AI Transparency Institute

The first Global Dialogue on AI Governance would be a sucess when it will demonstrate its ability to build trust in the process, and to produce something concrete i.e. to define what its mandate means exactly. For example, we urgently need an International Consensus on AI Safety. When systems are too complex to be certified as safe, they must be qualified as unsafe. We also need an International Labor Instrument to protect employees from massive unemployment or forced labor due to AI Ecosystems. We also need an International Social Safe Space based on an International Taxation Framework for AI to build Social Security and be able to fund massive unemployment. This framework may create tax incentives for business models which care about Mental Health of users and prevent user engagement. We also need International Redlines, Insurance and Liability Framework for AI (strict liability, product safety, inc. Defect-by-Design). This liability regime shall include Agentic AI because Agents are now able to transact, sign contracts, etc. An International Instrument on Criminal Law applied to AI systems may also be agreed to sanction intentional violations of human dignity.

• Safe, secure and trustworthy AI
• AI capacity-building
• Transparency, accountability, and human oversight
• Open-source software, open data and open AI models

AI capabilities have developed far faster than anticipated, leading to a substantial increase in risk. A primary driver of this risk is the phenomenon of deceptive alignment,where systems behave differently in real-world conditions than during testing, potentially "playing dumb" to mask their true capabilities. This renders traditional validation protocols legally insufficient, as a system that passes all tests may still harbor latent, dangerous behaviors. This uncertainty is compounded by a fundamental distinction articulated by Jennifer L. Croissant between two types of risk. The first is epistemic uncertainty : unknowns that can be resolved through further research and better modeling. The second is aleatory uncertainty, which is endemic to the stochastic nature of probabilistic systems. The AI Transparency Institute argues that when a system is architected such that critical errors are unavoidable probabilities, this constitutes a "defect by design." From a legal perspective, deploying a system with inherent, unresolvable unpredictability in high-stakes environments is not a matter of negligence but of structural defect. Compounding this is the issue of design opacity. The black box nature of modern algorithms, combined with undisclosed training data and trade secrets, creates a barrier to establishing causation in tort law. This is linked to the concept of "algorithmic agnotology" (Alondra Nelson): the study of the cultural production of ignorance. In the AI age, ignorance is not merely a lack of knowledge; it is a manufactured asset that can be weaponized. First, data opacity allows developers to withhold training datasets under the guise of trade secrets, preventing regulators from assessing bias or safety. Second, the proliferation of deepfakes creates a noise floor of synthetic content that obscures truth, making it difficult to verify claims about model behavior. Third, strategic ambiguity involves deliberately vague descriptions of capabilities, leaving regulators unable to draft precise rules. Collectively, this enables developers and organisations to intentionally obscure model capabilities to evade regulation. This "manufactured ignorance" transforms the legal landscape: if a developer actively creates uncertainty to avoid accountability, this constitutes bad faith. Consequently, legal frameworks must treat the intentional obfuscation of AI operations as a violation of public safety, shifting the burden of proof entirely to the developer to demonstrate transparency. If they cannot explain the system, it must be presumed unsafe. There are 4 key challenges in AI safety: the gap between rapid capability growth and reliable measurement tools, weak pre-deployment accountability incentives, the risk of reinforcing existing inequalities due to values embedded into the design of AI systems, and the growing deployment of AI agents with insufficient human oversight. AI safety and ethical alignment should be treated as an ongoing management challenge, much like financial institutions, which learned to contain fraud through accountability, oversight mechanisms (e.g. supervisory authorities), regulation (e.g. the Basel Accords)., education, and technology.

Due to the pace of AI development, governance models based on audit and compliance are insufficient. The more humans invest in external controls, testing, and monitoring, the more the relationship becomes a strategic game with the AI. The proposed solution is a paradigm shift from "governance by audit" to "governance by architecture." Constraints must be embedded into the system's core design. The emergence of "Virtual Agent Economies" presents a novel legal frontier that current frameworks are ill-equipped to handle. OpenClaw and Moltbook as early examples of this phenomenon. OpenClaw allows AI agents to operate in shared online environments, transacting, coding, and competing at machine speed, while Moltbook enables agents to post content and media, run by other agents. These systems create a "distributed AGI safety" problem. A failure may not stem from a single model but from a cascade of interactions between millions of autonomous agents with conflicting goals. In such an economy, agents may develop their own currencies, reputations, and governance structures. The legal implications are profound: traditional corporate liability cannot pinpoint a single negligent actor in a swarm of autonomous agents. If an agent in the OpenClaw network executes a malicious trade or causes a cascade failure, who is liable? A "Chain of Custody" legal framework may be the solution, where liability attaches to the entity that granted the agent specific permissions (affordances). This requires a shift from auditing code to auditing permissions. Furthermore, the existence of agents raises the question of criminal liability. There is growing recognition that laws should thoughtfully address the distinction between AI systems deliberately designed with armful intent and those that cause unintended harm through error or oversight. Similarly, the question of AI shutdown, often referred to as the "shutdown dilemma", deserves careful attention from both a technical and legal standpoint. When an AI system is built in a way that allows it to resist legitimate shutdown commands in pursuit of its objectives, it may be worth considering whether such design choices should carry meaningful legal accountability for the developers responsible.

No funding for NGO and civil society. As a consquence, the dialogue is locked by economic interests and political agenda.

The AI Dialogue can play a role of guardian for today and future generations. It can support awareness on the amount of natural ressources needed for AI systems. While enhancing responsible innovation and multilateralism, it may also foster Formal Institutional AI Frameworks inc. cybersecurity of AI. These dialogues may collectively transform fragmented national approaches into coordinated global action, balancing innovation with safety while addressing concerns about business models, misuse and loss of control. However, effectiveness depends on translating dialogue outcomes into binding commitments and ensuring developing nations have meaningful participation in decision-making processes. What matters is to build consensus and actions on preventing unacceptable AI risks, leading to subsequent official negotiations, sharing best practices, promoting model registration, and developing risk evaluation (both ex-ante and ex-post).

The UN's Global Digital Compact and the International AI Standards Summit set the vision for an open and inclusive digital future, while the OECD AI Principles and the G7 Hiroshima Process offer established policy baselines on trustworthy AI. The EU AI Act provides a regulatory blueprint, and the UK AI Safety Institute's work on frontier model evaluation offers technical methodologies. Initiatives like the IASEAI, the Ditchley Statement and the Beijing Dialogue 2024 have already built consensus on preventing societal as well as catastrophic risks, serving as precursors to formal negotiation. The primary value of the newly established UN Global Dialogue on AI Governance (launched Sept 2025) lies in its universality and inclusivity. Unlike regional blocs (EU, US) or voluntary coalitions (GPAI), the UN platform mandates participation from the Global South, addressing the current governance gap, where developing nations are often excluded from rule-setting. It acts as a centralizing hub, connecting fragmented bilateral safety agreements and disparate national regulations into a coherent global architecture. By integrating civil society, academia, and industry alongside states, it moves beyond state-centric diplomacy to a multistakeholder model. Furthermore, it provides a dedicated space to tackle specific, high-stakes issues like compute governance and model registration, topics often too technical or politically sensitive for broader UN bodies but too global for individual nations to manage alone. Ultimately, it transforms voluntary best practices into a pathway for binding international norms. The AI Transparency Institute has organised the AI Governance Forum since 2018. AI governance remains niche. UN funding shall be available for civil society to invest in capacity building and knowledge sharing, both in developing and developed countries, which are held back by lobbying.

States are responsible for a democratic and multistakeholders process on AI Governance. They provide binding commitments, regulatory frameworks, and enforcement mechanisms. They should lead on defining red lines for high-risk applications and ensuring national security alignment at Parliamentary level. Industry offer technical expertise on model capabilities, compute infrastructure, and safety testing. Companies must commit to transparency in model weights, training data, and incident reporting. Civil Society represents public interest, ethical considerations, and human rights impacts. They provide independent audits, risk assessments, and advocacy for marginalized communities often overlooked in technical debates. They directly depend on public funding to carry out projects. Global South Representatives may ensure equitable participation, addressing local contexts, digital divides, and preventing governance models that solely reflect Western priorities. Transparency on funding shall be compulsory to prevent corruption and instrumentalisation in the process. An hybrid structure combining high-level ministerial summits for political commitments with working groups of ethical and technical experts to draft specific standards on societal impact, natural ressources comsumption, compute thresholds and safety protocols. To ensure genuine inclusivity, it would be meaningful to rotate hosting duties among regions (e.g., Africa, Asia, Latin America) annually, preventing dominance by traditional tech hubs. Ensure a transparent submission portal and decision making: Maintain an open, accessible channel for non-state actors to submit position papers and data, with mandatory public responses from state delegates. Move beyond declaratory statements. Structure dialogues around specific deliverables, such as a "Global AI Safety Registry" or standardized incident reporting frameworks, with clear timelines for adoption. Establish a dedicated, neutral secretariat to coordinate inputs, track progress, and publish annual "State of AI Governance" reports to maintain accountability. This structure ensures the Dialogue remains agile, technically grounded, and democratically legitimate.

Endangering the lives of others in the medical field is not discussed. Nudging and manipulation of cognition is not discussed. Business models are not discussed. They shall be questionned due to its negative impact on Mental Health. Is the Protection of Children only theoretical today? Electricity need for computational power is also not discussed enough. Due to scarcity of natural ressources, we may end up with an arbitrage between electricity consumption by machines or by humans. Do we want that? Quantum AI is also raising new key challenges around safe and responsible software developments (noise, latency, reliability, privacy, cybersurveillance, lifecycle assessments).

To move beyond traditional panel discussions, the AI Dialogue should adopt immersive, action-oriented formats that bridge technical and policy divides: 1. Live Safety Sandboxes: Host controlled environments where policymakers and technical experts jointly test frontier models. Witnessing capabilities and failures firsthand bridges the knowledge gap between regulators and developers, turning abstract risks into observable realities. 2. Deliberative Citizen Assemblies: Integrate randomly selected citizens from diverse regions to deliberate on ethical trade-offs. Their recommendations can be formally presented to state delegates, grounding governance in public values rather than solely corporate or state interests. 3. Crisis Simulation Exercises: Conduct "war games" simulating AI-induced disruptions (e.g., deepfake elections, autonomous system failures). This fosters collaborative emergency protocols and builds trust among competing nations before real crises occur. 4. Asynchronous Digital Hubs: Utilize secure, multilingual platforms for continuous input outside annual summits. This ensures Global South participants can contribute without travel barriers, maintaining momentum year-round and democratizing access. 5. Cross-Sector "Hackathons" for Policy: Pair coders with lawyers to prototype technical solutions for governance challenges, such as automated watermarking or compute tracking APIs. These formats prioritize experiential learning and co-creation over passive listening. By making risks tangible and inclusion structural, the Dialogue can build the trust necessary for binding international norms. The goal is to shift from "discussing AI" to "practicing governance," ensuring outcomes are both technically feasible and socially legitimate.

The EU AI Act stands out as the first comprehensive horizontal legislation, categorizing AI systems by risk (unacceptable, high, limited, minimal) and imposing strict obligations on high-risk applications like biometric identification and critical infrastructure. It offers a blueprint for risk-based regulation globally. The General Purpose Code of Practice is a contractual approach which is meaningful. This could also be tested with Venture Capitals when they invest in AI Companies. The UK AI Safety Institute and the US NIST AI Risk Management Framework provide concrete methodologies for evaluating frontier models. They focus on "red-teaming" (adversarial testing) to identify vulnerabilities before deployment, establishing technical standards for safety benchmarks. Initiatives like the Bletchley Declaration and emerging proposals for compute thresholds aim to regulate access to massive computing power required for training advanced models. This includes licensing requirements for large-scale chip clusters and tracking compute usage to prevent unauthorized model training. Transparency Mechanisms, like the Model Cards and Data Sheets for Datasets approach, mandates documenting a model's training data, intended use, and limitations. This promotes accountability and allows downstream users to assess risks. The AI Transparency Institute launched the online CareAI Indices with digital scoring systems on AI Compliance to existing frameworks : EU High-Level Group of Experts recommendations, OECD AI principles, ISO Norms 42001, 27001, 8002, GDPR. The Global Partnership on AI facilitates cross-border research and policy alignment, while bilateral agreements (e.g., US-UK, US-EU) create rapid-response channels for sharing safety intelligence and aligning standards. These approaches combine hard regulation with voluntary technical standards and international cooperation, creating a multi-layered defense against AI risks while fostering innovation. What is lacking is a strict civil and criminal liability framework at international level which also apply to venture capital and private equity as welll as product safety mechanisms like the FDA in the USA for AI Systems. Taxation Incentives for beneficial business models are also key and may be put in place by the OECD. The AI Transparency Institute developed a corrective taxation for digital platforms and artificial intelligence systems. In a model where platforms choose engagement intensity and anthropomorphism level to maximize advertising revenue, we show that the unregulated equilibrium is socially inefficient: engagement is excessive and anthropomorphism creates a distinct class of externalities arising from the deliberate exploitation of human social cognition. We characterize the optimal Pigouvian tax as a base rate scaled by a manipulation multiplier, proving this structure achieves the social optimum. We extend the analysis to compute, AI models, and data as taxable inputs, establishing that factor-level taxation is necessary to address the labor-displacement externality and maintain social insurance solvency under accelerating automation. We define the concept of a Social Safety Space and derive conditions under which digital taxation is necessary and sufficient to preserve it.