New Audit Method Tracks How AI Ethics Principles Really Change When They Cross Borders

Artificial intelligence ethics has become one of the fastest-spreading normative projects in modern governance. Documents from the European Commission’s ethics guidelines to UNESCO’s 2021 recommendation circulate a shared vocabulary of transparency, fairness, accountability, and human oversight, and governments across Asia have adopted strikingly similar language in their own frameworks, from Japan’s Social Principles of Human-Centric AI to China’s governance principles for the new generation of AI and India’s Responsible AI strategy. On the surface, the world’s AI ethics documents look increasingly alike. But a new theoretical paper argues that this lexical convergence is deceptive: the same words can hide radically different allocations of power, responsibility, and remedy once principles are translated into institutions and operational rules.

The paper, published in the journal AI & Society by Masaya Ochiai, an independent researcher based in Nagoya, Japan, introduces a method called the normative translation audit, or NTA. Rather than treating AI ethics principles as fixed ideas that either spread or fail to spread, the framework treats them as material that is inevitably transformed as it moves between organizations, jurisdictions, and regulatory levels. The central claim is that comparative research on AI ethics has a structural blind spot: when analysts compare Asian frameworks against European or American ones, they often end up measuring deviation from a Western baseline rather than understanding how norms actually work within and between Asian societies.

Ochiai names this blind spot reference-point drift. The problem arises, he argues, when Euro-American frameworks silently become the reference point for comparison. In that situation, intra-regional comparison among Asian countries degenerates into deviation scoring: a framework is judged as more or less aligned with the OECD recommendation, the NIST AI Risk Management Framework, or the EU’s Assessment List for Trustworthy Artificial Intelligence, rather than being analyzed on its own terms. This matters because Asia contains enormous internal plurality, spanning vastly different political systems, legal traditions, and levels of AI development, while at the same time being tightly interconnected through cross-border data flows, supply chains, and regional institutions such as ASEAN. A comparison method that can only ask how closely a country matches a Western template cannot jointly handle that internal diversity and external interdependence.

The intellectual machinery behind the argument draws on established scholarship in policy transfer and norm diffusion. Ochiai cites work on norm localization in Asian regionalism, studies of how international human rights law is translated into local justice, and the sociology of standards, which shows that standardization is never a neutral technical act but a process that reorganizes social relationships. He also draws on the audit society literature, which describes how modern institutions increasingly rely on rituals of verification, and on critiques showing that principles alone cannot guarantee ethical AI. The recurring theme is translation: a principle such as fairness does not simply arrive intact in a new jurisdiction. It is redefined, reweighted, and re-embedded in local institutional arrangements, and those transformations are where governance actually happens.

To make these transformations visible, NTA redesigns the unit of comparison. Instead of comparing principle vocabularies, the method focuses on four branch points that determine practical governance outcomes. The first is operational definitions: how a principle like transparency or fairness is concretely specified, what counts as compliance, and what technical measures are named. The second is veto rights: who has the authority to stop or block an AI deployment, and under what conditions. The third is remedy lines: what channels exist for people harmed by an AI system, what compensation or correction is available, and who bears the burden of proof. The fourth is accountability assignment: which actors, institutions, or roles are held answerable when systems fail, and how that answerability is enforced.

These four branch points function as a diagnostic grid. Two jurisdictions may both endorse accountability in their official documents, yet one may assign it to a voluntary industry review board while the other embeds it in a statutory regulator with sanctioning powers. Vocabulary-based comparison would score both as aligned; branch-point comparison would reveal a divergence with real consequences for affected people. Similarly, the veto-rights branch point distinguishes frameworks in which civil society or affected communities can formally contest a deployment from those in which contestation is limited to internal corporate review. The remedy branch point captures whether harms trigger compensation, correction, or merely documentation. By fixing the comparison unit at these decision-relevant junctures, NTA aims to prevent the silent drift back to Western baselines.

NTA is deliberately positioned as a minimal procedure rather than a full compliance audit. Ochiai is explicit that it is not a pass/fail certification and not third-party assurance in the style of financial or security audits. Its purpose is narrower and more methodological: to fix comparison units before analysis begins and to trace whether register updates, meaning documented changes in how principles are specified, actually changed decisions in specifications, operations, or remedy design. The method requires three artifacts. The first is a Branch Register, a structured record of how each framework under study resolves each of the four branch points. The second is a Decision-Change Log, which tracks whether updates to that register corresponded to changes in actual decisions. The third is a multi-anchor memo, which documents which external frameworks were selected as comparison anchors and, crucially, which were excluded, so that no single external framework becomes the only baseline against which everything else is measured.

The paper includes a brief demonstration using the Association of Southeast Asian Nations and Singapore, drawing on the ASEAN Guide on AI Governance and Ethics, its 2025 expansion covering generative AI, and Singapore’s Model AI Governance Framework. The demonstration illustrates how branch-point comparison surfaces divergences, particularly around contestation and decision review, that vocabulary-based comparison tends to hide. Where a word-level analysis might conclude that ASEAN and Singaporean documents broadly echo global principles on transparency and human oversight, examining who can contest decisions and how reviews are triggered reveals a more textured picture of how responsibility is distributed across regional guidance and national implementation. The author presents this as an illustration rather than a completed empirical test, and the paper outlines a testable protocol, including a minimal coding sheet with fields for anchor identity, branch point, evidence locators, confidence levels, and disagreement flags for double-coded entries, for future studies to evaluate the method empirically.

The significance of the proposal extends beyond academic comparison. As AI regulation matures worldwide, regulators and companies increasingly benchmark their frameworks against one another, and the choice of benchmark shapes what counts as progress. If Asian frameworks are always scored against European or American templates, the analysis risks reproducing what postcolonial scholars have long criticized: the subaltern perspective is heard only insofar as it echoes the metropole. NTA’s multi-anchor requirement is a direct response to that risk. By forcing analysts to document anchor selection and exclusion, the method makes the politics of comparison explicit and creates space for intra-Asian baselines, such as ASEAN regional guidance, to serve as legitimate reference points in their own right.

The paper also arrives amid a broader wave of scholarship questioning whether the global AI ethics movement has delivered on its promises. Large-scale analyses of ethics guidelines have shown substantial convergence in stated principles alongside persistent disagreement about implementation, and researchers have documented how abstraction in sociotechnical systems can hollow out fairness commitments when they meet real-world engineering constraints. NTA contributes to this conversation by shifting attention from what frameworks say to what their words do: how they define, who they empower to block, what they promise the harmed, and whom they hold to account. Whether the method proves robust in empirical use remains to be tested, but it offers researchers and policymakers a concrete toolkit for asking a deceptively simple question with large consequences: when an AI ethics principle crosses a border, what exactly changes, and who benefits from the change?

Subject of Research: A normative translation audit design for comparing how AI ethics principles are translated into governance across Asian jurisdictions

Article Title: Normative translation audit (NTA): an audit design for AI ethics principle transplantation that enables intra-Asian comparison

Article References: Normative translation audit (NTA): an audit design for AI ethics principle transplantation that enables intra-Asian comparison. (n.d.). https://doi.org/10.1007/s00146-026-03362-6

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03362-6

Keywords: AI ethics, normative translation audit, policy transfer, accountability, ASEAN, Singapore, comparative governance, audit design, norm localization, reference-point drift, AI governance, intra-Asian comparison