Cardozo Law Institute in Holocaust and Human Rights

This submission addresses the Dialogue's fourth thematic cluster—respecting, protecting, and promoting human rights—from a specific and underrepresented analytical perspective: the relationship between artificial intelligence and atrocity prevention. While AI raises a broad range of governance concerns, this submission focuses on the challenges, gaps, and regulatory questions that arise specifically when AI intersects with the conditions that enable mass atrocity. The analysis draws on the United Nations Framework for Analysis for Atrocity Crimes as its primary analytical baseline, mapping documented patterns of AI use against the Framework's fourteen risk factors, and drawing on a wider mapping study as well as reports from the United Nations, Human Rights Watch, the International Committee of the Red Cross, Access Now, and Amnesty International. Analysis across the fourteen risk factors of the UN Framework of Analysis for Atrocity Crimes reveals a consistent pattern. AI technologies function as force multipliers for traditional atrocity risk factors rather than as wholly novel threats. They systematically undermine institutional safeguards, amplify discriminatory targeting capabilities, and create new mechanisms for mass surveillance and population control. The Framework identifies as risk factors conditions such as the absence of rule of law protections, the presence of enabling impunity and the capacity of would-be perpetrators to mobilize and target groups. AI intensifies each of these conditions where they already exist by entrenching and accelerating atrocity risk. Critically, the UN Framework of Analysis for Atrocity Crimes was drafted at a time when AI technologies were less widely available, and the capabilities of these systems were still unknown or at least unrealized. As AI becomes more deeply embedded, practitioners applying the Framework to real-world early warning will increasingly encounter AI-driven dynamics that the Framework provides no guidance on how to assess. Remediation of this gap is urgent.

• Protection and promotion of human rights
• Transparency, accountability, and human oversight
• Interoperability of governance approaches

International Criminal Law (ICL), as currently structured, is poorly suited to AI-enabled atrocity accountability. While ICL focuses on individual liability, most accidents in armed conflict arise from systemic failure and distributed decision-making. The mens rea requirement maps poorly onto systems whose unpredictability is foreseeable at the design stage even if specific harmful outcomes are not. Further, ICL produces perverse incentives, including the selection of individual "fall guys" that insulates systemic actors and obscures structural failures. Article 36 of Additional Protocol I requires legal review of new means and methods of warfare. Article 36 was designed for systems with fixed capabilities but is an inadequate framework for AI-enabled weapons systems capable of in-field learning, modification, and behavioral change after deployment. The international community should establish a lifecycle evaluation requirement in addition to the one-time Article 36 review. The ICRC has observed that with a software update or a change in military doctrine, systems initially constrained could readily become tomorrow's autonomous weapons systems, namely weapons systems that select and apply force to targets without human intervention. Autonomous systems whose capabilities evolve after deployment raise qualitatively different accountability questions than those whose capabilities are defined at manufacture. This concern has concrete present-day dimensions: a 2026 amicus brief filed by Access Now, the Center for Constitutional Rights, and allied organizations in ongoing litigation over AI's role in military targeting documented that AI-enabled kill-chain compression has reduced targeting decisions from days or weeks to mere seconds, rendering human authorization effectively a rubber stamp on AI-recommended strikes rather than a meaningful exercise of human control.

Every technology company that develops or deploys AI at scale in conflict-affected environments operates within an international human rights law framework that recognizes corporate responsibility to respect human rights under the UN Guiding Principles on Business and Human Rights. In practice, enforcement mechanisms remain weak, and human rights due diligence in the AI sector is inconsistent and insufficiently transparent. Technology companies providing AI capabilities to military actors frequently operate under government contracts that formally constrain what they can implement independently, limiting their ability to exercise a meaningful duty of care even where the will exists. Corporate concern in this space has been primarily reactive and reputation-driven rather than institutionalized. And the corporate accountability frameworks developed for atrocity prevention in other contexts have not advanced to meaningfully address AI. A direct threat to the multilateral governance framework is the current political environment in the United States, including executive pressure characterizing AI ethics as an obstacle to competitiveness and formally treating companies that decline to support certain military AI applications as supply chain risks, poses a direct threat to the multilateral governance framework. The ability of any single government to unilaterally de-legitimize human rights standards as constraints on AI development is a vulnerability that the Global Dialogue should address by creating durable, interoperable standards that do not depend on any one state's political will for their continued existence.

The central challenge for the atrocity prevention sector is that most existing legal and institutional frameworks were developed before modern AI capabilities were widely deployed. The UN Framework of Analysis for Atrocity Crimes was drafted when AI technologies were less widely available, leaving practitioners today without guidance for assessing AI-driven dynamics in early warning contexts. The Key Findings in our submission document identify three consequential gaps. First, AI surveillance systems built on commercially available components can be transferred across jurisdictions and redeployed against populations already facing compounding atrocity risk factors. The Integrated Joint Operations Platform, developed partly from U.S.-sourced technology, enabled the internment of an estimated 1.8 million Uyghurs. No governance mechanism addressed that transfer before deployment. Second, algorithmic amplification on social media platforms directly contributed to the Rohingya genocide. The UN Independent International Fact-Finding Mission on Myanmar found that Facebook played a significant role in the atrocities. The business models driving that amplification remain largely ungoverned in conflict contexts. Third, internet shutdowns are systematically used to conceal atrocity. Access Now documented 313 shutdowns in 52 countries in 2025, the highest on record, with 70 coinciding with documented grave human rights abuses. The ICC's December 2025 policy recognizing that shutdowns can facilitate international crimes provides a foundation the Dialogue must build on. Across all three areas, the corporate accountability gap is the connective problem: human rights due diligence in the AI sector remains voluntary, inconsistent, and applied with weaker protections in lower-income and conflict-affected markets where atrocity risk is highest

The AI Dialogue's most important contribution would be to create durable governance infrastructure that does not depend on any single state's political will. Current U.S. executive policy has formally characterized AI ethics standards as obstacles to competitiveness. Governance frameworks built through multilateral consensus are more resistant to that kind of unilateral erosion. Three concrete roles stand out. First, the Dialogue can catalyze coordination between the AI governance community and the UN Office on Genocide Prevention to develop AI-specific guidance under the Framework of Analysis for Atrocity Crimes. The Framework was drafted when current AI capabilities were not yet widely deployed; its practical application now requires supplementary guidance for surveillance, disinformation, and autonomous weapons contexts. Second, the Dialogue can advance a consensus position that internet shutdowns are an international law concern with atrocity prevention implications, building on the ICC's December 2025 cyber-enabled crimes policy. This would provide normative foundation for digital ceasefire provisions that civil society has documented as absent from peacemaking processes worldwide. Third, the Dialogue is well-positioned to bridge the human rights governance and AI governance communities, which continue to operate largely in parallel. The autonomous weapons accountability question, the corporate due diligence question, and the surveillance supply chain question are all simultaneously human rights problems and AI governance problems. The Dialogue should insist on that integration rather than treating human rights as one thematic cluster among equals. The Dialogue should also build in accountability mechanisms: a structured review process for commitments made and a formal civil society monitoring role. Without these, outputs risk becoming aspirational documents that states selectively cite without implementation.

Several existing frameworks provide foundations the Dialogue should build upon rather than replicate. The UN Framework of Analysis for Atrocity Crimes (HR/PUB/14/2) is the primary multilateral early warning tool for atrocity prevention practitioners. The Dialogue should work with the UN Office on Genocide Prevention to develop AI-specific supplementary guidance under the Framework, which predates modern AI deployment. This would give practitioners concrete tools they currently lack. The UN Guiding Principles on Business and Human Rights provide the normative architecture for corporate accountability. The Dialogue's added value is moving from voluntary application to binding due diligence obligations for AI companies in conflict-affected environments, a step the UNGPs invite but cannot compel. The ICC's December 2025 Policy on Cyber-Enabled Crimes Under the Rome Statute formally links internet shutdowns to international criminal liability for the first time. The Dialogue should treat this as a normative anchor for a digital ceasefire workstream rather than addressing shutdowns as a separate telecommunications matter. The Access Now #KeepItOn coalition, with over 366 member organizations in 106 countries, has produced the most comprehensive dataset on internet shutdowns globally and has litigated against them in multiple regional courts, including a successful 2025 ECOWAS ruling. The Dialogue should formally engage this network as an implementation and accountability partner. The ICRC's work on autonomous weapons and AI in armed conflict, alongside the Convention on Certain Conventional Weapons discussions, provides the legal and technical foundation for a lifecycle evaluation framework going beyond the one-time Article 36 review. Across all of these, the Dialogue's specific added value is multilateral political legitimacy: translating civil society and institutional work into governance commitments with broader state buy-in.

Meaningful participation requires moving beyond the default model in which technically sophisticated actors from high-income countries set the agenda while communities most affected by AI-enabled harm participate marginally or not at all. Civil society organizations with field expertise in atrocity prevention, human rights documentation, and digital rights, including those working in conflict-affected contexts in the Global South, should be structured participants in every thematic session. Organizations like Access Now, combining technical surveillance expertise with field documentation in over 100 countries, exemplify the civil society knowledge the Dialogue needs to be credible. The Independent International Scientific Panel should include expertise in atrocity prevention and international humanitarian law alongside technical AI expertise. Without it, the Panel's guidance will not be authoritative on the full range of risks AI presents. Affected communities, including diaspora groups subject to transnational AI surveillance, survivors of conflicts where AI played a documented role, and civil society from the Global South facing the sharpest governance asymmetries, should be involved in agenda-setting, not only invited to share testimony. Structurally, the Dialogue should include working sessions organized around documented case studies. A session examining how AI functioned in a specific atrocity context, with technical experts, human rights investigators, legal analysts, and affected-region civil society all present, would produce more concrete governance insights than a session on surveillance or disinformation in the abstract. Follow-up mechanisms matter as much as session format. The Dialogue should establish a clear process for tracking whether commitments translate into national and corporate action, with civil society reporting roles built in from the outset.

The communities most underrepresented in global AI governance discussions are those most exposed to AI-enabled harm. Conflict-affected populations in the Global South are governed by AI systems they had no role in designing, deployed under frameworks they had no role in shaping. Rohingya communities whose destruction was accelerated by algorithmic amplification on Facebook were not substantively consulted in the governance discussions that followed. Uyghur diaspora communities subject to transnational AI surveillance, documented across 44 states in over 7,000 verified cases, have limited formal access to the multilateral forums where surveillance governance is discussed. Atrocity prevention practitioners, including genocide scholars, early warning analysts, and field investigators, are largely absent from AI governance discussions, which tend to be populated by technical researchers, technology companies, and general human rights advocates. Their specific expertise in how mass atrocities begin and escalate is essential for identifying how AI changes those dynamics. Their absence means the Dialogue's outputs are likely to underestimate the most severe risks. Civil society organizations in the Global South with digital rights mandates face acute capacity constraints that limit sustained multilateral engagement. Inclusion requires more than formal invitation: funded travel, working interpretation, advance circulation of materials in accessible languages, and structured time for substantive participation rather than brief opening statements. The Dialogue should create a mechanism for civil society from conflict-affected regions to contribute thematic input between sessions. A standing advisory group with rotating regional civil society representation, supported with resources for meaningful participation, would be one practical model.

The most valuable format innovation would be organizing sessions around verified case studies rather than governance concepts in the abstract. A structured case study of how AI functioned in a specific atrocity context, with technical experts, human rights investigators, legal analysts, and civil society from the affected region all present, would generate insights that thematic panels cannot replicate. The Rohingya, Uyghur, Gaza and Sudan situations each offer documented evidence bases sufficient to anchor a substantive working session. Pre-session written consultation processes are valuable but need structured synthesis. If the Dialogue receives hundreds of civil society submissions and processes them only into a Co-Chairs' summary, the expertise in those submissions is effectively lost. An independent rapporteur producing a synthesis document, shared with participants before sessions convene, would make the consultation process substantively influential rather than formally inclusive. Cross-sector working groups organized around specific governance problems, such as the accountability gap for autonomous weapons, the technology transfer problem in surveillance, and the digital ceasefire question, would allow the Dialogue to produce technically grounded outputs that plenary discussion alone cannot achieve. These groups should include AI developers, international humanitarian law experts, atrocity prevention practitioners, and affected community representatives, with outputs feeding into plenary sessions. Regional preparatory meetings ahead of Geneva sessions, with dedicated rapporteurs synthesizing the discussions, would ensure that the perspectives most needed are substantively present. This is particularly important for states and civil society from regions with the highest AI-enabled atrocity exposure and the least structural power in multilateral governance processes.

Three examples from the atrocity prevention sector are instructive for the Dialogue. The ICC's December 2025 Policy on Cyber-Enabled Crimes Under the Rome Statute formally recognized that internet shutdowns can directly facilitate or contribute to international crimes. This created a normative foundation that did not previously exist and demonstrates that existing international law institutions can adapt to AI-enabled harm through policy development, without waiting for new treaty frameworks, when civil society documentation provides the evidentiary foundation. Access Now's #KeepItOn coalition, with over 366 member organizations in 106 countries, has produced the world's most comprehensive dataset on internet shutdowns, supported strategic litigation in multiple regional courts including a successful 2025 ECOWAS ruling against Senegal's shutdown practices, and driven the normative development that led to the ICC policy above. Its model of technical measurement, field documentation, legal advocacy, and targeted intergovernmental engagement is a replicable framework the Dialogue should engage formally. Anthropic's public refusal in February 2026 to support fully autonomous weapons applications, on the grounds that AI systems are not sufficiently reliable for such use, demonstrates that human rights-grounded governance positions are compatible with commercial viability. Following that refusal, Anthropic's Claude became the top-ranked free application on the Apple App Store. This directly challenges the premise embedded in current U.S. executive policy that human rights standards obstruct AI competitiveness, and the Dialogue should use it in making the case that responsible AI governance is a foundation for, not a constraint on, technological leadership.