—Ifeoma Ajunwa, Asa Griggs Candler Professor of Law, Emory University School of Law; Founding Director, A.I. and the Future of Work Program, LL.M., Yale Law School; Ph.D. in Sociology (Organizational Studies and Law & Society), Columbia University; and Cheng-chi “Kirin” Chang (張正麒), Research Fellow, U.S.-Asia Law Institute, NYU School of Law; Affiliate Research Fellow, A.I. and the Future of Work Program, Emory University School of Law
The question of A.I. provenance sits at a critical juncture of fundamental rights and emerging technology governance. Verifying the origins of both A.I. training data and generated content directly implicates constitutional protections for speech and the fight against misinformation. As synthetic content transcends national boundaries and erodes trust in democratic discourse, these challenges demand new frameworks that harmonize international standards while respecting state sovereignty. The stakes involve not merely technical verification, but the preservation of human rights, the integrity of public law systems, and the future of digital constitutionalism in an age where algorithms increasingly mediate access to information and opportunity.
When Marianna Vyshemirsky fled a bombed maternity hospital in Mariupol in 2022, Russian officials weaponized digital skepticism to dismiss photographs of her injuries as fabrications. They claimed an “actor” staged the wounds. When A.I. systems generated images showing immigrants supposedly eating pets in Springfield, Ohio during the 2024 U.S. presidential campaign, these false images spread across social media and triggered bomb threats endangering Haitian immigrants. These incidents reveal how synthetic content transcends national boundaries, creating an urgent need for international governance frameworks.
Recent research shows that consumers correctly identify A.I.-generated content only 50% of the time, performing no better than random chance. This authentication crisis affects democracies worldwide, enabling what scholars call the “liar’s dividend,” where bad actors dismiss genuine evidence by falsely claiming A.I. systems produced it. Meanwhile, A.I. developers train their systems on vast datasets whose origins and biases remain largely undocumented, perpetuating discrimination across borders and cultures. The absence of international standards for verifying both A.I. training data and generated outputs threatens to fragment global governance efforts, creating regulatory gaps that malicious actors will inevitably exploit.
Our forthcoming UCLA Law Review article “Provenance” proposes a comprehensive framework addressing these interconnected challenges through coordinated international action. This article examines why effective A.I. provenance governance requires global cooperation, explores recent developments in international standard setting, and outlines practical mechanisms for harmonizing verification requirements across jurisdictions. This article builds on Professor Ajunwa’s AI and Captured Capital Yale Law Journal forum article which demonstrated the need for cross-border collaborations to govern the development of AI technologies.
The Limits of National Solutions
California’s pioneering A.I. Transparency Act and Training Data Transparency Act represent significant advances in domestic regulation. These laws require providers to implement content markers and disclose training dataset information. The European Union’s A.I. Act establishes comprehensive authentication requirements including watermarks, metadata, and cryptographic verification. China implemented the world’s first binding regulations for generative A.I., mandating user verification, algorithm registries, and content labeling. Yet these divergent national approaches, while well intentioned, create three critical problems that only international coordination can solve.
First, synthetic content circulates globally regardless of where developers create systems. National authentication requirements prove ineffective when content crosses jurisdictional boundaries, as digital markers may not survive format conversions or platform transfers.
Second, fragmented regulations impose impossible compliance burdens on A.I. developers operating internationally. This particularly disadvantages smaller companies and researchers from developing nations, potentially concentrating power among established technology giants.
Third, divergent national standards create opportunities for regulatory arbitrage. Without harmonized baseline requirements, the least restrictive regulatory environment effectively sets global standards.
Building on Existing International Institutions
Rather than creating new bureaucratic structures, effective international A.I. governance should leverage existing institutional frameworks.
The World Trade Organization provides crucial infrastructure for harmonizing technical standards. However, the WTO’s Technical Barriers to Trade Agreement currently excludes services from its regulatory scope, likely encompassing A.I. systems. Expanding the TBT framework with provisions specifically addressing A.I. training data standards offers a practical path forward.
The World Intellectual Property Organization could coordinate output provenance verification through a tiered certification system. This framework would implement differentiated authentication requirements based on content risk levels and usage contexts. High risk commercial content, including political communications and financial advertisements, requires comprehensive verification protocols with immutable watermarks and complete provenance trails. Lower risk or non commercial content follows simplified authentication standards while maintaining essential traceability.
UNESCO would address the cultural and educational dimensions that national regulations often overlook. The organization could coordinate ethical guidelines and cultural preservation standards for A.I. training data, ensuring that provenance requirements respect diverse perspectives while preventing exploitation of cultural heritage. This becomes particularly important as current A.I. systems demonstrate consistent Western cultural bias across model versions, defaulting to Western values regardless of prompt variations.
The United Nations Development Programme would oversee technology transfer initiatives and capacity building programs, ensuring developing nations can effectively participate in the global A.I. ecosystem. This addresses concerns about a new divide in A.I. capabilities between developed and developing nations, establishing regional centers of excellence for A.I. development and verification similar to existing technology transfer initiatives.
Recent Progress and Remaining Gaps
The 2025 International A.I. Standards Summit in Seoul and the subsequent Paris A.I. Summit marked pivotal shifts toward action oriented implementation strategies. The Paris Summit secured commitments exceeding 109 billion euros for A.I. development. Sixty-one countries signed the declaration, including China but excluding the United States and United Kingdom, underscoring both progress and persistent challenges.
Technical standardization efforts parallel these diplomatic advances. ISO/IEC JTC 1/SC 42’s standards on data quality for machine learning provide foundations for documenting training data provenance. NIST’s A.I. Risk Management Framework offers guidance on black box model analysis and system observability. The European Commission’s 2025 training data template initiative demonstrates practical implementation pathways.
However, significant gaps remain. Current international efforts focus primarily on output provenance, verifying A.I. created content, while largely neglecting input provenance challenges concerning training data origins, quality, and ethical collection. The labor practices of data labeling workers, disproportionately located in the Global South and often paid less than two dollars daily, receive insufficient attention despite their fundamental role in A.I. system development. Kenyan workers who label content depicting child sexual abuse and other traumatic material to train systems like ChatGPT endure unaddressed psychological damage, highlighting how current frameworks fail to account for human costs embedded in A.I. supply chains.
Moreover, existing standards inadequately address systematic underrepresentation of minority groups in training datasets. Contemporary intervention studies reveal that only thirty-one percent recorded ethnicity information, with ethnic minorities severely underrepresented across documented cases. This deficiency creates cycles where A.I. systems trained on incomplete datasets perpetuate and magnify existing societal biases, manifesting as algorithmic discrimination in healthcare diagnostics, employment screening, and criminal justice applications.
Implementation Mechanisms and Enforcement
Translating broad principles into concrete action requires sophisticated implementation mechanisms that can adapt to diverse national contexts while maintaining interoperability. The proposed framework establishes what scholars term “regulatory coherence” zones, areas where nations with similar approaches develop deeper cooperation while maintaining broader framework compatibility. This allows regional variations in implementation while ensuring interoperability through common technical standards.
The European Union could maintain more stringent requirements under its A.I. Act while participating in the international framework, similar to how GDPR coexists with other privacy regimes. ASEAN nations might adopt simplified documentation standards reflecting their distinct technological capabilities and priorities. This flexible approach recognizes that strict harmonization proves neither feasible nor desirable, instead focusing on establishing baseline requirements that accommodate regional adaptation.
Enforcement combines traditional trade measures with digital mechanisms. A specialized A.I. Dispute Resolution Body operating under joint WTO and WIPO administration would handle provenance verification disputes through a three tiered system. Technical mediation resolves routine disagreements, binding arbitration addresses significant violations, and appellate review considers cases with broader implications. This structure enables rapid resolution of technical issues while maintaining consistent interpretation of international standards.
Member states could impose trade sanctions for serious violations while implementing technical measures such as blocking non compliant A.I. systems from accessing international networks. This dual approach provides both immediate and long term enforcement options while maintaining proportionality. The framework explicitly incorporates adjustment mechanisms inspired by the Montreal Protocol, establishing clear procedures for updating technical standards and verification requirements without requiring full treaty renegotiation.
Bridging the Development Gap
Effective international governance must prevent A.I. capabilities from creating new forms of technological inequality. The framework addresses these concerns through three specific mechanisms. First, it leverages the United Nations Technology Bank for Least Developed Countries, supported by UNCTAD, to assess and facilitate A.I. infrastructure development in developing nations. Second, it establishes a multi stakeholder A.I. capacity development network through the UN system, where nations collaborate on governance frameworks and technical assistance. Third, it implements preferential treatment provisions similar to those in WTO agreements, providing developing nations with extended implementation timelines and reduced compliance burdens.
These provisions recognize that developing nations face distinct challenges in implementing provenance requirements. Limited technical infrastructure, scarce expertise in A.I. systems, and competing development priorities require tailored support mechanisms. Without such provisions, international standards risk becoming barriers that exclude developing nations from meaningful participation in the global A.I. ecosystem, perpetuating rather than reducing technological inequality.
Technical assistance programs must extend beyond financial support to include knowledge transfer, capacity building, and shared computing resources. Regional centers of excellence could provide training for local developers, researchers, and policymakers while adapting international standards to local contexts. This approach ensures that provenance frameworks serve global interests rather than entrenching advantages for early adopting nations.
Looking Forward
The rapid evolution of A.I. technology demands governance frameworks that can adapt to continuous innovation while maintaining effective oversight. Recent developments underscore both the urgency and complexity of this challenge. DeepSeek’s emergence, rapidly surpassing established U.S. providers in various performance metrics, illustrates how quickly competitive advantages shift in the A.I. marketplace. New capabilities like fine tuning, which shapes model behavior with relatively small datasets while maintaining performance, create opportunities for implementing provenance controls after initial training, rather than relying solely on comprehensive documentation of original training data.
The Council of Europe’s Framework Convention enables parties to assess whether A.I. systems’ risks prove incompatible with human rights standards, potentially requiring moratoriums or bans. The UN’s newly established A.I. Advisory Body brings together expertise from governments, business, technology communities, civil society, and academia to provide rapid preliminary recommendations. These institutional innovations demonstrate growing recognition that effective A.I. governance requires ongoing adaptation rather than static regulatory frameworks.
Yet institutional mechanisms alone cannot ensure effective provenance verification. Success ultimately depends on sustained political will, adequate resources, and genuine commitment to addressing power asymmetries between technology producing and technology consuming nations. The December 2025 A.I. Standards Summit provides a crucial platform for advancing these reforms, building on momentum from ISO/IEC certification frameworks and the Global Digital Compact’s call for interoperable governance.
The stakes extend beyond technical standards or regulatory compliance. Provenance frameworks fundamentally concern whether we can maintain shared reality in an era of synthetic content, whether democratic discourse can survive the erosion of information trust, and whether A.I. development will reinforce or reduce global inequalities. International cooperation on provenance verification represents not merely a technical necessity but a moral imperative, essential for preserving human dignity and democratic governance in an increasingly digital world.
Just as food labeling evolved to protect consumers globally, establishing robust provenance requirements for A.I. content represents a crucial step in maintaining social trust across borders. The path forward requires balancing innovation with accountability, technical sophistication with practical implementation, and global coordination with respect for national sovereignty. Recent developments suggest that the international community increasingly recognizes these challenges and the urgent need for coordinated action. Whether this recognition translates into effective governance frameworks will determine not only the future of A.I. technology but the future of our shared information ecosystem and, ultimately, our ability to distinguish truth from fiction in the digital age.
Suggested citation: Ifeoma Ajunwa and Cheng-chi “Kirin” Chang, Governing A.I. Across Borders: Why Provenance Demands Global Cooperation, Int’l J. Const. L. Blog, Jan. 23, 2026, at: http://www.iconnectblog.com/governing-ai-across-borders-why-provenance-demands-globabl-cooperation/