New ethical matrix exposes bias and accountability gaps in near-term AI

A new paper develops a structured tool to evaluate the challenges surrounding artificial intelligence's accountability and fairness, arguing that both ethical values and technical realities must be assessed together if societies are to govern AI responsibly.

Titled “The bi-level ethical matrix and near-term AI” in AI & Society, the study authored by Clinton Castro expands on earlier models for ethical evaluation. The author refines the “ethical matrix,” first created by Ben Mepham for agricultural technologies and later adapted for AI by Cathy O’Neil and Hanna Gunn, into a bi-level system designed specifically for what he calls near-term AI: the algorithms already in use for credit scoring, predictive policing, risk assessments, and healthcare triage.

How can ethical values be retained in AI evaluation?

Earlier ethical matrices either focused too narrowly on descriptive issues, such as error rates and efficiency, or too broadly on abstract values, leaving out crucial details. His bi-level matrix resolves this by creating two distinct but connected tiers.

The moralized level captures foundational values, well-being, autonomy, and fairness, that allow regulators, developers, and the public to frame why certain AI outcomes matter. Importantly, the framework can also incorporate domain-specific values, such as legitimacy in policing, where constitutional rights and public trust are at stake.

The non-moralized level organizes empirical issues under the acronym B.I.A.S.T.: bias, impacts, accountability, security, and transparency. These categories track measurable realities such as predictive parity, error rates, data access, or security vulnerabilities. By separating facts from values, the framework clarifies debates where technical evidence and moral reasoning are often entangled.

The author argues this structure allows for greater clarity when evaluating controversies. For instance, in assessing predictive policing algorithms, the non-moralized tier records statistical disparities in error rates between racial groups, while the moralized tier addresses whether those disparities amount to unfairness. This dual approach ensures that empirical findings are not mistaken for ethical conclusions, but neither are ethical judgments left detached from measurable realities.

How does context shape ethical analysis of AI?

Another innovation in Castro’s framework is its explicit recognition of context sensitivity. Previous ethical matrices were often criticized for being too abstract, failing to adapt to the unique values and goals of different sectors. The bi-level model addresses this gap by embedding social domain values directly into the evaluation process.

In healthcare, this might mean prioritizing patient safety and trust; in education, fairness in learning opportunities; in policing, legitimacy defined by constitutional rights and proportionality. By linking ethical reflection to the actual purposes and standards of specific domains, the matrix prevents one-size-fits-all judgments and highlights where trade-offs differ by field.

The study demonstrates how this contextual sensitivity exposes new insights. For example, in policing, legitimacy requires not just competence and fairness but also signaling trustworthiness. When predictive policing systems are opaque, legitimacy itself is undermined. By adding legitimacy as a domain-specific value, the matrix surfaces how transparency in this setting is not merely a technical preference but a prerequisite for lawful and democratic policing.

This attention to context also improves participation. Affected parties, from defendants in criminal justice systems to patients in healthcare, can be engaged more meaningfully when ethical evaluations explicitly reflect the values central to their lived realities.

What role can the matrix play in regulation and governance?

While Castro stresses the potential of the bi-level matrix to sharpen ethical reflection, he is clear about its limits. The tool is not a replacement for democratic oversight, regulation, or journalistic scrutiny, but rather a way to improve both individual and collective reasoning about AI.

The study stresses that completing a matrix does not resolve trade-offs. Transparency may reduce efficiency, while bias mitigation may lower accuracy. Practical wisdom, informed by consistency, knowledge, and imagination, is still required to make final decisions. Organizations can cultivate these qualities through ethics officers, confidential reporting channels, employee training, and open discussions about values.

Castro also addresses accountability gaps often linked to AI. Systems may produce outcomes that feel as though responsibility should be assigned, yet no single actor appears accountable. By clarifying roles - developers, deployers, data subjects, decision subjects, and society - the matrix helps identify who can and should be held responsible. Mechanisms for appeal, consent withdrawal, or correction are positioned as critical for ensuring accountability does not collapse under the weight of automation.

Beyond institutions, the framework underscores the role of the public and press. Investigative reporting, such as ProPublica’s work on algorithmic bias, creates external pressure for reform, while citizens can engage through town halls, advocacy, and political participation. The matrix cannot compel action, but it can guide deliberation and provide a common language for diverse stakeholders.