INHR

Given our 8 years hosting international dialogues among experts from the leading AI states, INHR believes the UN sponsored global dialogue on AI will succeed if it focuses limited UN skill and attention in key areas and allows innovation and voluntary multi-stakeholder solutions to continue without over-regulation. The UN works best to codify international consensus, support global equity, and implement solutions where its agencies add value. Focus where the UN has expertise - such as ILO's future of work initiative, or building on climate protocols to think about how to power the AI era. Don't let every UN agency speak on AI just because they want to do so - e.g., the UNESCO code of ethics. Leave companies and militaries largely alone because the consensus that the UN requires for binding legal instruments does not yet exist. INHR focuses on AI in the military domain. We work successfully with experts from strategic rivals - China and the U.S.; Europe and Russia - to enhance safety and accountability in the most serious risk areas - agentic AI, weapons of mass destruction, and synthetic biology applications. The Dialogue should control the 50-plus UN agencies claiming to regulate aspects of AI and establish a central leadership under the Secretary General. The Independent International Scientific Panel and UNSYG high-level panels were a good start. Otherwise, the UN's scattershot approach confuses companies and technical experts. AI regulation requires a granular approach - you must understand the risks you wish to mitigate and what is technically possible before articulating solutions. This is no longer 1945. If Co-Chairs continue consulting the same old stakeholders - cross-regional diplomats clueless about how AI functions - while tech experts and AI safety NGOs struggle even to access UN premises, the Global Dialogue will be irrelevant. Technology and its designers will move on without you.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

INHR works to create AI applications that are safe, inclusive and rights-respecting, especially in the field of weapons of mass destruction. We prioritize addressing extreme risks from AI and encouraging global coordination and agreement on mitigation approaches. One way to achieve these goals is through secure, trustworthy AI under human oversight. Our policy recommendations help governments understand AI risk and show private and public developers how to design and operate advanced AI systems responsibly. Our technical AI governance work prioritizes ensuring that risky AI capabilities do not overwhelm human capacity. Engineers must build AI systems to be safe, secure, and trustworthy, regardless of the use of those systems. Technical safeguards like safe-by-design engineering and robust auditing can guarantee verifiably safe outputs and secure system defenses, but this must be accompanied by governance measures that are valid and relevant from technical, political, and national security perspectives. Interoperable and coherent governance approaches are essential. Governance structures across industries, economic sectors, and national borders must be able to regulate the development, use, and outcomes of AI, too. Standards limited to the national level will never be able to ensure the systemic security and safety that AI risk demands, and focused dialogue between countries is necessary to achieve effective governance on a global level. Finally, policy should ensure that AI systems engineered to work safely and securely also produce results aligned with human rights and UN SDGs. INHR's track record of advocacy at the Human Rights Council demonstrates our conviction and expertise. Broadly-diffused advanced AI will be transformative, but it must also be beneficial to humanity. Transparency and accountability are crucial to ensure human rights protection. They are also crucial to ensure that any AI errors are corrected and that remediation measures ensure major errors do not recur.

INHR focuses on how national security applications of AI intersect with these themes and prioritizes mitigating the greatest AI-related risks. In this regard, two cutting-edge issues in military AI are overlooked and deserve more attention: defense applications such as agentic AI in the military, and the highest consequence uses of military AI involving weapons of mass destruction. The international community has made progress in affirming that AI should never hold authority over the decision to use nuclear weapons, but these agreements between AI military superpowers are best reached in bilateral discussions rather than via UN agencies. Multilateral efforts like REAIM and UN Resolution 80/58 have advanced the conversation on military AI governance, and the UN and Global Dialogue can help by disseminating ideas to member states, as AI For Good does with AI security standards. Yet, no binding national law or international agreement exists to prevent AI-driven unintended military escalation, and the UN is not likely to secure one. These accords are more likely to be struck in smaller, more focused multilateral and bilateral meetings where progress is less easily sidetracked. Effective dialogue on military AI is essential to ensure military applications of AI are safe, secure, and trustworthy for the entire international community, including states without military AI capabilities. While states are unlikely to commit to broad transparency of their military AI programs, INHR is able to promote transparency about key elements of our AI governance strategies, encouraging international security. INHR's Recommendations for Governments on Mitigating AIxBio Risks, developed by a multilateral expert group from the U.S., China, India, and Europe, demonstrate that global competitors can support common security initiatives. These Recommendations offer a pathway for similar agreements to be reached on risks from agentic AI in the military and AI's entanglement with weapons of mass destruction.

Governance gaps on AI in the military, particularly AI and weapons of mass destruction, strongly impact the sector where INHR works. Absent bilateral and multilateral agreements, governments and militaries face a negative spiral where competitive pressures deter investment in safety and security. In the USA, OpenAI and Anthropic compete over providing classified LLM processing to the Pentagon and permitted uses. Countries at war (Russia-Ukraine; Israel-Palestine; US-Iran) are unwilling to spend on safety and security because they fear that such expenditures will penalize them in development and innovation of AI, giving a rival a lead on the battlefield or in an AI arms race. The UN is not well suited to fill these governance gaps. Consensus is unlikely where countries at war would need to agree on restraints. Geopolitical tensions prevent the P-5 from exercising leadership, and across the UNGA membership, many already use AI in their military applications, while others see themselves as potential victims of uncontrolled AI. Additionally, the UN lacks technical expertise necessary for making AI governance and operational decisions. In voluntary expert dialogues, INHR has successfully engaged leading AI and military states about restraining AI technology as deployed in the military. Like past arms control, conditions for restraint are ripe because all parties have fear the unknown - uncertainty about what AI can and will do, how adversaries will deploy it for advantage on the battlefield and in domestic economies, and how to prevent terrorists and other malevolent actors from getting their hands on AI tools that could make lethal drones more effective or make biological weapons more transmissible and virulent. These factors make it important for all actors to consider reasonable restraints, but we think multistakeholder initiatives like REAIM are more likely to succeed than formal UN processes.

INHR has three clear principles motivating our work on artificial intelligence. AI in the military should be safe, inclusive and rights-respecting. The Global Dialogue can contribute to all of these areas, but should be careful not to get in the way of progress on AI governance and recognize where it does not have all the answers. Rather, the UN should use its convening and publicity power to promote AI governance practices that are safe, inclusive, and rights-respecting, and make space for the most effective actors already working in this area. Private companies and the tech community will have more applicable, feasible ideas about how to handle AI safety than UN bureaucrats can develop. Private sector AI development and capabilities expertise is unparalleled, and national AI safety institutes have the necessary knowledge and talent to produce technically and operationally-relevant safety evaluations and other AI risk-mitigating measures. The UN is neither best-placed nor best-equipped to drive AI governance, and attempting to do so risks encumbering important progress underway. The UN should focus on sharing best practices developed by technical and political experts through events like AI For Good. The UN's main role should be on inclusivity, ensuring voices from the Global South are meaningfully brought into discussions and able to identify risks and solutions that existing safety initiatives are not yet adequately addressing. Finally, to ensure AI governance is rights-respecting, bodies like the Human Rights Council should consider AI as a risk area that cuts across their work, but the UN must prevent its many agencies from each pursuing workstreams that inefficiently reproduce AI safety work other national and subnational actors do better. The UN should leverage its convening and publicity power for AI governance, but if it tries to govern, it will cause more problems than it solves.

INHR's work to mitigate AI risk consists of several distinct initiatives running in parallel to address the AI risks with the most extreme risk potential: testing and evaluation of military AI, AIxBio security, strategic security issues from AI and nuclear weapons, and agentic AI. Each of these initiatives draws distinct expert partners in to develop policy solutions through intensive dialogue, and would be valuable for the Global Dialogue to build upon. Senior military experts, diplomats, computer scientists, and industry representatives helped to develop INHR's Testing & Evaluation Model Practices document. Chinese, American, and international experts convened under the shared belief that it is essential for the international community to establish a consensus on AI T&E principles and best practices, both to promote adherence to international humanitarian law and to reduce the global risks associated with unanticipated failures of AI-enabled military systems. Global biosecurity also hinges largely on the avoidance of unanticipated failures of AI systems to detect malicious user intent or highly capable novel pathogen creation. General purpose and bio-specific models can each make malicious actors more capable of creating catastrophic bioweapons. INHR's AIxBio Recommendations, endorsed by leading university organizations, biotech companies, and think tanks, would close several key gaps in the world's collective defense against such dangers. INHR's AIxNuclear Dialogue includes all N-5 countries, who agreed that they need to do more to address nuclear risks including AI decision support, ISR, and NC3. INHR and the Global AI Dialogue are complementary. Policies and practices are devised more efficiently by INHR whereas the Dialogue would better deliberate principles and solutions. Smaller, handpicked working groups like those INHR uses produce technically and politically valid recommendations, while more public and universal processes like the AI Dialogue are necessarily slower and subject to political discord.

INHR's dialogues on AI risk mitigation use a multi-stakeholder format to identify, analyze and prioritize among AI-military risks. We combine computer scientists, military experts, former officials, lawyers and specialists to understand risks and propose mitigation measures applying to the highest priority risks of concern. The multistakeholder process functions almost as an interagency process, across delegations including the largest AI and military powers, as well as an international all-star delegation representing a range of countries and regions. One overarching conclusion of our work is that there is no "one size fits all" solution to mitigating the risks of AI in the military. You need to understand the risks and then - from a national perspective - prioritize and decide which risk mitigation approaches are right for your country. INHR takes that process one step further and tries to identify common elements of such risk mitigation and common priorities where even competing powers can agree on restraints. The UN may not be the right party to convene this dialogue but some adjustments could include: - Ensuring civil society and technical experts have equal access to UN decision-makers on AI issues. This requires more than written submissions; it requires physical access to relevant Co-Chairs and diplomats. Environments where member states have better access to UN buildings, personnel and secretariat staff are unfair and discourage technical experts and the private sector. My technical colleagues no longer struggle through security and administrative hurdles needed to enter UN grounds. - Thematic breakout sessions should focus on specific risks and practical applications. Stop assigning speakers based on nationality. Focus on expertise in the thematic area of concern. - Remember that AI has huge benefits as well as risks. Encourage benefits while supporting existing structures to manage and mitigate risks, rather than replacing them with ineffective UN mandates.

Not all countries can be AI superpowers, but all should be AI governance superpowers. Global South views often go under-represented in AI governance. Countries lacking large language models or tech industries may be left out, further reinforcing inequities. All countries need to develop liability structures for AI-induced harm, policies on AI-displaced workers, and approaches to the cost of powering the AI era. INHR believes AI should be safe, inclusive and rights-respecting. We prioritize training developing states on-demand to strengthen their inclusion in global AI governance discussions. Under-representation can magnify risks from AI. A Baidu study showed that guardrails in English and European languages are easily circumvented if an LLM user switches to a language like Lao. This can have huge consequences – consider the results if a search on "How to make a biological weapon?" is blocked in English but permitted in Burmese. Similarly, datasets for medical applications of AI are notoriously based on white male patients - producing poor diagnoses and outcomes for women and minorities. INHR is proud to have been a finalist in the AI For Good's first global competition for its ideas on how to use legal cases on discrimination law to impose guardrails to check gender bias. During INHR's visit to the India AI Impact Summit, we witnessed how appropriate governance helps SMEs thrive, while other approaches suppress innovation. Each country will need its own inclusive structures - empowering women, minorities and civil society. The Global Dialogue is well placed to consider inequity based, for example, on impacts of climate-induced disaster and displacement, or to consider risks and carbon lock-in from AI infrastructure. INHR aims to convene the world's AI powers to solve this problem, not entrench it. Rather than complain about lack of inclusion, the Global Dialogue should encourage innovative solutions. If it does so, INHR can help.

The Dialogue should consider formats calibrated to the technical complexity and political sensitivity of AI governance. We find that proposals for risk mitigation work best when parties first understand the cause of AI risks of concern and then try to develop solutions. This requires depth of understanding about the causes and contexts for AI risks that a party believes require regulation. Absent this technical understanding, it remains challenging to find political consensus on solutions. INHR's AI dialogues prioritize research, pre-dialogue working groups, and scenario exercises to focus experts' attention on potentially mutually-acceptable risk mitigation solutions. If the Global Dialogue wants to learn from this experience, we can help. The UN Global Dialogue would do well to follow this example. Pre-Dialogue expert working sessions organized around specific technical governance problems would allow domain experts, civil society organizations, and government representatives to develop shared analytical frameworks before entering the plenary. INHR has found, in its multilateral engagement including the REAIM process and BWC intersessionals, that the gap between technical understanding and political deliberation is a primary obstacle to effective norm development. Structured pre-Dialogue sessions can reduce this gap before political positions have hardened. INHR is available to contribute to the design of domain-specific working sessions and scenarios that illustrate risks - particularly in the fields of weapons of mass destruction, agentic AI and AI-enabled biosecurity risks.

INHR operates a unique process of multidisciplinary and international discussion of AI-military risks that has produced concrete and actionable recommendations on risk mitigation and governance that are co-developed and validated by geostrategic competitors. INHR's Model Practices for Testing and Evaluation of Military AI, its AIxBio Recommendations to Governments, and our forthcoming multi-agent governance framework, were developed through expert dialogue with participation from China, Russia, Europe, India, and the United States. Our solutions work because they are concrete and focused on the risks all states face. The proposals make us all safer. For example, INHR's Model Practices for Testing and Evaluation encourage full-lifecycle practices to ensure AI operates predictably under effective governance. INHR's current work on multi-agent national security issues addresses loss of control risks, cyber security vulnerabilities, and auditability of inter-agent interactions, with a framework of recommendations forthcoming. Finally, our AIxBio Recommendations, endorsed by civil society organizations across five continents, address specific risks at the intersection of AI and synthetic biology with tailored solutions for governments and companies. Taken together, INHR's AI dialogues have a track record of generating concrete solutions for the challenges at AI's technical frontier. Our practices and approaches work in part because they acknowledge a political reality: states will continue augmenting their national defense and scientific discovery with AI. Because our dialogue focuses on using these technologies prudently rather than banning them, it is possible to include the industrial and political actors whose support is necessary to effect the necessary outcomes.