The University of Queensland

The Dialogue would be a success if it catalysed concrete international action toward a global ban on the development of artificial superintelligence (ASI). At minimum, it should produce an enforceable global moratorium. ASI here means systems substantially smarter than humans across all cognitively relevant domains. This aligns with the recent Statement on Superintelligence, signed by leading AI scientists, Nobel laureates, and public figures. That statement calls for a prohibition on superintelligence development until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in. Specifically, success looks like the following. First, Member States acknowledge that ASI poses a plausible risk of human disempowerment or extinction. They acknowledge that this risk is qualitatively distinct from, and far more severe than, other AI harms. Second, there is a clear roadmap toward a binding international instrument. It should be modelled on successful precedents such as the Biological Weapons Convention, the Montreal Protocol, and IAEA safeguards. The instrument should halt or strictly licence frontier training runs above defined capability thresholds. Third, compute-governance and hardware-verification mechanisms are established to make non-compliance detectable. Fourth, capability evaluations, incident reporting, and red-lines are agreed. Red-lines include autonomous replication, AI R&D automation, and large-scale cyber-offence. Crossing them should trigger coordinated pauses. Fifth, a standing scientific body is established, modelled on the IPCC, to maintain expert consensus on frontier risks. Success is not another declaration that lists every conceivable AI issue with equal weight. Diluting the agenda across dozens of concerns guarantees inaction on the one risk that could be civilisation-ending. The Dialogue's distinctive value is its ability to convene States on the catastrophic and existential tail. These are issues that no single jurisdiction can address alone. Other AI harms, while real, are tractable through existing domestic regulation, sectoral law, and standard-setting bodies.

Safe, secure and trustworthy AI

I selected "Safe, secure and trustworthy AI" as the single priority. It is the only theme that directly encompasses catastrophic and existential risks from frontier and superintelligent AI. This is the one class of risk that demands urgent, coordinated international action. It is also the one class of risk that cannot be managed by any State acting alone. I deliberately did not select the other themes. This is not because they are unimportant. It is because they are tractable through existing institutions. Capacity-building is being progressed through UNESCO, ITU, and development banks. Human-rights protections are progressing through OHCHR and existing human-rights instruments. Transparency and accountability are progressing through domestic regulators and emerging standards such as ISO/IEC 42001 and the NIST AI RMF. Open-source and interoperability questions are progressing through technical standards bodies. A UN-level Dialogue that spreads attention thinly across all seven themes will, in my assessment, fail to produce the binding coordination needed on the one issue where coordination is indispensable. That issue is preventing the development of artificial superintelligence before humanity can demonstrably control it. The recent Statement on Superintelligence was signed by Turing and Nobel laureates, former heads of state, and senior figures across the political spectrum. It reflects rapidly converging expert opinion that ASI could, within years, pose risks of large-scale human disempowerment or extinction. Surveys of AI researchers, including the SARA project I led informing Australia's Safe and Responsible AI discussion paper, support this view. So does the MIT AI Risk Repository. This concern is not fringe. "Safe, secure and trustworthy AI" must therefore be interpreted by the Dialogue as centrally including frontier-model risk governance, compute oversight, evaluation regimes, and a credible path to a global ban or moratorium on ASI. Treating it as a generic checklist of product-safety concerns would miss the historical moment.

Yes. The single most important cross-cutting issue not adequately captured is the governance of artificial superintelligence (ASI) and the catastrophic and existential tail of frontier-AI risk. The seven listed themes implicitly assume AI systems remain tools used by humans within existing institutional frameworks. They do not address the scenario, taken seriously by a growing share of leading AI researchers, Nobel laureates, and the signatories of the Statement on Superintelligence, in which AI systems become more capable than humans across virtually all cognitive tasks. In that scenario humanity could lose meaningful control. A few specific gaps the Dialogue should explicitly address: 1. A global ban or moratorium on superintelligence development. It should hold until there is (a) broad scientific consensus that it can be built safely and controllably, and (b) strong public buy-in. This should be the Dialogue's flagship deliverable. 2. Compute and hardware governance. Frontier training runs are detectable and verifiable in ways biological or chemical research often is not. International agreements on chip export controls, on-chip governance mechanisms, and registration of large training runs are technically feasible. They should be prioritised. 3. Red-lines and capability thresholds. There should be agreement on dangerous capability evaluations. These include autonomous self-replication, automated AI R&D, large-scale cyber-offence, and uplift to CBRN weapons. Crossing thresholds should trigger coordinated international pauses. 4. An IPCC-style scientific body for frontier AI risk, building on the International AI Safety Report. It should have a clear mandate to track progress toward dangerous capabilities. 5. Liability and accountability for developers of frontier systems whose deployment causes mass harm, including catastrophic and irreversible harms. Other AI harms, including bias, labour displacement, misinformation, and privacy, are serious but manageable through ordinary domestic policy. ASI is not. The Dialogue's comparative advantage is precisely on the issues that no nation can solve alone. Please use it accordingly.

Australia and the wider Asia-Pacific are exposed to frontier-AI risk without meaningful sovereign control over the technology's trajectory. The frontier is set by a handful of US (and some Chinese) labs racing toward systems that could, on credible timelines, exceed human capability across all cognitive domains. The most significant governance gap is the absence of any international mechanism to slow or halt that race. Domestic regulation is necessary but insufficient. Australia's Voluntary AI Safety Standard, the proposed mandatory guardrails for high-risk AI, and the Safe and Responsible AI process are all useful. None of them bind labs in California or Shenzhen. If artificial superintelligence is built without adequate control, no Australian regulation will protect Australians from dying. The most significant challenge is strategic. Middle powers like Australia have a strong national interest in a binding international moratorium on superintelligence development. They have limited individual leverage. We are price-takers on a technology that could determine whether humanity retains meaningful control of its future. More immediate concerns include AI-enabled cyber-offence against critical infrastructure, large-scale disinformation across regional election cycles, and erosion of an already fragile information environment. The most significant opportunity is that the same middle-power position makes Australia and its Asia-Pacific neighbours natural conveners. We are trusted by both blocs. We host significant compute and critical-minerals supply chains. We have credible scientific institutions, including CSIRO and the AI Safety Institutes being stood up across the region, that could anchor verification regimes. The AI Dialogue is an opportunity for our region to push hard for a global ban or moratorium on superintelligence development. For enforcement, that should be paired with compute governance and capability-evaluation regimes. This is where the Asia-Pacific can punch well above its weight. The cost of inaction is unbounded.

The AI Dialogue's most valuable role is to be the venue where Member States agree, for the first time, that the development of artificial superintelligence is a matter of international peace and security. Other forums can handle most AI policy questions. Only the UN can confer the legitimacy needed for a binding global instrument that constrains the most powerful States and their frontier developers. Concretely, the Dialogue can advance international cooperation in several ways. First, by setting a shared problem definition. The Dialogue can establish that catastrophic and existential risks from frontier AI are within scope of international cooperation. It can establish that the precautionary principle applies. The recent Statement on Superintelligence, signed by Nobel laureates and senior figures across the political spectrum, gives the Dialogue political cover to take this view. Second, by sequencing. The Dialogue can prioritise issues by tractability and stakes. A global ban or moratorium on superintelligence development should be the headline outcome. Compute governance, capability evaluations, and red-lines are the supporting infrastructure. Other AI governance issues, while real, can follow. Third, by building the institutional plumbing. The Dialogue can mandate an IPCC-style scientific body for frontier AI risk. It can support an international AI Safety Institute network. It can task the Secretary-General with proposing a verification regime for compute, training, and deployment. Fourth, by aligning incentives. The Dialogue can encourage State and developer commitments to share safety information, to halt training runs that cross agreed red-lines, and to accept independent third-party evaluations. Fifth, by including underrepresented States. Many countries that will bear the consequences of frontier AI have no seat at private governance tables. The Dialogue is one of the few places where they have voice. That legitimacy is what gives a future moratorium the durability to last.

The Dialogue should connect with, and build on, the following initiatives. On scientific consensus: the International AI Safety Report led by Yoshua Bengio, the network of national AI Safety Institutes (UK, US, Japan, Singapore, Canada, EU, and equivalents being stood up across the Asia-Pacific), and the OECD AI policy work. These provide the technical foundation for any binding instrument. On political agreements: the Bletchley, Seoul, and Paris AI Safety Summits and their declarations; the G7 Hiroshima AI Process; the Council of Europe Framework Convention on AI; the EU AI Act; Califonia's SB53; New York's RAISE; the African Union Continental AI Strategy; the ASEAN Guide on AI Governance and Ethics; and existing UNESCO Recommendations on the Ethics of AI. On precedents for catastrophic-risk governance: the IAEA, the Biological Weapons Convention and its proposed verification regime, the Chemical Weapons Convention and the OPCW, the Montreal Protocol, and ICAO. These show that intrusive verification of dangerous technologies is achievable when States accept the stakes. On civil society and academia: the Future of Life Institute, the Center for AI Safety, the MIT AI Risk Repository (which I co-author), the AI Risk Index, BlueDot Impact's AGI Safety Fundamentals programmes, and academic safety groups at MILA, Berkeley, Oxford, Cambridge, and elsewhere. These bodies host the technical expertise the Dialogue will need. The Dialogue's distinctive added value is fivefold. First, universal membership and political legitimacy that no summit or club can match. Second, the ability to bind States, not just developers, through international law. Third, a permanent home for frontier-AI risk on the international agenda, not a one-off summit. Fourth, an inclusive forum that gives Global South States, who will live with the consequences, a real voice. Fifth, the convening power to force a hard conversation that voluntary commitments and corporate self-governance have so far avoided: should superintelligence be built at all, and on whose terms?

Different stakeholders should play distinct, complementary roles, with the Dialogue's structure designed to keep frontier-AI risk at the centre rather than diluted across every interest group's preferred topic. Governments should set the agenda, negotiate binding instruments, and bring the political authority that no other actor has. They should be encouraged to send delegations that include both diplomatic and technical staff, with explicit briefs from heads of government on catastrophic-risk priorities. Frontier developers should attend, but with disclosure obligations. Their voluntary commitments have not been sufficient. They should be expected to share information on training runs above defined compute thresholds, on capability evaluations, and on incidents. They should not be permitted to set the agenda, given the conflict of interest. Independent scientists and AI safety researchers, including those working on alignment, evaluations, and governance, should have privileged input. They are the equivalent of the IPCC's working groups and should be treated as such. National AI Safety Institutes are the natural channel. Civil society, including human-rights, labour, and digital-rights organisations, should be funded to participate, especially from low- and middle-income countries. Affected communities, including workers in sectors facing rapid automation and citizens of countries with limited frontier-AI capacity, should have dedicated channels. Recommended structure: 1. A standing plenary of Member States with annual sessions. 2. A scientific advisory body modelled on the IPCC, producing periodic reports on frontier-AI risk and capability trajectories. 3. Working groups on (a) a moratorium or ban on superintelligence, (b) compute and verification, (c) red-lines and capability evaluations, and (d) liability and incident response. 4. A multi-stakeholder consultative forum that is inclusive but advisory, not agenda-setting. 5. Regional preparatory meetings to build Global South capacity for substantive participation. 6. A clear two-year deliverable: a draft framework convention text.

Several voices are systematically underrepresented in current global discussions on AI governance, and the gap matters most for the catastrophic and existential risks that will affect them disproportionately. First, the Global South. Countries in Africa, South Asia, Southeast Asia, the Pacific, and Latin America are largely absent from frontier-AI governance forums dominated by the US, EU, UK, and China. They are also the States least able to absorb shocks from AI-enabled cyber-offence, labour displacement, disinformation, or a loss-of-control scenario. The Dialogue should fund delegate participation, run regional preparatory meetings, and translate technical materials. Second, independent AI safety researchers outside the major labs. Frontier developers have strong incentives to downplay extreme risks. The scientists who study alignment, evaluations, and dangerous capabilities, often in academia or non-profits, should have privileged channels equivalent to the IPCC working groups. Third, future generations. The decisions taken now about superintelligence will affect people not yet born. Existing institutions have no formal mechanism to represent their interests. Some States have begun appointing future-generations commissioners. The Dialogue could invite them to participate or could establish an analogous standing role. Fourth, workers and communities facing rapid automation. Trade unions, vocational educators, and affected communities are largely absent from international AI fora. Fifth, AI safety advocates from non-Western traditions. Concerns about powerful technologies and human flourishing are present in every philosophical and religious tradition. Engaging these voices builds legitimacy and reduces the perception that frontier-AI risk is a niche Western preoccupation. Sixth, smaller frontier developers and open-source contributors who often see governance gaps that incumbents prefer not to surface. Inclusion mechanisms: dedicated funding for travel and translation, regional consultations, observer status for relevant civil-society networks, formal IPCC-style scientific working groups, and a public submission process with transparent published responses.

Innovative formats should be judged by one criterion: do they help States reach a binding agreement on frontier-AI risk faster and with broader legitimacy than business-as-usual diplomacy? Several formats meet that test. 1. Structured expert elicitation. Use Delphi processes and forecasting tournaments to surface expert views on capability timelines, dangerous-capability thresholds, and risks from superintelligence. This produces transparent, traceable inputs to political deliberation. I have used these methods in the SARA project and in published Delphi work on crisis exercises (Wallace et al., 2025, Safety Science). 2. Scenario and tabletop exercises. Run cross-government tabletop exercises on plausible frontier-AI incidents, including loss-of-control scenarios, model-enabled cyber-offence, and rapid capability gains. Tabletops align officials' mental models faster than briefings and create relationships that pay off in a real crisis. See, for example https://www.intelligencerising.org/ or https://ai-2027.com/about?tab=tabletop-exercise#tab-box-tabletop-exercise 3. Capability demonstrations. Have national AI Safety Institutes present live or recorded demonstrations of frontier capabilities and red-team findings. Concrete demonstrations move the conversation from abstract to operational. 4. Rotating co-chairs from the Global South alongside major powers. Rotating leadership signals that the Dialogue is not a club of incumbents. 5. Public consultation with transparent published responses. Invite written submissions from civil society, academia, and industry. Publish them. Require the Secretariat to respond to substantive points. This is what the IPCC and major regulators already do. 6. Citizen assemblies on superintelligence. Convene representative citizen panels in willing Member States to deliberate on whether and how superintelligence should be developed. The Statement on Superintelligence calls for strong public buy-in. Citizen assemblies are the established way to test for it. 7. Live capability and incident reporting. A standing dashboard of frontier-model incidents, near-misses, and capability evaluations would give the Dialogue a shared evidence base updated continuously rather than every two years.

The most useful precedents are international agreements where humanity decided that some technologies are too inherently dangerous to develop, deploy, or commercialise, and coordinated to prohibit them. These offer the right template for superintelligence. 1. Human reproductive cloning. The UN Declaration on Human Cloning (2005) and dozens of national bans show that States can coordinate to prohibit a specific application even where the underlying science continues for legitimate purposes. The same logic applies to superintelligence: prohibit the dangerous frontier while permitting beneficial narrow AI. 2. Biological Weapons Convention (1972). A near-universal ban on an entire class of technology. Its weakness is verification, which is precisely the problem the AI Dialogue should solve from the outset by leveraging the verifiability of frontier compute. 3. Chemical Weapons Convention and the OPCW. Demonstrates that intrusive, technical verification of dangerous-technology bans is achievable when States accept the stakes. The OPCW won the Nobel Peace Prize for this work. 4. CCW Protocol IV on Blinding Laser Weapons (1995). A pre-emptive ban agreed before the technology was widely deployed. This is the closest precedent for banning superintelligence before it exists. 5. Montreal Protocol (1987). Coordinated phase-out of ozone-depleting substances, with trade measures against non-parties. Shows enforceable global coordination is possible on a tight timeline when the science is clear. 6. Nuclear non-proliferation regime and IAEA safeguards. The model for compute and training-run verification. Useful but insufficient: the EU AI Act, NY's RAISE Act, and California's SB 53. These advance transparency, incident reporting, and safety frameworks for frontier developers. They are not bans on developing superintelligence and should not be mistaken for one. They regulate how frontier AI is built; they do not address whether superintelligence should be built at all. The Dialogue's job is the latter.