Independent Researcher

The first Global Dialogue on AI Governance succeeds if it produces what institutional submissions cannot: an honest account of why AI governance is failing politically, not a description of where it should arrive. Most institutional actors will describe the destination. What's missing is a reckoning with why the political conditions for getting there don't exist, and what historical experience tells us about what changes that. Nuclear governance is instructive. The Cuban Missile Crisis of October 1962 catalysed the political momentum behind the Limited Test Ban Treaty of 1963 and the Non-Proliferation Treaty of 1968. Crisis generated the political will that seventeen years of proposals hadn't. I'm not convinced that pattern has broken for AI governance, and the Dialogue cannot afford to proceed as if it has. There are four documented domains where AI is already producing operational harm: the use of personal data to persuade populations without their knowledge, to enforce state power outside warrant-based accountability, to exclude people from housing, employment, and credit through invisible algorithmic systems, and to manufacture false reality at a scale and speed that outruns any correction mechanism. None yet has a coherent cross-border governance framework with meaningful reach and enforcement. A successful first session gives independent researchers and civil society genuine floor time with written response rights, names the specific gaps where no framework exists, and produces a Co-Chairs' summary that distinguishes confirmed governance commitments from aspirational positions. It should also make a clear recommendation on AI integration into nuclear command, control, communications, and intelligence infrastructure as one of the most consequential governance issues before the Dialogue.

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

My research spans two bodies of work: Bridging the Wisdom Gap, a historical analysis of AI and nuclear weapons governance published on SSRN, and an investigative series on commercial and state surveillance across four operational domains. On safe and trustworthy AI: open-source defence reporting and peer-reviewed research indicate that multiple nuclear-armed states appear to be exploring or integrating AI into military decision-support systems relevant to nuclear posture. No dedicated multilateral framework appears to govern that convergence. The highest-stakes application of AI is receiving among the least dedicated governance attention. On human rights and human oversight: these are inseparable in practice. My research on AI Meta-Bias examines how independent evaluations of police facial recognition have reached conflicting conclusions on demographic bias, including evaluations of the Essex Police deployment. Same system, same data, opposite conclusions. The measurement layer carries the assumptions of whoever built it. Transparency without access to that layer is not transparency, and human oversight without understanding what is being overseen is not oversight. On interoperability: fragmentation is the structural problem, not a peripheral one. Global trackers identify more than 1,000 AI policy initiatives across dozens of countries, most non-binding and incompatible. AI governance is repeating the structure of the failed Baruch Plan of 1946-47, where trust and verification mechanisms were absent. Agreeing what is being measured has to precede arguing about what is being governed. One concrete recommendation: the Dialogue should propose a baseline taxonomy for AI incidents across jurisdictions, supported by a shared incident reporting mechanism, making cross-border comparison of harm possible for the first time.

The most urgent gap across all four listed themes is AI integration into nuclear command, control, communications, and intelligence infrastructure. Multiple nuclear-armed states appear to be exploring or integrating AI into military decision-support systems relevant to nuclear posture, documented in open-source defence strategies, peer-reviewed academic research, and ongoing strategic stability dialogues. No dedicated multilateral framework appears to comprehensively govern that convergence. Nuclear governance took twenty-three years to build the verification mechanisms, inspection regimes, and binding agreements that eventually constrained proliferation. AI governance is at an earlier stage, and the two technologies are already converging. Stanislav Petrov's 1983 decision to distrust a Soviet early-warning system and not report an apparent United States launch worked because a human being had both the authority and the time to make that judgement. AI-assisted decision support in nuclear contexts compresses the time available and shifts the authority structure in ways existing human oversight frameworks don't appear to account for. A second cross-cutting issue is the civilian-to-military pipeline in commercial AI infrastructure. Open-source reporting has linked commercially available robotics platforms, including Unitree systems, to Chinese military experimentation through university procurement channels under China's military-civil fusion directive. The line between civilian and military application can dissolve without anyone formally crossing it, and current governance frameworks are designed for line-crossing rather than dissolution. Nearly a decade of negotiations at the Convention on Certain Conventional Weapons has not produced binding global restrictions on autonomous weapons. The Dialogue convenes in Geneva in July 2026. I'd recommend the Co-Chairs request that the Independent International Scientific Panel include a dedicated assessment of AI-nuclear integration risks before the 2027 session, and that the Dialogue establish minimum no-launch-authority automation norms for nuclear-armed states. Historical experience suggests major governance advances follow crisis rather than precede it. The Dialogue has a narrow window to be different.

The governance gaps in the priorities I've selected are affecting the United Kingdom and wider European region in concrete ways, with a consistent pattern across them: deployment is moving faster than democratic oversight, legal clarity, and independent scrutiny. In the United Kingdom, live facial recognition has expanded rapidly through police use in public spaces, including hundreds of deployments in London and fixed camera installations in Croydon. Millions of faces have reportedly been scanned before Parliament has established a clear statutory framework. The Home Office launched its first consultation on a legal framework in December 2025, by which point the technology was already operating across multiple police forces in England and Wales. The deployment preceded the governance. A similar pattern is emerging in predictive public-sector analytics. Recent UK proposals to use machine learning for early intervention involving children, drawing on datasets collected for other purposes, raise serious questions about consent, proportionality, accuracy, and redress. Across Europe, the challenge is increasingly one of fragmentation. The EU AI Act establishes a baseline, but implementation capacity, enforcement, and national exemptions vary across jurisdictions. For practitioners working in public-sector AI implementation, the burden of decoding how the EU AI Act, emerging UK frameworks, sector-specific guidance, and any future global instrument from this Dialogue interact falls on individual teams rather than being resolved at policy level. For independent civil society researchers, governance gaps create an evidence-access problem. Procurement details, training data, error rates, and operational use often sit inside state agencies or private vendors, limiting meaningful external scrutiny. The opportunity is concrete: common standards for incident reporting, stronger transparency obligations for high-risk public-sector AI, interoperable rights protections across borders, and structured inclusion of practitioners and independent researchers in governance processes.

International cooperation on AI governance keeps stalling for structural reasons that most governance forums are not designed to name. I've called them the 3 Ps: power, profit, and politics. Power: the organisations building the most capable AI systems have become the primary source of expertise about those systems. The regulators, the standards bodies, the advisors to governments all draw from the same pool. The IAEA doesn't ask weapons manufacturers to write nuclear safeguards, and the reason isn't that weapons scientists lack knowledge. It's that independence of assessment from the assessed is a governance principle. We haven't made that decision for AI. Profit: voluntary commitments made in genuine good faith dissolve when commercial pressure reaches a sufficient threshold, because that is what voluntary means. Self-regulation works until it costs something. The pattern is documented across the industry, most recently when one major lab held its red lines on Pentagon contracts and a competitor took the same contract with the conditions removed. Politics: geopolitical competition burns the international agreements needed to make anything binding. This Dialogue was adopted by consensus without a vote. The fractures between the Member States who supported it are the reason a decade of autonomous weapons negotiations has produced nothing enforceable. The Dialogue's role is to be the one forum where all three can be discussed together, by people who aren't funded by the entities being discussed, and to convert that diagnosis into practical outputs: shared terminology, interoperable minimum standards, conflict-of-interest transparency, and implementation guidance public institutions can actually use. The gap between what's agreed at international level and what gets implemented locally is where cooperation fails silently, every day.

Pockets of serious work on AI governance exist everywhere I look. UNIDIR on AI in security, the ICRC on autonomous weapons and meaningful human control, SIPRI on AI-nuclear integration, the OECD AI Incidents Monitor, the Council of Europe Framework Convention on AI, Stop Killer Robots building Member State coalitions over nearly a decade. The quality of work across these forums is genuine. Coordination between them remains limited and fragmented. CCW discussions address conventional autonomous weapons without touching nuclear applications. Nuclear arms control treats AI integration as domestic implementation. AI governance initiatives concentrate on civilian bias, privacy, commercial accountability. Three diplomatic tracks running alongside each other, none looking sideways. The most urgent failures sit in the spaces between them, and nobody has a clear mandate over those spaces. My surveillance research documents operational harm across four domains, persuasion, enforcement, exclusion, and the simulation of the dead, that no single existing initiative covers in full. The Dialogue can own those gaps. Every Member State at the table, not just those who can afford Bletchley or OECD summits. Civil society and independent researchers on the same footing. And a UN General Assembly mandate that none of the existing forums carry. I think there's something the governance conversation hasn't caught up with yet: AI could enable the coordination it currently lacks. Cross-jurisdictional incident mapping, pattern recognition across fragmented regulatory landscapes, real-time tracking of where frameworks conflict. I'm building governance tooling for local government, and the irony of using AI to manage AI governance is not lost on me. But it works, and the scale of the coordination problem may need the technology the coordination is about. Whether the Dialogue is brave enough to try that is a different question, and I hope it is.

I read the proposed July programme and my first reaction was simple: who is this for? If the answer is diplomats talking to diplomats, the structure is fine. If the answer is supposed to include the people AI governance affects, and the people inside organisations watching ethics get overruled in real time, it needs to change. The people who know what AI systems do to communities should speak first in every thematic session, before the experts, before the scene-setting. Welfare claimants scored by algorithms, families subject to facial recognition, workers inside AI content moderation pipelines. Their evidence changes the conversation because they're describing what the system did to them, not what it was designed to do. There is currently no global mechanism for people inside AI companies, procurement teams, or public-sector deployments to report ethical concerns safely. I've heard from people who've witnessed governance being bypassed, ethics frameworks overruled, harms dismissed as acceptable cost. They have nowhere to take that. The Dialogue should recommend a protected disclosure mechanism for AI governance concerns, designed for an industry where the people who see the harm are often contractually prevented from naming it. Independent researchers and civil society need written response rights to plenary statements, included in the Co-Chairs' summary. Floor time during a session disappears, text in the record doesn't. The filtering problem I've described elsewhere, where institutional affiliation determines whose contribution gets weight, needs to be addressed in how the Dialogue selects speakers and reviewers. The Dialogue also needs to produce something practitioners can use. I build AI governance tooling for local government, playbooks, risk assessments, oversight frameworks. If it produces another summary, it joins the pile. If it produces tools, people can use it.

Global AI governance is an almost impossibly large problem, and the voices shaping it are a vanishingly small subset of the people affected by it. There is an enormous amount of AI research and evidence-building happening outside institutions. Independent researchers, charity workers, community organisations, and activists are documenting AI harms and building responses, often because they experienced the harm directly or found it in other work. They do it without funding, without institutional backing, and the quality is frequently as rigorous as anything from funded labs. Mushtaque Ahmed Rajput built the Visual Inverse Turing Test from Karachi because AI systems were failing non-Latin scripts and nobody with institutional resources had looked. I've published a working paper on SSRN and built governance tooling for local government without university affiliation. Neither case is unique, and the consultation infrastructure doesn't reach this work. The question is how to change that at scale, and citizen science offers a model. The Covid Symptom Study enrolled 4.7 million participants because someone built the tool and gave people a reason to use it. iNaturalist has 4 million observers contributing 250 million biodiversity observations. The Citizen Science Global Partnership operates as a network-of-networks connecting grassroots researchers across borders. AI governance has no equivalent infrastructure, and it needs one: a global reporting network where communities can document what AI systems are doing to them, in their own languages, feeding directly into the governance process. Beyond independent researchers: data labellers, content moderators, and trust and safety workers hold operational knowledge governance processes never hear. Children are governed by systems they didn't consent to. Global South communities bear concentrated harms while the conversations happen elsewhere. The Dialogue could pilot a call for community-submitted evidence routed through networks like Global Voices, which has 1,200 contributors across 180 countries reporting on digital rights.

The most effective engagement format for the Dialogue isn't confined to two days in Geneva, it's the infrastructure between sessions. The Covid Symptom Study reached 4.7 million participants because someone built a tool that took 30 seconds, worked on a phone, and influenced public health decisions. iNaturalist has 4 million observers contributing 250 million biodiversity observations because the friction is near zero and the data goes somewhere that matters. These aren't consultation formats, they're evidence infrastructure, and they work because they meet people where they already are. AI governance has no equivalent. The Dialogue could build one: a multilingual reporting tool where communities document what AI systems are doing to them, accessible on a phone, feeding into a structured database the Independent International Scientific Panel and the Co-Chairs can draw on. A living evidence base, updated continuously, designed so that a welfare claimant in Lagos or a content moderator in Nairobi can contribute as easily as a professor in Geneva. The format innovation is making the Dialogue a year-round process rather than a two-day event. The July session becomes the moment where accumulated evidence gets heard and acted on, not the only moment where participation is possible. Between sessions, the Dialogue could pilot collaborative working sprints where practitioners work on shared governance problems using AI-assisted coordination tools. I build this kind of tooling for local government, and the model works: bring people with the same operational problem together, give them a shared workspace, produce something usable by the end. That produces implementation guidance faster than any drafting committee. The principle is borrowed from every successful global data initiative: reduce friction, build in local languages, make the contribution visible, make sure it gets used. Coca-Cola reaches every country on earth through distribution infrastructure, not summits. The Dialogue should learn from that.

There are good examples of AI governance that work, and they share a feature: they produce consequences or usable infrastructure rather than principles nobody enforces. The Dutch SyRI ruling in 2020 struck down an algorithmic welfare fraud system by applying existing human rights law directly, no AI legislation needed. That model is replicable anywhere with equivalent legal infrastructure, and civil society organisations are already building the litigation strategies. The OECD AI Incidents Monitor provides cross-jurisdictional evidence allowing pattern recognition across borders. The Council of Europe Framework Convention on AI anchors governance in human rights rather than technical definitions that date within months. What's missing is the knowledge infrastructure underneath. Tens of thousands of AI governance papers exist across arXiv, SSRN, Zenodo, and institutional repositories, growing faster than peer review can absorb, with no central register, no shared categorisation, no way to surface the independent researcher in Karachi alongside the lab in San Francisco. The conversation draws on whatever fraction the people in the room happen to have read. Peer review has a structural problem in this domain. The pool qualified to review work on autonomous weapons in nuclear contexts is vanishingly small, most sitting inside military institutions with positions on the outcome. Inside review carries institutional bias, outside review has gaps where information is classified. The answer might be something AI governance hasn't tried: a transparent, AI-assisted research commons that ingests work from all sources, categorises against explicit published rules, and surfaces patterns across the full evidence base rather than whatever made it through institutional gatekeeping. The alternative is pretending manual processes can keep pace with the technology they're meant to govern, and they can't.