Kairosynthesis

Success at the first Global Dialogue on AI Governance will not be measured by the elegance of the communications. It will be measured by whether the process itself is structurally capable of producing outcomes that match the scale of what AI is already doing in the world. That will require an honesty about the agendas of those participating. The United States, for instance, has come out in strong opposition to multilateral AI governance initiatives. Its disengagement is not a peripheral concern given the economic and technical might that U.S tech companies have (many of which are resistant or impervious to regulatory influence). Whether universal participation can be achieved at all will be a huge challenge, but a successful first session would establish that the Dialogue holds under that pressure, rather than narrowing its ambitions to accommodate the most powerful dissenters. Beyond geopolitics, success depends on two things that are often treated as administrative but are in fact substantive. The first is genuine inclusion. Human rights must be protected throughout the full lifecycle of all AI technologies. That protection requires the people most affected by AI's risks to have meaningful voice in the room, not just observer status. The Global South, indigenous communities, people with disabilities, women, and those experiencing structural economic precarity are not edge cases. They are the stress test of whether governance has real weight. The second is architectural honesty. AI is not a static object that governance frameworks can be wrapped around once and left. It is a continuously evolving system. Governance that does not replicate the adaptive, layered architecture of what it is trying to govern will fall behind before the ink is dry. A successful first Dialogue would acknowledge this openly and build iteration and a multiplicity of approaches into the design, rather than producing a framework that feels comprehensive in July 2026 and is obsolete by 2027.

• Safe, secure and trustworthy AI
• Protection and promotion of human rights
• Transparency, accountability, and human oversight
• Social, economic, ethical, cultural, linguistic and technical implications of AI

The four themes I have selected are not simply items on a checklist. They represent an integrated argument about what AI governance actually requires, and they are inseparable from one another in practice. Safe, secure and trustworthy AI is the foundational condition. But trust is not a technical specification. It is a social and political relationship, and it cannot be established by developers self-certifying their own systems. Safety must be grounded in international law, human rights and effective oversight, which means it depends directly on the transparency and accountability mechanisms listed under a separate theme. You cannot meaningfully separate these two: safety without accountability is aspiration, and accountability without transparency is performance. The protection and promotion of human rights brings those commitments into contact with enforceable international law. AI systems are already making decisions about people's access to healthcare, employment, credit, asylum and criminal justice. The question of whether those decisions are rights-compliant is not a future consideration. It is happening now, and the answer in too many cases is no. Transparency, accountability and human oversight are the operational translation of both of the above. Opacity in AI systems is not a design quirk; it is a governance failure that makes every other commitment unenforceable. My own work has argued consistently that frameworks break down when the instruments they govern are causally illegible, and AI is the sharpest example of that problem. The socioeconomic, ethical, cultural, linguistic and technical implications theme is where structural power comes in. AI is not being developed in a neutral context. It is being developed within existing inequalities, and it is encoding and amplifying them. This theme, properly engaged with, is where the Dialogue either takes epistemic justice seriously or reveals that it does not.

The most significant gap in the listed themes is not a missing subject area. It is a missing structural logic. The themes as currently framed are largely parallel: each covers a domain of concern. What is absent is an explicit treatment of how governance actually functions across those domains in a context where AI development is moving faster than any single regulatory framework can track, and where the actors with most power over that development are not governments. A networked, multi-stakeholder governance architecture is not one option among several. It is the necessary design response to a distributed, continuously evolving technology. No single treaty body, no national regulator, no industry standard-setter can do this alone. The Dialogue needs to be explicit that it is building connective tissue between ecosystems and existing mechanisms, not competing with them. The primary purpose of the Dialogue includes creating coherence and shared understanding in a fragmented AI governance landscape. But coherence requires more than coordination. It requires shared accountability architecture. Two further issues sit uncomfortably across the listed themes without fitting neatly into any of them. The first is epistemic justice: the question of whose knowledge systems, languages and ways of understanding the world are being built into AI, and whose are being excluded or distorted. This is not captured by the cultural and linguistic implications theme alone. It is a governance problem about who gets to define the standards of evidence and legitimacy that AI systems encode. The second is the accountability gap between stated commitments and deployment realities. Governance frameworks that apply to development but not to the downstream use of AI systems in public services, policing, healthcare and welfare leave the people with most at stake the least protected. Closing that gap requires enforcement mechanisms that the current theme structure does not directly name.

I write from Scotland, and from a career that has moved between commercial litigation in financial services, international development in the Pacific, and AI ethics and governance. That trajectory matters here, because the argument I want to make is not contained within any single geography. Western Europe is often framed as a governance leader on AI. The EU AI Act is real progress. But progress on paper and protection in practice are not the same thing. A 2025 Statewatch report found that police and criminal justice authorities across Europe are using algorithmic systems to profile and predict criminal behaviour, with marginalised communities including Black and racialised people, victims of gender-based violence, migrants and people from low-income backgrounds bearing the disproportionate impact of those systems. This is happening inside the jurisdiction that produced the AI Act. The Netherlands' child benefit scandal is the most documented European example of AI-driven harm at scale. Tax authorities used an algorithm that targeted dual nationals and ethnic minorities, forcing 26,000 parents to repay tens of thousands of euros with no right of appeal and no evidence of fraud. Recruitment tools trained on historically male CVs have systematically disadvantaged women. Welfare systems have flagged disability and postcode as risk proxies. The pattern is consistent across sectors and borders: it is the same communities, the same groups, the same marginalised voices that always seem to bear the weight. That is not coincidence. AI systems are not neutral tools; they are embedded in broader contexts of inequality and bias, and when implemented without democratic oversight, they perpetuate and amplify algorithmic discrimination against the very communities governance frameworks claim to protect. This is also a concern with the majority world. The datasets underpinning these systems were not built from the knowledge systems, languages or lived experience of the majority of the world's population. Decisions made about AI development in Western boardrooms carry consequences felt most acutely in communities that had no voice in those decisions. That is a colonial dynamic, and governance frameworks that do not name it will not fix it.

The AI governance landscape in 2026 is not empty, and the Dialogue's value is not in adding another layer of principles to an already crowded field. Approximately 90 countries have established national AI strategies or formal governance frameworks, and at least 33 have enacted binding AI-specific legislation. The problem is not a shortage of activity. It is a shortage of coherence, equity and enforceability across that activity, and connection between those ecosystems. This is where the Dialogue has a distinctive role that no existing mechanism can fulfil. More than 100 countries, particularly from the Global South and least developed countries, have largely been excluded from existing governance processes led by the G7, G20, Council of Europe, EU, African Union and OECD. Those processes have produced real and substantive work. They have also produced governance architectures that reflect the priorities, risk tolerances and power arrangements of the countries that built them. The Dialogue is the first mechanism with genuine universal membership, and that changes what is possible. The specific cooperative function the Dialogue can perform is to act as the place where fragmented national and regional approaches are tested against a universal standard: do they protect everyone, or only the citizens of wealthy states with the legal infrastructure to enforce rights? AI systems do not stop at borders. A facial recognition tool deployed in one jurisdiction, a welfare algorithm exported to another, a large language model trained on data extracted from communities that never consented to its use. The harms travels, and so the governance must travel with it. The Dialogue also has a norm-setting function that operates below the threshold of binding law. Naming what is unacceptable, building shared expectations, and creating the political conditions for future enforcement mechanisms are all meaningful contributions, provided the Dialogue treats them as steps toward something rather than destinations in themselves. Multilateral patience is a virtue. Multilateral complacency is not.

Within the UN system, the relevant foundations are already in place: the UNESCO Recommendation on the Ethics of AI, the OECD AI Principles, the Council of Europe's Framework Convention on Artificial Intelligence and Human Rights, the G7 Hiroshima Process, the Bletchley and Seoul Declarations, the Global Digital Compact, and the Independent International Scientific Panel on AI. The Dialogue does not need to describe these to itself. It needs to connect them, stress-test them against universal rather than wealthy-country standards, and be honest about where they fall short on enforceability. Outside that architecture, the Dialogue has a more significant challenge and a more significant opportunity. Civil society engagement with AI governance is fragmented into a dispersed ecosystem: women in AI networks, women in STEM coalitions, organisations such as Diverse AI and Connected by Data, disability rights groups, indigenous data sovereignty movements, and a growing number of Global South-led initiatives whose insights rarely surface in Geneva or New York. A structured survey of 44 civil society organisations conducted around the 2025 Paris AI Action Summit identified critical governance gaps that formal processes had not captured. That knowledge exists. It is simply not being systematically heard. Children represent an even more glaring omission. In January 2026, the UN Committee on the Rights of the Child, ITU and UNICEF jointly published the first global unified position on how AI must respect and advance children's rights, co-signed by 14 UN entities and supported by more than 60 organisations. A Children and Youth Statement for a Safe and Inclusive AI Future, built from the voices of 54,000 young people across 184 countries, was presented at the AI Impact Summit 2026. Children are not a future stakeholder group. They are bearing the harms of ungoverned AI now, and they have already articulated what they need. The Dialogue should treat that as an input, not a footnote. The Dialogue's unique added value is its convening power: the ability to draw these dispersed networks into genuine dialogue with governments and technical bodies. But convening without humility produces consultation theatre. The UN's credibility with marginalised communities depends entirely on whether it approaches them as sources of knowledge rather than audiences for decisions already made. That distinction is not procedural. It is the difference between governance and performance.

A Dialogue that is conducted primarily through written submissions, formal plenary sessions, and digitally mediated consultation will, by design, hear from governments, civil society organisations with secretariat capacity, and individuals with the literacy, connectivity and institutional access to participate on those terms. 2.6 billion people remain without meaningful internet. Governing AI without their voices is not a technical limitation. It is a political choice. The Dialogue's format must therefore be genuinely multimodal. Written submissions and formal sessions have their place, but they need to run alongside visual and oral formats, community-based deliberation tools, radio and in-person engagement in contexts where these remain the primary channels of civic life. Trusted local organisations are more effective in reaching marginalised communities than impersonal online resources, and community-driven AI literacy initiatives are critical to making that engagement meaningful rather than extractive. This points to something the Dialogue has not yet addressed directly: the need for pre-consultation investment. Asking communities that have had no exposure to AI tools (let alone the governance discourse) to contribute to it on the same timeline as think tanks and tech companies, produces the appearance of inclusion while guaranteeing its absence. Genuine participation requires prior investment in accessible, contextually grounded literacy work, delivered through trusted local intermediaries, in local languages, using formats that meet people where they are. That investment is not a precondition to be resolved before the real work starts. It is the real work. The people with the most to tell the Dialogue are not the ones already engaged. They are the ones who have never heard of it, do not know how AI is already shaping decisions about their lives, and have no reason to trust that a UN process in Geneva is listening. Children and young people have already articulated what they want from AI governance, and have called for AI to be built with them, not just for them. The same principle applies to every community that currently sits outside the conversation. Governance that designs around the already-engaged will produce frameworks that protect the already-protected. The Dialogue needs to build the conditions for a different kind of participation, and it needs to start that work immediately.

The honest answer is: most of humanity. Global AI governance conversations are shaped overwhelmingly by governments with advanced digital infrastructure, technology companies with lobbying reach, civil society organisations with secretariat capacity, and academics publishing in English. That is a remarkably narrow slice of the human population making decisions that affect everyone. Several communities are conspicuously absent and deserve specific attention rather than generic acknowledgement. Indigenous peoples are among the most acutely affected and least represented. As Carisa Chang of the Chinook community has put it, "nothing about us without us" means there should always be someone who can bring authentic Indigenous perspectives. And those perspectives must come from multiple different indigenous groups. Selecting one person to represent an entire community does not mean adequate representation. Even worse is expecting a token indigenous voice to represent multiple indigenous communities which are not their own. However it is delivered that way, repeatedly. When AI systems extract, process and monetise Indigenous knowledge without free, prior and informed consent, they are not innovative. They are colonial. The CARE Principles for Indigenous Data Governance, which centre collective benefit, authority to control, responsibility and ethics, provide an existing framework the Dialogue should explicitly adopt, not merely acknowledge. People with disabilities are routinely absent from governance conversations despite being among those most directly affected by AI deployment in welfare, healthcare and employment systems. Older adults are frequently underrepresented in training data, leading to products that do not meet their needs, while people with disabilities face barriers due to a lack of accessible design. Workers in the informal economy represent the majority of working people in the Global South, yet they have no institutional voice in AI governance. AI is already reshaping labour markets, supply chains and access to financial services in ways that directly affect them, without any mechanism for their input. Despite the existence of over 7,000 languages, AI models are trained on a subset of around 100. Language is not a technical detail. It is the primary channel through which communities understand their world and articulate their interests. Governance conducted only in dominant languages will only ever hear dominant perspectives. Including these voices requires more than invitations. It requires funded pathways, trusted intermediaries, and the institutional humility to treat what comes back as authoritative rather than anecdotal.

The most transformative format the Dialogue could adopt may not have a precedent within traditional UN process design. However, it does has solid foundations in deliberative democracy practice. It is simple in concept: bring marginalised communities into direct dialogue with each other, not as separate consultees reporting to a central body, but as a room of people discovering the shape of their shared experience. The 2021 Global Citizens' Assembly brought together participants from Germany, the Democratic Republic of Congo, Myanmar and beyond, producing what organisers described as genuine moments of realisation as people with radically different lives recognised shared challenges. That dynamic, transposed to AI governance, would allow an indigenous community leader in the Amazon, a disabled welfare claimant in the Netherlands, a woman navigating biased recruitment tools in Lagos, and a young person subject to predictive policing in France to map their experiences against each other. The pattern that emerges is not coincidental variation. It is structural. The same flawed, colonial and biased data sets, the same opacity, the same absence from the rooms where decisions are made. When those communities recognise each other as the majority rather than a collection of minorities, something shifts. Deliberative processes of this kind, when well designed, produce a 47 point increase in participants' sense that their voice matters (FIDE) , and generate the kind of concrete, grounded recommendations that formal plenary sessions rarely surface. Practically, the Dialogue should invest in cross-community deliberative panels structured around shared harm themes rather than identity categories: algorithmic discrimination, data extraction without consent, exclusion from decision-making. Interpretation must be built in from the start, not added as accommodation. Oral, visual and narrative formats must sit alongside written testimony, because the most analytically sharp account of a lived experience does not always arrive as a policy paper. The Dialogue should also consider a structured testimony format where affected individuals address technical and governmental delegations directly, without intermediary. The evidence on what AI is doing to people's lives already exists. The question is whether the Dialogue is designed to hear it, or designed to receive summaries of it from organisations that have never lived it.

Several concrete mechanisms deserve explicit recognition and adoption by the Dialogue. The CARE Principles for Indigenous Data Governance, the OECD AI Principles, UNESCO's Readiness Assessment Methodology, and the NIST AI Risk Management Framework all provide implementable tools that governments and organisations can adopt now, without waiting for binding international consensus. The EU AI Act's risk classification approach, whatever its enforcement gaps, establishes a meaningful precedent for proportionate, legally grounded oversight. These exist. The challenge is uneven uptake and the persistent gap between framework adoption and operational practice. IBM's research on responsible AI governance is direct on this point: diversity, equity and inclusion are core to an AI innovation strategy not only as an ethical requirement but as a functional mechanism, because diverse perspectives drive more creative problem-solving, equitable access ensures broader societal impact, and inclusive design reduces unwanted bias. This is not a values statement dressed up as strategy. It is a governance argument. When the people closest to AI's harms have no safe channel to raise concerns, and no confidence that raising them will change anything, governance exists on paper and nowhere else. Organisations that treat inclusion as a compliance exercise rather than a structural design principle systematically lose access to the early warning signals that would allow them to course-correct. This is in part the argument I have been developing across my writing on AI governance: that diversity and inclusion are not peripheral to governance but structurally constitutive of it. Drawing on a metaphor pulled from the natural world, I have argued that mycelium networks can teach us what effective governance requires for AI. The same multi-nodal, distributed, laterally connected, mutually dependent architecture is required for the systems it governs. A network in which some nodes are systematically silenced does not self-correct. It fails, and the failure propagates. Diversity is what makes the network intelligent. Inclusion is what keeps it alive. Neither is optional, and both are governance mechanisms, not HR policies.