 12 experts in AI ethics, governance, and information systems, followed by a pilot test (n=120) to assess reliability (Cronbach’s alpha) and construct validity (confirmatory factor analysis). Phase III presents a nationwide survey of 420 professionals involved in AI governance across manufacturing, financial services, and healthcare sectors, complemented by 30 in-depth interviews with ethics officers, data scientists, and compliance managers. Data collection utilizes structured questionnaires and semi-structured interview guides. Phase IV employs quantitative analysis—partial least squares structural equation modeling (PLS-SEM) to test the theoretical model and moderation analyses for organizational size and sector—alongside thematic analysis of interview transcripts (NVivo) to enrich the quantitative findings with contextual insights. Robustness checks include multi-group analysis, measurement invariance testing, and sensitivity analyses to address potential common-method bias. Expected findings include (a) a validated, integrative theoretical framework that links ethical dimensions to measurable organizational outcomes such as perceived trust, regulatory risk exposure, and operational performance; (b) evidence that higher levels of explainability and human oversight positively influence perceived fairness and trust, while stronger privacy protections correlate with lower risk exposure; (c) sectoral variations indicating that financial services place greater emphasis on transparency and accountability, whereas healthcare prioritizes privacy and explainability due to patient safety considerations; (d) identification of imperfect mediation effects where transparency mediates the relationship between explainability and trust, and accountability moderates the impact of fairness on employee engagement. The study contributes to knowledge by offering a concrete, testable theory of AI ethics for enterprises, bridging normative ethics with governance practice, and providing a validated measurement model for periodic ethics audits of AI systems. It advances understanding of how ethical principles translate into governance arrangements, risk management, and organizational performance, thereby informing policy design, regulatory compliance, and standard-setting. Practical implications include a blueprint for AI ethics dashboards, audit checklists, and governance structures that integrate ethical evaluation into life-cycle management (data collection, model development, deployment, and monitoring). The conclusions anticipate that embedding a formal ethical framework within corporate governance enhances stakeholder trust, reduces susceptibility to ethical and regulatory risk, and fosters sustainable AI adoption. Recommendations for practice include implementing cross-functional ethics committees, periodic independent audits, and sector-specific guidelines for transparency and explainability, while future research should explore longitudinal effects and the integration of AI ethics with sustainability and human-centered design paradigms.