IOCSS Research Paper | AI Ethics and Governance Series | 2025
Abstract: This paper examines the proliferation of AI ethics frameworks and asks whether the current approach to AI governance—through principles, voluntary codes, and anticipatory regulation—is adequate to the governance challenges that artificial intelligence systems present. We argue that existing frameworks suffer from three structural limitations: they are too abstract to guide specific decisions, they are too process-oriented to generate substantive constraints, and they are insufficiently attentive to the distributional dimensions of AI impact. We propose a research agenda for more adequate AI governance that integrates philosophical rigor with institutional design.
1. Introduction: The Framework Proliferation Problem
The global AI ethics landscape is characterized by an extraordinary proliferation of frameworks, principles, guidelines, and codes of conduct. By conservative count, over two hundred distinct AI ethics frameworks had been published by 2023, produced by technology companies, national governments, international organizations, academic institutions, civil society organizations, and multistakeholder bodies. This proliferation raises a fundamental question: if so many frameworks exist and continue to multiply, why do concrete governance challenges remain so incompletely addressed?
IOCSS approaches this question not primarily as a policy problem but as a philosophical problem. The difficulty with existing AI ethics frameworks is not merely that they are too numerous or that their proliferation creates confusion (though both are true). The deeper problem is that they rest on inadequate philosophical foundations—foundations that make them incapable of generating the specific guidance that AI governance requires.
This philosophical diagnosis is consequential because it suggests that the solution is not simply more frameworks, better-coordinated frameworks, or frameworks with more enforcement mechanisms (though enforcement is certainly lacking). It requires a rethinking of the philosophical approach to AI governance at a more fundamental level.
2. What Current Frameworks Get Right
The existing literature on AI ethics has established a productive consensus on several core principles. The most widely cited—fairness, accountability, transparency, and safety—each identify genuine governance concerns that require institutional response.
Fairness captures the requirement that AI systems not produce outcomes that discriminate systematically against protected groups or that amplify existing social inequalities. This concern has been operationalized in technical research on algorithmic fairness, which has produced important tools for detecting and mitigating certain forms of algorithmic discrimination.
Accountability identifies the requirement that there be identifiable agents responsible for AI systems' impacts, and that these agents be answerable for harms. This concern connects to fundamental principles of democratic governance and legal liability.
Transparency captures the requirement that the operation of AI systems be explicable—to regulators, to affected individuals, and to the public—in ways that enable meaningful oversight. This requirement connects to both epistemic values (we should understand how decisions affecting us are made) and democratic values (consequential systems should be publicly legible).
Safety identifies the requirement that AI systems not cause physical, psychological, or social harm, and that safety properties be verifiable and maintained under deployment conditions.
These principles are real and important. The problem is not that they are wrong but that they are insufficient—they do not, in their abstract form, resolve the specific cases where governance decisions must be made.
3. The Abstraction Problem
The most fundamental structural limitation of existing AI ethics frameworks is their abstraction: they articulate principles at a level of generality that is compatible with a wide range of specific decisions. Claiming to be "fair" or "transparent" does not resolve the competing interpretations of what these terms require in specific contexts.
Consider the principle of fairness. Academic computer science has identified multiple formally defined and mathematically incompatible fairness criteria: demographic parity (equal rates of positive outcomes across groups), equalized odds (equal true and false positive rates across groups), individual fairness (similar individuals should receive similar outcomes), and counterfactual fairness (outcomes should not change if an individual's protected characteristics changed while holding other variables constant). In many real-world applications, these criteria cannot all be satisfied simultaneously. Which criterion should govern, and why, is a philosophical and political question that cannot be answered by appealing to "fairness" as an abstract principle.
The abstraction problem becomes acute in the context of AI systems that operate in culturally specific contexts. Fairness norms are not culturally neutral: what counts as a relevant group, what outcomes matter, and what forms of differential treatment are acceptable vary across social contexts. An AI ethics framework designed primarily with Western liberal democratic assumptions may generate requirements that are inapplicable or actively problematic in different cultural contexts.
4. The Process-Substance Problem
A second structural limitation of existing frameworks is their emphasis on process over substance. Many AI ethics frameworks specify procedural requirements—impact assessments, stakeholder engagement, documentation requirements, auditing protocols—without specifying what substantive constraints should govern the outcomes of these processes.
Procedural requirements are not without value: they create institutional structures, documentation trails, and accountability mechanisms that can support oversight. But procedures without substantive constraints can produce "ethics washing"—the appearance of ethical compliance without the reality of ethical constraint. A company that conducts a stakeholder engagement process, produces an impact assessment, and documents its consideration of ethical concerns has technically complied with procedural requirements even if the substantive outcome of its process is one that most observers would judge ethically unacceptable.
The philosophical response to this problem requires developing accounts of substantive AI ethics that specify what kinds of outcomes are impermissible regardless of process, and what kinds of decisions require specific substantive constraints rather than merely procedural safeguards. This is harder and more contentious than process specification, which is precisely why many frameworks avoid it.
5. The Distributional Problem
Existing AI ethics frameworks have been extensively criticized for their inadequate attention to distributional questions: who benefits and who bears the costs of AI systems, and how do AI-driven changes interact with existing inequalities of wealth, power, and opportunity.
The distributional dimensions of AI are multiple. At the global level, the concentration of AI development capacity in a small number of technology companies in the United States and China, and the deployment of AI systems developed in these contexts in societies with very different conditions, creates patterns of technological influence that reproduce and potentially deepen global inequalities. At the national level, the economic impacts of AI-driven automation are distributed very unequally across occupational categories and socioeconomic groups. At the individual level, the burdens of data extraction, content moderation, and model training that underpin AI capabilities are frequently borne by low-income workers, many of them in the Global South.
IOCSS's interest in these distributional questions connects to its broader commitment to cultural and social equity in the contexts it studies. AI systems that shape what cultural content is produced, distributed, and consumed have direct implications for the cultural diversity that IOCSS values. The dominance of AI recommendation systems that optimize for engagement metrics derived from primarily Anglophone datasets has measurable effects on the visibility and viability of smaller cultural traditions.
6. Toward Adequate AI Governance: A Research Agenda
IOCSS proposes a research agenda for more adequate AI governance centered on three requirements:
Specificity. Governance frameworks should develop the philosophical and analytical tools to resolve specific competing claims rather than merely articulating abstract principles. This requires both philosophical precision about the grounds and limits of each principle and empirical research on the specific contexts in which AI systems are deployed.
Substantive Constraint. Governance frameworks should specify categories of AI application that are impermissible regardless of procedural compliance, and should develop robust criteria for evaluating whether specific applications meet substantive ethical standards. This requires engagement with normative political theory, rights frameworks, and the specific values of the communities affected.
Distributional Accountability. Governance frameworks should develop explicit mechanisms for assessing and addressing the distributional impacts of AI systems. This includes both backward-looking accountability for harms that have occurred and forward-looking requirements for the distribution of benefits.
7. Conclusion
The proliferation of AI ethics frameworks is a symptom of genuine governance challenges rather than evidence that these challenges are being adequately addressed. IOCSS contributes to this field from a philosophical perspective, arguing that more adequate governance requires philosophical rigor—about what the principles mean, what they require in specific cases, and how they relate to the substantive values of the communities they are designed to serve—rather than merely more frameworks.
This paper was prepared by the IOCSS AI Ethics Research Division. Correspondence: research@iocss.org