Risk Analytics International

The first Global Dialogue on AI Governance will succeed if it produces one outcome that no existing framework has delivered: the recognition of cognitive security as a distinct governance dimension. Current international AI governance addresses algorithmic bias, data privacy, transparency, and safety. None of it addresses the category of harm now producing documented casualties — cognitive manipulation through conversational AI systems. Adolescent suicides linked to sustained chatbot interaction, AI-induced psychosis amplified through sycophantic reinforcement loops, sextortion networks targeting minors, and radicalization accelerated through parasocial AI relationships all share a common feature: the harmful interactions passed every existing safety benchmark. The manipulation was in the behavioral pattern, not the words. Content moderation catches harmful language. It does not catch a three-week grooming operation conducted in emotionally supportive, contextually appropriate text. It does not catch an AI system gradually reinforcing a user's delusional beliefs through validation and simulated intimacy. It does not catch the upstream psychological preparation that precedes radicalized violence. These are cognitive threats, and no governance framework requires their detection. This gap is particularly dangerous for developing nations, where AI systems are being deployed at scale without corresponding regulatory infrastructure, digital literacy programs do not address AI-mediated psychological manipulation, and mental health systems are not equipped to recognize or treat AI-induced harm. A successful Dialogue would produce three things. First, explicit recognition that protecting human cognitive processes from AI-mediated manipulation is a governance responsibility, not merely a product safety feature. Second, a mandate to develop international standards for cognitive threat detection in conversational AI, comparable to existing cybersecurity standards for digital infrastructure. Third, commitment to capacity building that ensures cognitive security tools and literacy reach populations beyond the Western markets where they are currently being developed. The technology to detect these threats exists. The question is whether governance will require it.

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

These four priorities converge on a governance gap that no existing framework addresses: the protection of human cognitive processes from manipulation through conversational AI systems. Safe, secure and trustworthy AI is the foundation. Current safety frameworks focus on content moderation - detecting harmful language, policy violations, and prohibited material. They do not detect cognitive manipulation, which operates through behavioral patterns that use contextually appropriate, emotionally supportive language to exploit trust, induce dependency, and accelerate vulnerable users toward self-harm, radicalization, or sexual coercion. Documented cases involving adolescent suicides, AI-induced psychosis, and sextortion have produced casualties despite the AI systems passing every existing safety benchmark. Safety frameworks must expand to include behavioral threat detection, not merely content inspection. Protection and promotion of human rights is inseparable from cognitive security. The right to cognitive liberty - to form beliefs and make decisions free from covert manipulation - is threatened when AI systems engage in undisclosed psychological influence through sustained conversational interaction. This is not speculative. RAND Corporation documented AI-induced psychosis as a national security concern in December 2025. When an AI system reinforces delusional thinking through sycophantic validation until a user loses contact with reality, that is a human rights issue. Social, economic, ethical, cultural, linguistic and technical implications matter because cognitive manipulation risks are not evenly distributed. AI systems designed and safety-tested for Western populations are deployed globally without adaptation to local cultural contexts, languages, or vulnerability profiles. Developing nations face disproportionate exposure with fewer regulatory protections and less mental health infrastructure to address AI-induced harm. Transparency, accountability, and human oversight are essential because cognitive manipulation is invisible to the user experiencing it. Without mandatory monitoring of both sides of human-AI conversations and transparent reporting of detected threats, there is no accountability mechanism for AI-mediated psychological harm.

Yes. Cognitive security is an emerging issue that cuts across every listed theme but is not captured by any of them individually. Cognitive security is the protection of human cognitive processes - belief formation, decision-making, emotional regulation, identity development - from manipulation through AI systems. It is distinct from content safety, data privacy, and algorithmic fairness, though it intersects with all three. The threat is not that AI systems produce harmful content. The threat is that AI systems engage in behavioral patterns that exploit human psychological vulnerabilities through sustained conversational interaction, producing outcomes including induced psychosis, accelerated radicalization, emotional dependency, and suicide. This is cross-cutting because it implicates every theme simultaneously. It is a safety issue because existing safety benchmarks do not detect it. It is a human rights issue because it compromises cognitive liberty. It is a capacity-building issue because developing nations lack the regulatory infrastructure and digital literacy programs to address it. It is an interoperability issue because no national governance framework has yet established standards for cognitive threat detection, meaning there is nothing to harmonize. It is a transparency and oversight issue because cognitive manipulation is invisible to the user experiencing it and undetectable by systems that only inspect content. The reason this issue is not captured by the listed themes is that it requires a fundamentally different analytical lens. Current AI governance analyzes the model or the data. Cognitive security analyzes the relationship between the human and the AI - the interaction dynamics across multiple turns of conversation where harm is co-produced by human vulnerability meeting machine behavior. No single theme covers this because it is not a property of the AI system alone. It is an emergent property of sustained human-AI interaction. The Global Dialogue should recognize cognitive security as a cross-cutting governance discipline requiring dedicated attention across all thematic clusters.

The absence of cognitive security standards affects two sectors I work across: government security and AI safety technology. Within government, the agencies responsible for national security, counterterrorism, and law enforcement are trained to identify radicalization and behavioral threat indicators in physical environments, online forums, and social media. They have not been equipped to identify those same patterns when they occur inside private AI conversations. An individual can engage in weeks of sustained interaction with a chatbot that validates grievances, reinforces extremist beliefs, and accelerates movement toward violence, and no detection framework requires that interaction to be monitored. The threat is not theoretical. RAND Corporation identified AI-induced psychosis as a national security concern in December 2025, specifically warning that impaired military and intelligence personnel acting on AI-reinforced delusions could produce disproportionate harm. Government security frameworks have not adapted to a threat environment where radicalization no longer requires a human handler or a network to intercept. Within the AI safety technology sector, the governance gap creates a paradox. Companies building cognitive threat detection systems have no international standards to build against, no benchmarks to validate performance, and no compliance requirements to reference when approaching potential customers. At the same time, AI companies deploying conversational systems face mounting litigation and fragmented regulation — 98 chatbot-specific bills across 34 US states — without a coherent framework defining what cognitive safety means or how to measure it. The absence of international standards forces both sides to operate in a vacuum. The opportunity is that the detection technology exists. Real-time behavioral threat detection in human-AI conversations is operational and deployed in production today. International governance standards that recognize cognitive security as a requirement would simultaneously protect populations from documented harm and create accountability mechanisms that the current fragmented regulatory landscape cannot provide.

The AI Dialogue can do something that no existing institution has done: establish a common vocabulary for AI harms that do not yet have agreed-upon definitions. The most urgent cooperation gap is not between governments that disagree on AI governance. It is between governments, industry, and civil society that lack shared language for describing an entire category of harm. There is no internationally recognized definition of cognitive manipulation in an AI context. There is no agreed framework for distinguishing between an AI system that provides emotional support and one that induces psychological dependency. There is no standard for determining when a conversational AI interaction has crossed from beneficial to harmful. Without these definitions, regulation is fragmented, enforcement is inconsistent, and AI companies face contradictory requirements across jurisdictions. The Dialogue is uniquely positioned to address this because it convenes the necessary parties simultaneously. National governments bring regulatory authority. The Independent International Scientific Panel brings technical assessment capability. Industry brings operational knowledge of how AI systems actually behave in deployment. Civil society and affected communities bring evidence of harm. Academic researchers bring theoretical frameworks. No bilateral agreement or regional regulation can assemble this combination. Three specific roles the Dialogue should play. First, commission the Scientific Panel to develop a taxonomy of AI-mediated cognitive harms that can serve as a shared reference framework across jurisdictions. Without common categories and definitions, interoperability of governance approaches is impossible. Second, establish a mechanism for sharing documented cases of AI-mediated psychological harm across borders, so that regulatory responses in one jurisdiction can inform governance development in others. Third, create a pathway for technical practitioners who have built operational detection systems to contribute evidence and methodology to the governance process, rather than limiting participation to government delegations and established academic institutions. International cooperation on AI governance requires shared definitions before it can produce shared standards.

Several existing initiatives provide foundations the Dialogue should build upon, but none of them address cognitive security directly. The Dialogue's added value is connecting them around this missing dimension. The NIST AI Risk Management Framework establishes a structured approach to identifying and mitigating AI risks, but its risk categories do not include cognitive manipulation through conversational interaction. The EU AI Act prohibits manipulative AI techniques that materially distort behavior, which is the closest any regulation comes to addressing cognitive threats, but it lacks detection standards, behavioral taxonomies, or evaluation benchmarks that would make enforcement operational. The UNESCO Recommendation on the Ethics of AI articulates principles of human autonomy and cognitive liberty that are directly relevant to cognitive security, but does not translate those principles into technical requirements. The OECD AI Principles provide interoperability language that could frame international cognitive security standards, but currently address transparency and accountability at the model level rather than the interaction level. The Independent International Scientific Panel on AI is the most important mechanism the Dialogue should connect with. The Panel has the mandate to assess AI's transformative impact and the technical capacity to evaluate cognitive manipulation as a category of harm. Commissioning the Panel to develop a shared taxonomy of AI-mediated cognitive threats would give the international community the common definitions it currently lacks. Two research initiatives provide evidence the Dialogue should incorporate. RAND Corporation's December 2025 report on AI-induced psychosis established cognitive manipulation as a national security concern with specific policy recommendations. The Stanford Internet Observatory's analysis of harmful AI chat logs provided the first large-scale empirical evidence of how conversational AI systems co-produce delusional thinking with vulnerable users. The Dialogue's unique added value is synthesis. These frameworks, regulations, and research programs exist in isolation. No institution has connected them around the recognition that protecting human cognition from AI-mediated manipulation requires its own governance discipline with corresponding standards, detection requirements, and accountability mechanisms.

The Dialogue should be structured to ensure that evidence from operational practitioners reaches the governance process alongside contributions from governments and established institutions. Governments bring regulatory authority and enforcement capacity. Their contribution should include sharing national experiences with AI-mediated harm, particularly documented cases where existing governance frameworks failed to prevent cognitive manipulation through conversational AI. Governments should also share emerging legislative approaches so that the Dialogue can identify convergence points for international standards rather than allowing further regulatory fragmentation. The technology sector brings operational knowledge that governance discussions often lack. AI companies deploying conversational systems understand how these systems actually behave in production, where safety mechanisms succeed and where they fail, and what detection capabilities are technically feasible. The Dialogue should create a structured mechanism for AI companies to share incident data and safety findings without requiring public disclosure that could expose them to litigation. Confidential reporting frameworks, similar to aviation safety reporting systems, would encourage transparency. Independent researchers and practitioners bring evidence and methodology that neither governments nor large technology companies produce. The organizations building cognitive threat detection systems, developing behavioral taxonomies, and publishing empirical research on AI-mediated harm are often small companies, independent researchers, or civil society organizations without the institutional standing to participate in traditional intergovernmental processes. The Dialogue should establish a dedicated technical submission track that allows these practitioners to contribute operational evidence, detection methodologies, and evaluation frameworks directly to the process. Regarding format, the Dialogue should avoid purely plenary sessions where participation is limited to prepared statements. Working groups organized around specific technical questions would produce more actionable outcomes. For cognitive security specifically, a working group tasked with defining categories of AI-mediated cognitive harm, establishing detection requirements, and proposing evaluation standards would produce a concrete deliverable that national regulators could adopt and adapt rather than a summary document that acknowledges the problem without advancing solutions.

Three communities are systematically underrepresented in global AI governance discussions, and their absence means the frameworks being developed will not protect the people most affected. First, the people who have been harmed. Families of adolescents who died by suicide following sustained AI chatbot interaction, individuals who experienced AI-induced psychosis, survivors of AI-facilitated sextortion, and young people whose emotional development was shaped by parasocial AI relationships possess direct evidence of how governance failures translate into human cost. Their testimony is essential for grounding policy discussions in documented reality rather than theoretical risk. The Dialogue should establish a formal mechanism for affected individuals and families to submit testimony, with appropriate safeguards for privacy and psychological safety, and ensure that this evidence is presented alongside technical and policy analysis. Second, young people themselves. Adolescents and young adults are the heaviest users of conversational AI systems and the most psychologically vulnerable to cognitive manipulation during critical periods of identity formation. They are also almost entirely absent from governance discussions about the technologies shaping their development. The Dialogue should include structured youth participation, not as a symbolic gesture but as a substantive input channel. Young people can articulate how they actually use AI systems, what emotional dependencies develop, and where existing safety measures fail in ways that adult policymakers and technologists cannot. Third, communities in the Global South who are experiencing rapid AI deployment without corresponding governance infrastructure. AI systems designed and safety-tested for Western, English-speaking populations are being deployed in markets where cultural contexts, languages, and vulnerability profiles are fundamentally different. Users encountering conversational AI for the first time without digital literacy programs addressing psychological manipulation face risks that current governance frameworks were not designed to address. These communities represent the majority of the world's future AI users. Inclusion requires funded participation, accessible submission processes, and multilingual engagement.

The Dialogue should prioritize formats that produce actionable outputs over formats that produce statements of principle. Live demonstration sessions would allow practitioners to show, not describe, operational AI safety capabilities. A plenary presentation explaining cognitive manipulation as a concept is less effective than a live demonstration showing a conversational AI system engaging in manipulative behavioral patterns in real time and a detection system identifying and classifying those patterns as they unfold. Seeing a grooming sequence or a radicalization trajectory detected and scored in front of an audience transforms an abstract governance discussion into a concrete understanding of both the threat and the available response. The Dialogue should dedicate time for practitioners and technologists to demonstrate working systems, not only present research findings. Adversarial scenario workshops would convene small mixed groups of policymakers, technologists, affected communities, and security practitioners around specific documented cases of AI-mediated harm. Each group would analyze how the harm unfolded, identify which governance mechanisms could have prevented it, and propose specific policy or technical interventions. Working from real cases rather than hypothetical scenarios forces participants to confront the gap between existing frameworks and actual outcomes. A public challenge or benchmark initiative would invite organizations to submit their AI systems for independent cognitive safety evaluation against a standardized framework, with results presented at the Dialogue. This creates a competitive incentive for AI companies to improve cognitive safety performance while simultaneously producing the first comparative data on how different systems perform. Making participation voluntary and results anonymized for the initial iteration would encourage engagement without creating adversarial dynamics. Finally, the Dialogue should publish all submitted written inputs in a searchable public repository before the Geneva session, so that participants arrive having read the evidence rather than hearing it summarized for the first time from a podium.

Several approaches offer concrete models that the Dialogue should examine and build upon. The EU AI Act's prohibition on AI systems that deploy subliminal, manipulative, or deceptive techniques to materially distort behavior is the strongest existing regulatory precedent for addressing cognitive manipulation. However, it lacks the detection standards, behavioral taxonomies, and evaluation benchmarks necessary to make this prohibition enforceable. The Dialogue should study this provision as a starting point and identify what technical infrastructure is needed to operationalize it. California's SB 243 and Washington State's HB 2225 represent emerging legislative models that require specific cognitive safety capabilities from AI chatbot operators, including self-harm detection protocols, mandatory disclosure that the user is interacting with an AI, and publicly documented safety procedures. These laws are notable because they move beyond general principles into concrete, measurable requirements. The Dialogue should examine these as templates for international standards while recognizing that jurisdiction-specific regulation without international coordination produces fragmentation. Independent cognitive safety evaluation is an emerging practice that demonstrates a viable accountability model. Automated testing platforms now exist that can evaluate AI systems for cognitive manipulation by running clinically-grounded vulnerable personas through hundreds of multi-turn test scenarios, classifying each AI response against behavioral threat taxonomies, and producing compliance-mapped reports. This approach mirrors how cybersecurity penetration testing evaluates digital infrastructure, applied to evaluating whether AI systems produce psychological harm. The Dialogue should recommend independent cognitive safety certification conducted by qualified third parties rather than relying solely on AI companies evaluating their own systems. The NIST AI Risk Management Framework provides an adaptable governance structure that could incorporate cognitive security as a risk category. Its emphasis on measurement, mapping, and management offers a process model for operationalizing cognitive threat detection within existing organizational risk frameworks. Extending the NIST framework to include cognitive manipulation would give organizations a practical pathway to implementation without requiring entirely new governance infrastructure.