Legitimacy in Open-Source AI Governance: Authority, Capability, and the Tension Between Constraint and Participation

In all three systems (China, the United States, and the European Union), openness is no guarantee of legitimacy in authority and participation.

“In all three systems, openness is no guarantee of legitimacy in authority and participation.”

Open-source AI governance systems in China, the United States, and the European Union differ in how they balance authority and participation.

“The study finds that open-source AI governance systems differ in how they balance authority and participation: centralized authority and low contestability in China, dispersed governance and high concentrated private capacity in the United States, and procedural safeguards and high participation thresholds in the European Union.”

The Interim Measures assign supervisory responsibility to a range of existing government departments, including those responsible for cyberspace governance, industrial policy, public security, and media administration, each acting within its established mandate.

“The Interim Measures for the Management of Generative Artificial Intelligence Services assign supervisory responsibility to a range of existing government departments, including those responsible for cyberspace governance, industrial policy, public security, and media administration, each acting within its established mandate (Interim Measures, Art. 16).”

• cites Interim Measures, Art. 16Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The EU AI Act establishes specific disclosure obligations for certain uses of AI, including cases in which individuals interact with AI systems or are exposed to AI-generated or manipulated content.

“In addition, the Act establishes specific disclosure obligations for certain uses of AI, including cases in which individuals interact with AI systems or are exposed to AI-generated or manipulated content.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The United States structures transparency through evaluation, reporting, and technical standards.

“The United States adopts a different approach, structuring transparency through evaluation, reporting, and technical standards.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Article 17 of the Interim Measures mandates security assessments and algorithm filing procedures for services with certain levels of social impact.

“while Article 17 mandates security assessments and algorithm filing procedures for services with certain levels of social impact (Art. 17).”

• cites Art. 17Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The EU AI Act establishes a harmonized legal framework governing the development, placement on the market, and use of AI systems across Member States.

“The Artificial Intelligence Act establishes a harmonized legal framework governing the development, placement on the market, and use of AI systems across Member States (EU AI Act, Art. 1).”

• cites EU AI Act, Art. 1Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The European Union constructs transparency as a formal legal requirement embedded within the regulatory framework itself.

“The European Union constructs transparency as a formal legal requirement embedded within the regulatory framework itself.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Executive Order 14110 calls for mechanisms to identify and label AI-generated content, enabling users and institutions to distinguish synthetic outputs from human-generated information.

“It also calls for mechanisms to identify and label AI-generated content, enabling users and institutions to distinguish synthetic outputs from human-generated information.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Executive Order 14110 requires developers of certain advanced AI systems to provide federal authorities with information regarding model development, training processes, and associated risks.

“At the same time, the Order requires developers of certain advanced AI systems to provide federal authorities with information regarding model development, training processes, and associated risks (Sec. 4).”

• cites Sec. 4Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

U.S. responsibility is distributed across institutions such as the Department of Commerce, NIST, the Department of Homeland Security, and other federal bodies, each governing a different aspect of technological development, economic regulation, or national security.

“Responsibility is distributed across institutions such as the Department of Commerce, NIST, the Department of Homeland Security, and other federal bodies, each of which governs a different aspect of technological development, economic regulation, or national security.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The EU AI Act identifies transparency obligations as a core component of its governance architecture.

“The Artificial Intelligence Act identifies transparency obligations as a core component of its governance architecture (EU AI Act, Art. 1).”

• cites EU AI Act, Art. 1Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Executive Order 14110 emphasizes reliable and standardized testing, including ongoing monitoring of AI system performance.

“Executive Order 14110 emphasizes the importance of reliable and standardized testing, including ongoing monitoring of AI system performance (EO 14110, Sec. 2).”

• cites EO 14110, Sec. 2Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

A growing body of research reframed transparency as a set of technical and administrative practices.

“A growing body of research therefore reframed transparency as a set of technical and administrative practices.”

• cites Felzmann et al. 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Haibe-Kains et al. 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Hernandez-Orallo et al. 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Visibility alone does not resolve questions about who sets standards, whose contributions shape governance decisions, or how responsibility should be distributed within collaborative systems.

“Yet visibility alone does not resolve questions about who sets standards, whose contributions shape governance decisions, or how responsibility should be distributed within collaborative systems.”

Transparency was institutionalized as a bureaucratic requirement embedded in governance procedures, rather than functioning primarily as a democratic principle enabling public oversight.

“Rather than functioning primarily as a democratic principle enabling public oversight, transparency was institutionalized as a bureaucratic requirement embedded in governance procedures.”

Transparency was progressively redefined within AI ethics and governance scholarship as a problem of documentation and procedural oversight.

“However, as algorithmic systems became more opaque and technically sophisticated, transparency was progressively redefined within AI ethics and governance scholarship as a problem of documentation and procedural oversight.”

The result is a governance model in which transparency is increasingly measured by the completeness of documentation rather than by the extent to which power, decision-making authority, or responsibility are genuinely made visible.

“The result is a governance model in which transparency is increasingly measured by the completeness of documentation rather than by the extent to which power, decision-making authority, or responsibility are genuinely made visible.”

Governance within OSS communities was frequently described as emerging organically from transparent collaborative practices, including version control systems, visible pull requests, and distributed peer review.

“Governance within these communities was frequently described as emerging organically from transparent collaborative practices, including version control systems, visible pull requests, and distributed peer review (Dabbish et al. 2012; Markus 2007).”

• cites Dabbish et al. 2012Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Markus 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Transparency, initially framed as a mechanism for democratic oversight and epistemic accountability, was gradually institutionalized as documentation and compliance procedures.

“Transparency, initially framed as a mechanism for democratic oversight and epistemic accountability, was gradually institutionalized as documentation and compliance procedures.”

These reinterpretations show how debates over open-source AI governance have been reshaped by the practical challenges of regulating decentralized technological systems.

“They show how debates over open-source AI governance have been reshaped by the practical challenges of regulating decentralized technological systems.”

Transparency within AI governance evolved from a political value into a procedural instrument.

“In this sense, transparency within AI governance evolved from a political value into a procedural instrument.”

The normative transformations of openness, transparency, and accountability generate tensions within the institutional organization of open-source AI ecosystems.

“As values such as openness, transparency, and accountability are translated into technical and administrative practices, they generate tensions within the institutional organization of open-source AI ecosystems.”

When applied to AI systems, openness was frequently understood as the accessibility of model artifacts-such as source code, model weights, or training procedures-rather than as a broader commitment to participatory governance or institutional accountability.

“When applied to AI systems, openness was frequently understood as the accessibility of model artifacts—such as source code, model weights, or training procedures—rather than as a broader commitment to participatory governance or institutional accountability.”

Openness, originally associated with collective participation and shared agency in open-source communities, increasingly became equated with the technical visibility of model artifacts.

“Openness, originally associated with collective participation and shared agency in open-source communities, increasingly became equated with the technical visibility of model artifacts.”

In influential studies of OSS communities, openness was operationalized through practices such as code visibility, peer production, and participatory maintenance.

“In influential studies of OSS communities, openness was operationalized through practices such as code visibility, peer production, and participatory maintenance (Crowston, Howison, & Annabi 2007; de Laat 2007).”

• cites Crowston, Howison, & Annabi 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites de Laat 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Media reports document cases in which companies release partially open models or datasets under the banner of open source while simultaneously disclaiming liability for downstream harms.

“Media reports document cases in which companies release partially open models or datasets under the banner of open source while simultaneously disclaiming liability for downstream harms (Wired 2023; Financial Times 2024).”

• cites Wired 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Financial Times 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Accountability, long understood as a coherent moral obligation tied to identifiable actors, became diffused across complex networks of contributors, platforms, and downstream users.

“Accountability, long understood as a coherent moral obligation tied to identifiable actors, became diffused across complex networks of contributors, platforms, and downstream users.”

OSS governance structures often developed through informal community norms rather than through explicitly articulated regulatory principles or normative frameworks.

“As O’Mahony and Ferraro (2017) demonstrate, OSS governance structures often developed through informal community norms rather than through explicitly articulated regulatory principles or normative frameworks.”

• cites O’Mahony and Ferraro 2017Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Governance challenges arise not only from the technical properties of AI systems, but also from the institutional arrangements through which access, accountability, and authority are organized in open technological ecosystems.

“Governance challenges arise not only from the technical properties of AI systems, but also from the institutional arrangements through which access, accountability, and authority are organized in open technological ecosystems (Shevlane & Dafoe 2020).”

• cites Shevlane & Dafoe 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Open-source AI systems enable a broader range of actors to examine, adapt, and deploy model artifacts compared to proprietary AI systems.

“In contrast to proprietary AI systems, whose internal architectures and training resources are typically restricted for reasons of intellectual property, security, or competitive advantage, open-source AI enables a broader range of actors to examine, adapt, and deploy model artifacts (Bommasani et al. 2021; Maas & Villalobos 2023).”

• cites Bommasani et al. 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Maas & Villalobos 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

As openness, transparency, and accountability migrated from software communities into large-scale AI ecosystems, their meanings shifted significantly.

“Yet as these values migrated from software communities into large-scale AI ecosystems, their meanings shifted significantly.”

When algorithmic systems are developed through distributed collaboration and open dissemination, existing anti-discrimination law and liability frameworks struggle to determine who should be held accountable for biased outcomes, misuse, or security failures.

“When algorithmic systems are developed through distributed collaboration and open dissemination, these frameworks struggle to determine who should be held accountable for biased outcomes, misuse, or security failures.”

Open-source AI governance is characterized by a structural tension between limiting power and facilitating participation.

“Study concludes that open-source AI governance is characterized by a structural tension between limiting power and facilitating participation.”

• Also stated in: § Introduction · plain

  “Finally, open-source AI governance is characterized by a constant structural tension between the need to constrain power and the need to enable participation.”

Documentation practices can produce a form of "procedural transparency," in which the presence of documentation signals compliance without necessarily improving interpretability, accountability, or meaningful oversight.

“As Johnson and Green (2022) observe, these practices can produce a form of “procedural transparency,” in which the presence of documentation signals compliance without necessarily improving interpretability, accountability, or meaningful oversight.”

• cites Johnson and Green 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Anti-discrimination law and liability regimes are typically designed to identify responsible entities within centralized systems of decision-making.

“As Lopez and Kim (2020) and Carter and White (2021) observe, anti-discrimination law and liability regimes are typically designed to identify responsible entities within centralized systems of decision-making.”

• cites Lopez and Kim (2020)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Carter and White (2021)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

In such cases, openness becomes a mechanism through which responsibility is redistributed across developer communities rather than clearly assigned to institutional actors.

“In such cases, openness becomes a mechanism through which responsibility is redistributed across developer communities rather than clearly assigned to institutional actors.”

Early scholarship on open-source software tended to conceptualize openness primarily as a technical property rather than as a political or ethical ideal.

“Early scholarship on open-source software (OSS) tended to conceptualize openness primarily as a technical property rather than as a political or ethical ideal.”

As AI models became larger, more complex, and more widely distributed, earlier normative commitments were reformulated into technical and administrative practices that institutions could more easily operationalize.

“As AI models became larger, more complex, and more widely distributed, earlier normative commitments were reformulated into technical and administrative practices that institutions could more easily operationalize.”

Governance mechanisms designed for centralized technology firms or clearly identifiable developers struggle in ecosystems where models are continuously modified, redistributed, and repurposed across organizational and national boundaries.

“Governance mechanisms designed for centralized technology firms or clearly identifiable developers struggle in ecosystems where models are continuously modified, redistributed, and repurposed across organizational and national boundaries.”

The diffusion of authority within open ecosystems complicates traditional models of accountability and regulatory control.

“Yet the diffusion of authority within open ecosystems simultaneously complicates traditional models of accountability and regulatory control.”

Transparency increasingly became operationalized as compliance work: the structured production of documentation designed to make algorithmic systems inspectable by regulators, auditors, and other institutional actors.

“Through these developments, transparency increasingly became operationalized as compliance work: the structured production of documentation designed to make algorithmic systems inspectable by regulators, auditors, and other institutional actors.”

As AI systems become more visible, they do not necessarily become more intelligible.

“as systems become more visible, they do not necessarily become more intelligible”

Document-based analysis cannot fully evaluate the effectiveness of policy enforcement.

“As a result, document-based analysis cannot fully evaluate the effectiveness of policy enforcement.”

The tensions in open-source AI governance do not resolve at the level of conceptual definition but reappear as practical challenges concerning how authority is exercised, how responsibility is assigned, and how participation is structured in open technological ecosystems.

“These tensions do not resolve at the level of conceptual definition, but instead reappear as practical challenges concerning how authority is exercised, how responsibility is assigned, and how participation is structured in open technological ecosystems.”

Open technological ecosystems often lack clearly defined governance institutions capable of coordinating oversight and accountability.

“Scholars studying digital infrastructures therefore emphasize that open technological ecosystems often lack clearly defined governance institutions capable of coordinating oversight and accountability (O’Mahony & Ferraro 2017; Scott et al. 2023).”

• cites O’Mahony & Ferraro 2017Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Scott et al. 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Authority in China is embedded in a pre-existing administrative order rather than newly constructed or institutionally differentiated.

“In this sense, authority is neither newly constructed nor institutionally differentiated; it is embedded.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Technological openness can coexist with significant capability disparities.

“Development research and AI ethics scholarship therefore emphasize that technological openness can coexist with significant capability disparities (UNDP 2020; Pratti & Graves 2025).”

• cites UNDP 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Pratti & Graves 2025Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Values such as openness, transparency, and accountability remain central to discussions of digital technologies, but their practical meaning has been reshaped as they are translated into increasingly complex sociotechnical systems.

“Values such as openness, transparency, and accountability remain central to discussions of digital technologies, yet their practical meaning has been reshaped as they are translated into increasingly complex sociotechnical systems.”

As development of open-source AI becomes decentralized, conventional mechanisms for assigning liability and authority are destabilized.

“as development becomes decentralized, conventional mechanisms for assigning liability and authority are destabilized.”

Actors with greater access to infrastructure and data may benefit disproportionately from open technological ecosystems.

“As a result, actors with greater access to infrastructure and data may benefit disproportionately from open technological ecosystems (Caton & Haas 2024).”

• cites Caton & Haas 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The rules governing AI development, deployment, and oversight shape the distribution of rights, opportunities, and risks across social groups.

“The rules governing AI development, deployment, and oversight shape the distribution of rights, opportunities, and risks across social groups.”

Open-source AI governance is characterized not simply by competing values, but by the difficulty of institutionalizing them simultaneously within coherent regulatory structures.

“Open-source AI governance is therefore characterized not simply by the presence of competing values, but by the difficulty of institutionalizing them simultaneously within coherent regulatory structures.”

Even when systems are formally open, the benefits of technological innovation may remain concentrated among actors with the resources needed to use and adapt them effectively.

“Even when systems are formally open, the benefits of technological innovation may remain concentrated among actors with the resources needed to use and adapt them effectively.”

Open-source AI systems frequently emerge from networks of contributors who may not share formal institutional relationships.

“Research on dataset provenance and collaborative model development demonstrates that open-source AI systems frequently emerge from networks of contributors who may not share formal institutional relationships (Choksi et al. 2024).”

• cites Choksi et al. 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Open-source AI complicates the model of attributing responsibility for technological harms to identifiable actors because development often involves distributed contributions from multiple individuals and organizations.

“Open-source AI complicates this model because 21 development often involves distributed contributions from multiple individuals and organizations.”

Accountability shifts from a coherent institutional obligation to a fragmented network of partial responsibilities.

“Taken together, these developments reveal a third normative transformation in the evolution of open-source governance: accountability shifts from a coherent institutional obligation to a fragmented network of partial responsibilities.”

As participation in open-source AI broadens, responsibility becomes more diffuse.

“As participation broadens, responsibility becomes more diffuse”

Access to open-source AI does not necessarily translate into equal participation, because developing and deploying advanced AI systems requires substantial computational resources, technical expertise, and institutional support.

“However, scholars increasingly note that access does not necessarily translate 22 into equal participation. Developing and deploying advanced AI systems requires substantial computational resources, technical expertise, and institutional support.”

The same features of open-source AI that expand access and enable distributed innovation simultaneously complicate the institutional conditions required for oversight and accountability.

“The same features that expand access and enable distributed innovation simultaneously complicate the institutional conditions required for oversight and accountability.”

When AI models are developed collaboratively and reused across multiple contexts, existing liability frameworks struggle to determine who should be held accountable for biased outcomes, misuse, or security failures.

“When AI models are developed collaboratively and reused across multiple contexts, these frameworks struggle to determine who should be held accountable for biased outcomes, misuse, or security failures.”

Debates over open-source AI governance cannot be resolved through technical standards or regulatory mechanisms alone.

“The governance challenges identified in the previous section reveal that debates over open-source AI cannot be resolved through technical standards or regulatory mechanisms alone.”

Policy analyses highlight an emerging "accountability gap" in open AI ecosystems.

“Policy analyses therefore highlight an emerging “accountability gap” in open AI ecosystems (Scott et al. 2023).”

• cites Scott et al. 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Open-source AI complicates the traditional regulatory assumption because development frequently occurs across decentralized networks of contributors and platforms.

“Open-source AI complicates this assumption because development frequently occurs across decentralized networks of contributors and platforms.”

Rather than operating as stable normative commitments, openness, transparency, and accountability are progressively rearticulated through technical and administrative practices, altering how they function within governance frameworks.

“Rather than operating as stable normative commitments, these values are progressively rearticulated through technical and administrative practices, altering how they function within governance frameworks.”

Under the EU AI Act, providers are required to ensure that systems meet regulatory requirements prior to being placed on the market, including preparation of technical documentation and conformity assessment procedures.

“Providers are required to ensure that systems meet regulatory requirements prior to being placed on the market, including the preparation of technical documentation and conformity assessment procedures.”

• cites EU AI ActNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Executive Order 14110 emphasizes that actors involved in the development and deployment of AI systems should be held accountable to standards designed to prevent harm, discrimination, and misuse.

“Executive Order 14110 emphasizes that actors involved in the development and deployment of AI systems should be held accountable to standards designed to prevent harm, discrimination, and misuse (EO 14110).”

• cites EO 14110Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The Chinese model treats transparency as a condition of governability, embedded within supervision and administrative control.

“Transparency is therefore embedded within supervision.”

• Also stated in: § 4.2 Transparency · plain

  “In this sense, the Chinese model treats transparency as a condition of governability: AI systems must be rendered visible in ways that allow them to be incorporated into existing structures of administrative control.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

In China, participation is not structured solely by access to technological resources, but also by the ability to operate within a regulated administrative environment.

“Participation is therefore not structured solely by access to technological resources, but also by the ability to operate within a regulated administrative environment.”

• AssumptionThe authors state this as an assumption: Actors participating in AI development possess the organizational capacity to implement compliance systems, monitor outputs, and manage risks associated with generated content.

Transparency in AI governance is not a single institutional practice but a set of arrangements through which systems are made legible for distinct regulatory purposes.

“These differences demonstrate that transparency in AI governance is not a single institutional practice but a set of arrangements through which systems are made legible for distinct regulatory purposes.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Under the EU AI Act, deployers are subject to additional obligations concerning the use, monitoring, and, in some cases, assessment of the impacts of AI systems.

“Deployers are subject to additional obligations concerning the use, monitoring, and, in some cases, assessment of the impacts of AI systems.”

• cites EU AI ActNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Executive Order 14110 frames AI governance as a collaborative undertaking involving government, industry, academia, and civil society.

“Executive Order 14110 frames AI governance as a collaborative undertaking involving government, industry, academia, and civil society (EO 14110).”

• cites EO 14110Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Article 21 of the Interim Measures provides that violations may result in administrative penalties imposed by competent authorities, including warnings, orders to rectify, suspension, or termination of services.

“Article 21 further provides that violations of the regulatory framework may result in administrative penalties imposed by competent authorities, including warnings, orders to rectify, suspension, or termination of services (Art. 21).”

• cites Art. 21Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Article 15 of the Interim Measures requires the establishment of complaint and reporting mechanisms.

“Article 15, which requires the establishment of complaint and 51 reporting mechanisms (Art. 15).”

• cites Art. 15Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The EU model embeds transparency directly within the legal structure governing AI systems, linking it to compliance, enforcement, and the functioning of the internal market.

“Rather than relying on administrative supervision or distributed reporting alone, the EU model embeds transparency directly within the legal structure governing AI systems, linking it to compliance, enforcement, and the functioning of the internal market.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

China's accountability for AI is organized around direct provider responsibility combined with administrative enforcement, producing a centralized accountability model.

“In China, accountability is organized around direct provider responsibility combined with administrative enforcement.”

• Also stated in: § 4.3 Accountability · boosted

  “As a result, governance operates through a model of centralized accountability in which responsibility is clearly attached to identifiable actors and enforced through administrative mechanisms.”

• Also stated in: § 4.3 Accountability · plain

  “China concentrates responsibility at the level of service providers and enforces it through administrative authority.”

U.S. authority is not unified but is a coordinated system of overlapping institutional competencies, dispersed across a network of actors.

“What emerges from this arrangement is not a unified authority structure but a coordinated system of overlapping institutional competencies.”

• Also stated in: § 4.1 Authority · plain

  “Authority is therefore not concentrated in a single institutional site, nor fully codified through a unified legal framework. It is instead dispersed across a network of actors whose collective activity constitutes governance.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

Executive Order 14110 does not establish a unified accountability regime specific to artificial intelligence.

“However, the Order does not establish a unified accountability regime specific to artificial intelligence.”

• cites EO 14110Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Under Executive Order 14110 Section 4, developers may be required to provide information regarding model development and risks to federal authorities.

“Developers may be required to provide information regarding model development and risks to federal authorities (Sec. 4)”

• cites Sec. 4Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Executive Order 14110 supports the development of research infrastructure, workforce training, and technological resources, thereby expanding the conditions under which actors may participate in AI development.

“It supports the development of research infrastructure, workforce training, and technological resources, thereby expanding the conditions under which actors may participate in AI development.”

• cites EO 14110Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Within the EU framework, transparency is a legally defined condition of market participation; documentation, traceability, and disclosure are codified as obligations that must be satisfied for AI systems to be developed, deployed, and placed on the market within the Union.

“Within this framework, transparency is not simply a tool for regulatory oversight but a legally defined condition of market participation. Documentation, traceability, and disclosure are codified as obligations that must be satisfied in order for AI systems to be developed, deployed, and placed on the market within the Union.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The Interim Measures' provisions assume that actors participating in AI development possess the organizational capacity to implement compliance systems, monitor outputs, and manage risks associated with generated content.

“These provisions assume that actors participating in AI development possess the organizational capacity to implement compliance systems, monitor outputs, and manage risks associated with generated content.”

• AssumptionThe authors state this as an assumption: Actors participating in AI development possess the organizational capacity to implement compliance systems, monitor outputs, and manage risks associated with generated content.

Executive Order 14110 establishes reporting and evaluation requirements for certain advanced AI systems, requiring developers to provide information regarding training processes, model capabilities, and associated risks.

“The Order establishes reporting and evaluation requirements for certain advanced AI systems, requiring developers to provide information regarding training processes, model capabilities, and associated risks (EO 14110, Sec. 4).”

• cites EO 14110, Sec. 4Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Article 9 of the Interim Measures requires service providers to assume responsibilities associated with network information content production, extending existing content governance obligations to generative AI systems.

“Article 9 requires service providers to assume the responsibilities associated with network information content production (Interim Measures, Art. 9), thereby extending existing content governance obligations to generative AI systems.”

• cites Interim Measures, Art. 9Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The EU structures authority through legislative codification, defining authority in advance through binding legal rules that specify the scope of regulation and the actors responsible for its implementation.

“The European Union adopts a third approach, structuring authority through legislative codification.”

• Also stated in: § 4.1 Authority · plain

  “In contrast to both administrative embedding and executive coordination, the EU framework defines authority in advance through binding legal rules that specify both the scope of regulation and the actors responsible for its implementation.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

The Interim Measures for the Management of Generative Artificial Intelligence Services specify that providers bear primary responsibility for the content generated by their systems.

“The Interim Measures for the Management of Generative Artificial Intelligence Services specify that providers bear primary responsibility for the content generated by their systems.”

• cites Interim Measures, Art. 9Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

China, the United States, and the European Union illustrate distinct institutional configurations of accountability.

“Taken together, these three approaches illustrate distinct institutional configurations of accountability.”

• Also stated in: § 4.3 Accountability · plain

  “These differences highlight how governance regimes vary not only in who exercises authority, but also in how responsibility for AI systems is assigned and operationalized.”

In the U.S. framework, participation is not primarily structured through ex ante authorization or centralized regulatory approval, but is shaped by access to technological infrastructure, expertise, and institutional resources.

“Within this framework, participation is not primarily structured through ex ante authorization or centralized regulatory approval.”

• Also stated in: § 4.4 Participation · plain

  “Instead, it is shaped by access to technological infrastructure, expertise, and institutional resources.”

Article 14 of the Interim Measures obliges providers to take corrective measures when unlawful content is generated, including stopping generation or transmission and implementing subsequent remediation.

“Article 14, which obliges providers to take corrective measures when unlawful content is generated, including stopping generation or transmission and implementing subsequent remediation (Art. 14)”

• cites Art. 14Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

U.S. transparency emerges from a combination of reporting requirements, technical evaluation processes, and standards-setting activities carried out by institutions such as NIST, producing distributed visibility.

“Transparency in this framework is not organized as a single, unified obligation. Instead, it emerges from a combination of reporting requirements, technical evaluation processes, and standards-setting activities carried out by institutions such as NIST.”

• Also stated in: § 4.2 Transparency · plain

  “These mechanisms produce a form of distributed visibility in which different aspects of AI systems become legible to different institutional actors.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

AI governance is not organized according to a single institutional model but reflects divergent approaches to structuring regulatory power within technologically mediated systems.

“These differences demonstrate that AI governance is not organized according to a single institutional model but instead reflects divergent approaches to structuring regulatory power within technologically mediated systems.”

• AssumptionThe authors state this as an assumption: Policy texts function as institutional artifacts rather than merely statements of political intention.

In the United States, participation is shaped less by formal restriction than by differential access to resources and the capacity to meet evolving regulatory requirements.

“Participation is 56 therefore shaped less by formal restriction than by differential access to resources and the capacity to meet evolving regulatory requirements.”

In the United States, responsibility for AI accountability is not concentrated in a single institutional site.

“Responsibility is therefore not concentrated in a single institutional site.”

The United States structures AI accountability in a dispersed manner, with responsibility distributed across multiple actors and overlapping institutional domains rather than concentrated in a single site.

“The United States structures accountability in a more dispersed manner.”

• Also stated in: § 4.3 Accountability · plain

  “As a result, governance operates through a model of distributed accountability in which responsibility is allocated across overlapping institutional domains.”

• Also stated in: § 4.3 Accountability · plain

  “The United States distributes responsibility across multiple actors and relies on existing regulatory institutions to oversee compliance.”

Public justification cannot be reduced to the articulation of reasons; it depends on the conditions under which those reasons can be contested.

“Rawls’s idea of public reason requires that political decisions be defensible through reasons accessible to free and equal citizens (Rawls 2001).”

• Also stated in: § 5.11 Rawlsian Legitimacy in the Age of Algorithmic Governance · plain

  “Public justification therefore cannot be reduced to the articulation of reasons; it depends on the conditions under which those reasons can be contested.”

• cites Rawls 2001Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

In the American context, fragmentation does not produce silence; it produces multiplicity.

“In the American context, fragmentation does not produce silence; it produces multiplicity.”

Participation depends upon institutional conversion factors: access to data and technical expertise, regulatory literacy, compliance capacity, and political security.

“Participation depends upon institutional “conversion factors”: access to data and technical expertise, regulatory literacy, compliance capacity, and political security.”

Applying Rawlsian criteria to governance systems that do not fully share these institutional premises requires analytical caution.

“Applying Rawlsian criteria to governance systems that do not fully share these institutional premises therefore requires analytical caution.”

China's generative AI regime satisfies the transparency criterion in a straightforward and robust manner.

“In this respect, the regime satisfies the transparency criterion in a straightforward and even robust manner.”

Under Rawls's criteria, China's regime generates not opacity but insulation.

“Under Rawls’s criteria, this generates not opacity but insulation.”

When review is structurally internal to enforcement, contestation depends on the responsiveness of the same authority that exercised discretion.

“When review is structurally internal to enforcement, contestation depends on the responsiveness of the same authority that exercised discretion.”

In the Chinese regime, a systematic outward-facing practice of reason-giving that subjects regulatory determinations to publicly accessible scrutiny is less institutionally prominent.

“What is less institutionally prominent is a systematic outward-facing practice of reason-giving that subjects regulatory determinations to publicly accessible scrutiny.”

In China's generative AI regime, regulatory authority is explicitly situated within designated administrative organs embedded in a vertical hierarchy connecting rule articulation, inspection, and enforcement.

“Rather, regulatory authority is explicitly situated within designated administrative organs, embedded in a vertical hierarchy that connects rule articulation, inspection, and enforcement.”

In the EU, participation is structured not only by formal access to the market, but also by the capacity to satisfy legal and procedural requirements associated with responsible AI development.

“Participation is therefore structured not only by formal access to the market, but also by the capacity to satisfy legal and procedural requirements associated with responsible AI development.”

When justificatory structures are predominantly administrative rather than publicly dialogical, the risk is insulation, not necessarily abuse.

“When justificatory structures are predominantly administrative rather than publicly dialogical, the risk is not necessarily abuse, but insulation.”

In the U.S., revisability depends on invoking general administrative or sector-specific legal mechanisms rather than engaging a single, structured oversight body.

“Revisability depends on invoking general administrative or sector-specific legal mechanisms rather than engaging a single, structured oversight body.”

The U.S. AI Safety Institute, established within NIST, contributes to standardization and evaluation of advanced AI systems through testing protocols and safety benchmarks.

“The U.S. AI Safety Institute, established within the National Institute of Standards and Technology, contributes to the standardization and evaluation of advanced AI systems through testing protocols and safety benchmarks (National Institute of Standards and Technology 2024).”

• cites National Institute of Standards and Technology 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The European Union constructs a pattern of juridified and structured AI governance, with authority, visibility, responsibility, and participation systematically organized through law and implemented through a multi-level regulatory architecture.

“The European Union constructs a third pattern characterized by juridified and structured governance.”

• Also stated in: § 4.5 Institutional Patterns of AI Governance · plain

  “These elements combine to form a governance model in which authority, visibility, responsibility, and participation are systematically organized through law and implemented through a multi-level regulatory architecture.”

Welfare-oriented reasoning helps explain why different institutional arrangements are adopted under similar technological pressures.

“As Floridi et al. (2018) observe, policy discussions surrounding artificial intelligence are often structured around innovation and economic development goals rather than explicit theories of political legitimacy.”

• Also stated in: § 5.0 Normative Authority and Method of Adjudication · plain

  “These arguments reflect a form of welfare-oriented reasoning that helps explain why different institutional arrangements are adopted under similar technological pressures.”

• cites Floridi et al. (2018)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionRests on an assumption: Governments frequently justify regulatory choices by appealing to economic growth, innovation capacity, global competitiveness, or systemic stability.
• AssumptionRests on an assumption: Different institutional arrangements are adopted under similar technological pressures.

Multiplicity can dilute clarity.

“Yet multiplicity can dilute clarity.”

This dispersion can complicate the identification of a final adjudicatory locus.

“This dispersion can complicate the identification of a final adjudicatory locus.”

The United States organizes AI governance through a model of distributed and coordination-based governance, with authority, oversight, and responsibility distributed across a network of actors.

“The United States organizes these dimensions through a model of distributed and coordination-based governance.”

• Also stated in: § 4.5 Institutional Patterns of AI Governance · plain

  “In combination, these features produce a governance model in which institutional authority, oversight, and responsibility are distributed across a network of actors, and participation is shaped by access to resources and institutional capacity rather than centralized authorization.”

The result of U.S. AI governance is not institutional opacity but institutional pluralization.

“The result is not institutional opacity, but institutional pluralization.”

In the U.S., when harms occur, redress may depend on which sector is implicated-employment, finance, healthcare, consumer protection, or national security.

“When harms occur, redress may depend on which sector is implicated—employment, finance, healthcare, consumer protection, or national security.”

The U.S. AI Safety Institute does not constitute a unified mechanism through which contestability is institutionally structured.

“It therefore does not constitute a unified mechanism through which contestability is institutionally structured.”

Rawls developed his account of political legitimacy primarily within a constitutional liberal society characterized by free and equal citizens cooperating under publicly justified institutions.

“Rawls developed his account of political legitimacy primarily within the context of a constitutional liberal society characterized by free and equal citizens cooperating under publicly justified institutions (Rawls 2001).”

• cites Rawls 2001Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The same vertical structure that generates authority in China also mediates its reconsideration.

“The same vertical structure that generates authority also mediates its reconsideration.”

In China's generative AI regime, rule-making, supervision, and review remain closely integrated within executive institutions.

“In China’s generative AI regime, rule-making, supervision, and review remain closely integrated within executive institutions.”

The Chinese generative AI regime does not suffer from the kind of institutional ambiguity that often characterizes technologically complex governance environments.

“the Chinese generative AI regime does not suffer from the kind of institutional ambiguity that often characterizes technologically complex governance environments.”

This integration reduces structural separation between discretion and oversight.

“Yet it reduces structural separation between discretion and oversight.”

Responsibility for AI oversight in the U.S. is dispersed across agencies whose mandates intersect but do not fully converge.

“Responsibility for AI oversight is dispersed across agencies whose mandates intersect but do not fully converge.”

Rawlsian procedural justice provides a coherent framework for evaluating whether AI governance arrangements preserve the conditions of legitimate authority.

“In this way, Rawlsian procedural justice provides a coherent framework for evaluating whether AI governance arrangements preserve the conditions of legitimate authority.”

China, the United States, and the European Union illustrate distinct institutional configurations of participation.

“Taken together, these three approaches illustrate distinct institutional configurations of participation.”

• Also stated in: § 4.4 Participation · boosted

  “These differences demonstrate that participation in AI governance is not determined solely by technological openness, but by the institutional conditions that shape how actors enter and operate within the ecosystem.”

Authority may remain efficient, decisive, and coherent while becoming structurally insulated from reciprocal challenge.

“Authority may remain efficient, decisive, and coherent while becoming structurally insulated from reciprocal challenge.”

The EU AI Act requires Member States to establish AI regulatory sandboxes that allow developers to test innovative systems under supervisory conditions.

“It requires Member States to establish AI regulatory sandboxes that allow developers to test innovative systems under supervisory conditions.”

• cites EU AI ActNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

In the EU, entry into the AI ecosystem is conditioned by compliance with legal requirements that govern development, deployment, and market access.

“Entry into the AI ecosystem is therefore conditioned by compliance with legal requirements that govern development, deployment, and market access.”

The institutional authorship of governance decisions in China's generative AI regime is readily identifiable.

“The institutional authorship of governance decisions is therefore readily identifiable.”

Concentration becomes normatively problematic when it is not accompanied by institutionalized constraint.

“Concentration becomes normatively problematic when it is not accompanied by institutionalized constraint.”

In the Chinese regime, justificatory processes are primarily oriented toward compliance within the administrative hierarchy.

“In the Chinese regime, justificatory processes are primarily oriented toward compliance within the administrative hierarchy.”

In the Chinese regime, regulatory emphasis centers on providers' obligations to satisfy supervisory expectations.

“Regulatory emphasis centers on providers’ obligations to satisfy supervisory expectations.”

In China's case, responsibility is not indeterminate; it is centralized.

“In China’s case, responsibility is not indeterminate; it is centralized.”

In the U.S. model, transparency is pursued through extensive reporting obligations, agency inventories, and publicly available guidance documents.

“On one level, transparency is pursued through extensive reporting obligations, agency inventories, and publicly available guidance documents.”

In the Chinese regime, the justificatory flow is predominantly vertical (provider to regulator) rather than horizontal (authority to citizen).

“The justificatory flow is predominantly vertical—provider to regulator—rather than horizontal— authority to citizen.”

Technical opacity may generate an accountability gap in which harms diffuse across actors while institutional power remains insulated from meaningful review.

“As scholars of AI accountability have argued, technical opacity may generate an “accountability gap,” in which harms diffuse across actors while institutional power remains insulated from meaningful review (Raji et al. 2020; Novelli et al. 2022).”

• cites Raji et al. 2020; Novelli et al. 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

In China's generative AI framework, complaint and reporting mechanisms are embedded within the same administrative hierarchy that promulgates and enforces standards.

“However, these mechanisms are embedded within the same administrative hierarchy that promulgates and enforces standards.”

The distributed nature of U.S. governance introduces ambiguity in identifying who exercises binding authority.

“Here the distributed nature of U.S. governance introduces ambiguity.”

Actors capable of sustained engagement with supervisory authorities exercise greater influence over evolving standards than peripheral participants in China.

“Actors capable of sustained engagement with supervisory authorities exercise greater influence over evolving standards than peripheral participants.”

In the American AI ecosystem, power is shaped by concentrated private investment and technological infrastructure.

“In addition to fragmented public authority, power in the American AI ecosystem is also shaped by concentrated private investment and technological infrastructure.”

Training state-of-the-art systems requires enormous computational expenditure, often available only to a small number of technology firms and well-resourced research institutions.

“Empirical analyses demonstrate that training state-of-the-art systems requires enormous computational expenditure, often available only to a small number of technology firms and well-resourced research institutions (OECD 2022; Stanford AI Index 2024).”

• cites OECD 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Stanford AI Index 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The American approach minimizes formal restrictions on model development and release.

“At the level of entry conditions, the American approach minimizes formal restrictions on model development and release.”

• AssumptionRests on an assumption: The United States federal government has not imposed blanket pre-authorization requirements for generative model deployment.

Threshold design in the Chinese context operates through formal authorization and continuous oversight requirements.

“Threshold design in the Chinese context operates through formal authorization and continuous oversight requirements.”

Capability in China is mediated not through market-based infrastructural concentration alone, but through the ability to align with administrative oversight.

“Capability is mediated not through market-based infrastructural concentration alone, but through the ability to align with administrative oversight.”

Formal openness can coexist with material exclusion; unequal access to computational resources, technical expertise, and institutional infrastructures systematically constrains who can meaningfully engage in AI development and governance.

“Contemporary scholarship on AI access and participation demonstrates how formal openness can coexist with material exclusion, as unequal access to computational resources, technical expertise, and institutional infrastructures systematically constrains who can meaningfully engage in AI development and governance (Gabriel 2022; Prabhakaran et al. 2022).”

• cites Gabriel 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Prabhakaran et al. 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

U.S. power is diffused across agencies and levels of government.

“Power is diffused across agencies and levels of government.”

Under heterogeneous conversion environments, identical formal permission does not guarantee equal opportunity.

“Under heterogeneous conversion environments, identical formal permission does not guarantee equal opportunity.”

• cites Robeyns 2005Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Effective participation in China requires the capacity to satisfy documentation expectations, security audits, and ongoing reporting obligations.

“Effective participation thus requires the capacity to satisfy documentation 85 expectations, security audits, and ongoing reporting obligations.”

The US governance environment assumes, rather than systematically cultivates, baseline technical and organizational capacity.

“The governance environment further assumes, rather than systematically cultivates, baseline technical and organizational capacity.”

This outward-facing orientation aligns, at least rhetorically, with Rawls's requirement that coercive power be publicly justifiable.

“This outward-facing orientation aligns, at least rhetorically, with Rawls’s requirement that coercive power be publicly justifiable (Rawls 2001).”

• cites Rawls 2001Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Regulatory compliance in China often demands integration of state directives into corporate governance structures.

“Research on platform governance in China suggests that regulatory compliance often demands integration of state directives into corporate governance structures (Sacks 2021; Creemers 2022).”

• cites Sacks 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Creemers 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

In the EU, authority is legislatively articulated; roles of providers, deployers, conformity assessment bodies, national supervisory authorities, and EU-level institutions are specified in binding legal text.

“Authority is not merely exercised; it is legislatively articulated. The roles of providers, deployers, conformity assessment bodies, national supervisory authorities, and EU-level institutions are specified in binding legal text.”

Frontier model development is increasingly dependent upon access to large-scale computational infrastructure, advanced engineering teams, and high-quality proprietary datasets.

“Frontier model development is increasingly dependent upon access to large-scale computational infrastructure, advanced engineering teams, and high-quality proprietary datasets.”

In China, influence tends to be mediated through actors already embedded within state-recognized networks.

“While public consultation mechanisms exist, influence tends to be mediated through actors already embedded within state-recognized networks (Zhang 2023).”

• cites Zhang 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionRests on an assumption: Public consultation mechanisms exist in China.

Dispersion introduces a different structural vulnerability: oversight responsibilities may become diluted.

“However, dispersion introduces a different structural vulnerability. When authority is spread across multiple institutions, oversight responsibilities may become diluted.”

China conditions deployment on formal alignment with state-defined regulatory criteria.

“China conditions deployment on formal alignment with state-defined regulatory criteria.”

Compliance obligations in China presume the existence of internal content moderation systems, data governance controls, and technical safeguards capable of filtering prohibited outputs.

“Compliance obligations presume the existence of internal content moderation systems, data governance controls, and technical safeguards capable of filtering prohibited outputs.”

In the U.S. model, justification risks becoming informational rather than normatively binding.

“Justification, therefore, risks becoming informational rather than normatively binding.”

In the Chinese case, the structural risk lies not in infrastructural scarcity or compliance overload alone, but in authorization-mediated differentiation.

“In the Chinese case, the structural risk lies 87 not in infrastructural scarcity or compliance overload alone, but in authorization-mediated differentiation.”

The American approach places significant emphasis on public reporting and disclosure.

“The American approach places significant emphasis on public reporting and disclosure.”

Development of frontier AI models increasingly depends on large-scale capital, proprietary datasets, cloud computing infrastructure, and specialized technical expertise, creating structural advantages for a small number of firms.

“The development of frontier AI models increasingly depends on large-scale capital, proprietary datasets, cloud computing infrastructure, and specialized technical expertise, creating structural advantages for a small number of firms.”

It introduces a structural concern that authority over technological development may be exercised through market structures not themselves subject to the same forms of public justification and institutional review.

“However, it introduces a different structural concern: the possibility that authority over technological development may be exercised through market structures that are not themselves subject to the same forms of public justification and institutional review.”

Smaller or decentralized actors in China may possess technical competence yet lack the organizational infrastructure necessary to navigate registration procedures and sustain compliance.

“Smaller or decentralized actors may possess technical competence yet lack the organizational infrastructure necessary to navigate registration procedures and sustain compliance.”

In the EU, institutional authorship is defined through statutory allocation rather than inferred from administrative practice.

“Institutional authorship is therefore not inferred from administrative practice but defined through statutory allocation.”

Meaningful auditing and documentation require sustained technical expertise, cross-disciplinary teams, and institutionalized oversight structures.

“Research on algorithmic accountability indicates that meaningful auditing and documentation require sustained technical expertise, cross-disciplinary teams, and institutionalized oversight structures (Raji et al. 2020; Novelli et al. 89 2022).”

• cites Raji et al. 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Novelli et al. 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

On concentration of authority, the United States appears normatively advantaged.

“Finally, on the question of concentration of authority, the United States appears normatively advantaged.”

Participation in rule-shaping processes in China reflects structured coordination rather than decentralized pluralism.

“Participation in rule-shaping processes likewise reflects structured coordination rather than decentralized pluralism.”

Governance in the United States presumes a degree of organizational sophistication that is unevenly distributed across participants.

“Governance thus presumes a degree of organizational sophistication that is unevenly distributed across participants.”

In the U.S. model, answerability may require navigating dispersed legal pathways rather than engaging a singular review structure.

“Answerability, however, may require navigating dispersed legal pathways rather than engaging a singular review structure.”

In the U.S. model, traceability is often achievable through transparency measures.

“In the U.S. model, traceability is often achievable through transparency measures.”

China's governance architecture for generative and open-source AI organizes participation through a structured system of authorization, registration, and ongoing supervisory oversight.

“China’s governance architecture for generative and open-source AI organizes participation through a structured system of authorization, registration, and ongoing supervisory oversight.”

The AI Act embeds revisability within its compliance architecture.

“The AI Act embeds revisability within its compliance architecture.”

Transparency reports, guidance documents, and policy statements contribute to an ecosystem of visible reasoning in the U.S.

“Transparency reports, guidance documents, and policy statements contribute to an ecosystem of visible reasoning.”

No single centralized AI regulator in the U.S. consolidates rule-making, enforcement, and review.

“No single centralized AI regulator consolidates rule-making, enforcement, and review.”

In the U.S. model, contestability exists but is mediated through fragmentation.

“Contestability exists, but it is mediated through fragmentation.”

It embeds authority within publicly promulgated legal norms that are themselves subject to judicial interpretation.

“It embeds authority within publicly promulgated legal norms that are themselves subject to judicial interpretation.”

This concentration of market power complicates Rawlsian concerns about domination, as authority over technological development may be exercised not only by public institutions but also by private actors operating outside traditional structures of democratic accountability.

“This concentration of market power complicates Rawlsian concerns about domination, as authority over technological development may be exercised not only by public institutions but also by private actors operating outside traditional structures of democratic accountability.”

Juridification enhances examinability by translating governance into codified criteria rather than discretionary instruction.

“Juridification enhances that examinability by translating governance into codified criteria rather than discretionary instruction.”

From a Rawlsian standpoint, U.S. dispersion mitigates the risk of arbitrary domination associated with concentrated discretion.

“From a Rawlsian standpoint, this dispersion mitigates the risk of arbitrary domination associated with concentrated discretion.”

Influence within formally open structures correlates strongly with sustained technical contribution, infrastructural centrality, and reputational capital.

“However, empirical studies of open-source communities suggest that influence within formally open structures correlates strongly with sustained technical contribution, infrastructural centrality, and reputational capital (Crowston 2007; O’Mahony & Ferraro 2017).”

• cites Crowston 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites O’Mahony & Ferraro 2017Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Capability in China is channeled through administrative conformity rather than purely economic scale or regulatory literacy alone.

“Capability is therefore channeled through administrative conformity rather than purely economic scale or regulatory literacy alone.”

The EU's AI governance architecture takes a third path: juridification, understood as institutionalization of governance authority through binding legal rules embedded within a rule-of-law framework.

“the European Union’s AI governance architecture takes a third path: juridification, understood here as the institutionalization of governance authority through binding legal rules embedded within a rule-of-law framework.”

This form of transparency differs qualitatively from both administrative clarity and executive coordination.

“This form of transparency differs qualitatively from both administrative clarity and executive coordination.”

Compliance in China is often dynamic and iterative, requiring ongoing interaction between firms and supervisory bodies.

“Scholarship on regulatory governance in China highlights how compliance is often dynamic and iterative, requiring ongoing interaction between firms and supervisory bodies (Creemers 2022).”

• cites Creemers 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Public-facing deployment and commercial scaling in China are contingent upon institutional authorization.

“However, public-facing deployment and commercial scaling are contingent upon institutional authorization.”

The absence of regulatory barriers does not eliminate differentiation; it shifts the locus of differentiation from legal restriction to infrastructural concentration.

“In this context, the absence of regulatory barriers does not eliminate differentiation; it shifts the locus of differentiation from legal restriction to infrastructural concentration.”

Governments operationalize innovation and safety considerations through specific practices of measurement, prioritization, and policy translation.

“Governments do not merely assert that innovation or safety matters; they operationalize these considerations through specific 95 practices of measurement, prioritization, and policy translation.”

Governments rely on forms of welfare-oriented reasoning that assess technological systems in terms of expected economic benefits, innovation capacity, and systemic stability.

“Instead, they rely on forms of welfare-oriented reasoning that assess technological systems in terms of expected economic benefits, innovation capacity, and systemic stability.”

• AssumptionRests on an assumption: Governments do not design regulatory systems by applying theories of justice.

Formal openness alone is insufficient if access to computational infrastructure, technical expertise, and regulatory capacity remains unevenly distributed.

“Formal openness alone is insufficient if access to computational infrastructure, technical expertise, and regulatory capacity remains unevenly distributed.”

In the United States, welfare-oriented reasoning is strongly oriented toward innovation and economic growth.

“In the United States, welfare-oriented reasoning is strongly oriented toward innovation and economic growth.”

Formally inclusive systems often rely on tacit expectations of technical competence and organizational stability.

“Open-source governance scholarship has long demonstrated that formally inclusive systems often rely on tacit expectations of technical competence and organizational stability (Crowston 2007; O’Mahony & Ferraro 2017).”

• cites Crowston 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites O’Mahony & Ferraro 2017Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

As algorithmic systems increasingly mediate access to employment, information, economic opportunity, and public discourse, the institutions governing them become part of what Rawls describes as the basic structure of society.

“As algorithmic systems increasingly mediate access to employment, information, 103 economic opportunity, and public discourse, the institutions governing them become part of what Rawls describes as the basic structure of society.”

• cites RawlsNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Formal openness does not guarantee inclusive participation in AI ecosystems.

“Third, formal openness does not guarantee inclusive participation in AI ecosystems.”

Participation in commercially consequential or higher-risk domains in the EU is conditioned upon conformity assessments, documentation requirements, and ongoing monitoring duties.

“Yet participation in commercially consequential or higher-risk domains is conditioned upon conformity assessments, documentation requirements, and ongoing monitoring duties.”

Institutions designed to constrain power must simultaneously remain attentive to how structural rules shape unequal capacities for participation.

“institutions designed to constrain power must simultaneously remain attentive to how structural rules shape unequal capacities for participation (Claassen 2011; Floridi et al. 2018).”

• cites Claassen 2011Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Floridi et al. 2018Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The AI Act's risk-tiered structure explicitly differentiates obligations according to system impact, seeking to avoid uniform burdens across heterogeneous applications.

“The AI Act’s risk-tiered structure explicitly differentiates obligations according to system impact, seeking to avoid uniform burdens across heterogeneous applications.”

• cites AI ActNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Mechanisms designed to enhance accountability and constrain discretionary power may also generate administrative or infrastructural barriers to participation.

“These tensions reflect a broader institutional dilemma frequently identified in governance scholarship: mechanisms designed to enhance accountability and constrain discretionary power may also generate administrative or infrastructural barriers to participation (Novelli et al. 2022; Veale and Borgesius 2021).”

• cites Novelli et al. 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Veale and Borgesius 2021Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Institutional legitimacy also depends on the distribution of substantive opportunities within the system of cooperation.

“Capability analysis therefore emphasizes that institutional legitimacy also depends on the distribution of substantive opportunities within the system of cooperation.”

Advances in fairness evaluation, interpretability, and safety alignment increasingly demand resource-intensive processes.

“Advances in fairness evaluation, interpretability, and safety alignment increasingly demand resource-intensive processes (Hardt et al.; Molnar 2022).”

• cites Hardt et al.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Molnar 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Within the EU context, regulatory literacy itself becomes a decisive factor in shaping participation.

“Within the EU context, regulatory literacy itself becomes a decisive factor in shaping participation.”

The European Union embeds risk-based obligations directly into the conditions of participation.

“Rather than relying primarily on decentralized innovation norms or post hoc accountability, the EU embeds risk-based obligations directly into the conditions of participation.”

Influence in governance participation correlates with organizational resources and sustained institutional presence.

“Research on governance participation consistently finds that influence correlates with organizational resources and sustained institutional presence (Markus 2007).”

• cites Markus 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

None of the three governance models fully resolves the institutional tensions inherent in governing open technological ecosystems.

“None fully resolves the institutional tensions inherent in governing open technological ecosystems.”

Governance systems that appear formally open may still produce structured inequalities in participation.

“As a result, governance systems that appear formally open may 109 still produce structured inequalities in participation.”

Even risk-based obligations entail documentation, evaluation, and monitoring practices that demand significant institutional infrastructure.

“Nonetheless, even risk-based obligations entail documentation, evaluation, and monitoring practices that demand significant institutional infrastructure.”

Evaluating institutions solely through formal procedural criteria is limited: even when governance arrangements satisfy procedural requirements, actors may remain unable to participate meaningfully if they lack the resources, expertise, or institutional access needed to translate formal permissions into effective opportunities.

“Even when governance arrangements satisfy procedural requirements, actors may remain unable to participate meaningfully if they lack the resources, expertise, or institutional access necessary to translate formal permissions into effective opportunities (Sen 2009; Robeyns 2005).”

• cites Sen 2009Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Robeyns 2005Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

China's governance framework reflects another configuration of welfare-oriented reasoning, one that integrates technological development with state coordination and system stability.

“China’s governance framework reflects yet another configuration of welfare-oriented reasoning, one that integrates technological development with state coordination and system stability.”

Procedural safeguards are structural requirements for legitimate governance institutions, not optional design features.

“For this reason, procedural safeguards cannot be treated as optional design features; they represent structural requirements for legitimate governance institutions.”

Governance arrangements that strengthen procedural oversight without addressing capability inequalities may inadvertently reinforce structural barriers to participation.

“Governance arrangements that strengthen procedural oversight without addressing these capability inequalities may inadvertently reinforce structural barriers to participation.”

Welfare-oriented reasoning functions as a causal mechanism shaping institutional divergence.

“Taken together, these cases demonstrate that welfare-oriented reasoning functions as a causal mechanism shaping institutional divergence.”

• AssumptionRests on an assumption: Governments do not design regulatory systems by applying theories of justice.

Meaningful participation in AI development and governance often depends on access to computational infrastructure, technical expertise, and regulatory literacy.

“At the same time, meaningful participation in AI development and governance often depends on access to computational infrastructure, technical expertise, and regulatory literacy.”

Normative evaluation of AI governance requires institutional criteria rather than purely technological or economic metrics.

“First, normative evaluation of AI governance requires institutional criteria rather than purely technological or economic metrics.”

Conformity assessments and documentation regimes impose nontrivial administrative burdens.

“Empirical analyses of AI governance implementation indicate that conformity assessments and documentation regimes impose nontrivial administrative burdens (Alan Turing Institute 2024).”

• cites Alan Turing Institute 2024Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The European Union constructs welfare reasoning through a different institutional logic, one centered on risk differentiation and the protection of fundamental rights.

“The European Union constructs welfare reasoning through a different institutional logic, one centered on risk differentiation and the protection of fundamental rights.”

The EU regulatory framework presumes familiarity with risk classification methodologies, data governance protocols, and interpretive engagement with evolving standards.

“The regulatory framework further presumes familiarity with risk classification methodologies, data governance protocols, and interpretive engagement with evolving standards.”