<p>Large-scale AI platforms continuously shape the information environments of billions of people through the cumulative delivery of countless outputs in the form of recommendations, advertisements, and generated text. Existing regulatory approaches fall into three categories, each with its own inherent limitations. Input regulation cannot escape the trade-off between privacy and effectiveness; internal regulation is blocked by the epistemological barrier of the black box. Both allow for continuous improvement through technological innovation, but innovation alone cannot achieve complete regulation. Output regulation, by contrast, is the regulatory layer capable of achieving formal completeness—through the matching of declared metadata against democratically established Rules—and points in the right direction. However, it currently lacks a mechanism for continuous, verifiable enforcement at societal infrastructure scale. This difficulty stems not from a principled constraint but from an insufficiency of institutional and technical infrastructure. Building on this recognition, this paper proposes a framework for applying VRAIO (Verifiable Record of AI Output) to the governance of platform AI outputs. VRAIO mandates structured metadata declarations for all output candidates, has an independent third-party body (the Recorder) verify formal alignment with democratically established Rules, and records the results in a tamper-proof ledger. By establishing output regulation as the foundational infrastructure, VRAIO realizes an integrated governance base that connects the continuous improvement of input and internal regulation to society in a verifiable form. By shifting the focus of governance from model internals to the outputs that actually appear in society, VRAIO achieves institutional accountability for AI outputs without presupposing that the black box has been opened. Using YouTube as the primary case study, the paper demonstrates that VRAIO can govern not only individual outputs but also sequential and cumulative output patterns such as radicalization pathways, engagement-maximization loops, and targeted advertising directed at minors. VRAIO further functions as connective infrastructure that translates the obligations declared by the General Data Protection Regulation (GDPR), the EU AI Act, and the Digital Services Act (DSA) into real-time, output-level enforcement, providing a unified verification layer that cuts across fragmented regulatory regimes. The paper examines the major objections—the reliability of metadata declarations, the international fragmentation of regulation, and the impact on innovation—and argues that VRAIO’s incentive design renders non-compliance an irrational choice from the outset. VRAIO is a safety valve. Inconspicuous in normal operation, it is institutionally powerful—infrastructure that makes accountability for AI platform outputs a reality.</p>

VRAIO as a safety valve: governing AI platform outputs at societal scale

期刊论文
发表日期：

2026年6月16日

查看全文

Yusaku Fujii

摘要

Large-scale AI platforms continuously shape the information environments of billions of people through the cumulative delivery of countless outputs in the form of recommendations, advertisements, and generated text. Existing regulatory approaches fall into three categories, each with its own inherent limitations. Input regulation cannot escape the trade-off between privacy and effectiveness; internal regulation is blocked by the epistemological barrier of the black box. Both allow for continuous improvement through technological innovation, but innovation alone cannot achieve complete regulation. Output regulation, by contrast, is the regulatory layer capable of achieving formal completeness—through the matching of declared metadata against democratically established Rules—and points in the right direction. However, it currently lacks a mechanism for continuous, verifiable enforcement at societal infrastructure scale. This difficulty stems not from a principled constraint but from an insufficiency of institutional and technical infrastructure. Building on this recognition, this paper proposes a framework for applying VRAIO (Verifiable Record of AI Output) to the governance of platform AI outputs. VRAIO mandates structured metadata declarations for all output candidates, has an independent third-party body (the Recorder) verify formal alignment with democratically established Rules, and records the results in a tamper-proof ledger. By establishing output regulation as the foundational infrastructure, VRAIO realizes an integrated governance base that connects the continuous improvement of input and internal regulation to society in a verifiable form. By shifting the focus of governance from model internals to the outputs that actually appear in society, VRAIO achieves institutional accountability for AI outputs without presupposing that the black box has been opened. Using YouTube as the primary case study, the paper demonstrates that VRAIO can govern not only individual outputs but also sequential and cumulative output patterns such as radicalization pathways, engagement-maximization loops, and targeted advertising directed at minors. VRAIO further functions as connective infrastructure that translates the obligations declared by the General Data Protection Regulation (GDPR), the EU AI Act, and the Digital Services Act (DSA) into real-time, output-level enforcement, providing a unified verification layer that cuts across fragmented regulatory regimes. The paper examines the major objections—the reliability of metadata declarations, the international fragmentation of regulation, and the impact on innovation—and argues that VRAIO’s incentive design renders non-compliance an irrational choice from the outset. VRAIO is a safety valve. Inconspicuous in normal operation, it is institutionally powerful—infrastructure that makes accountability for AI platform outputs a reality.