Main Insight
At The Seventh Edition of Athens Roundtable in London in December 2025, a closed-door dialogue convened 76 experts to discuss operational AI red lines. Participants identified major risk clusters and emphasized enforceable technical thresholds, verification mechanisms, and multilateral standards grounded in existing international law to prevent severe harms from powerful AI.
Defining and Governing Unacceptable AI Risks
February 12, 2026
On December 4, 2025, leading global experts on AI governance gathered in London for the Seventh Edition of The Athens Roundtable on AI and the Rule of Law. While plenary discussions examined the broader evolution of AI governance, one of two closed-door dialogue sessions focused on a more specific and urgent set of questions:
Which AI behaviors and uses should be prohibited? What would meaningful red lines look like in practice? How could compliance be demonstrated? And what diplomatic mechanisms could support international agreement and verification?
This working session brought together policymakers, technical experts, diplomats, and civil society leaders to address what is, at its core, a global risk management problem. As AI systems grow more capable and more autonomous, the consequences of failure grow as well. The dialogue informed the report Defining and Governing Unacceptable AI Risks, which synthesizes insights from 76 experts working on the frontlines of AI safety and governance.
Several risk clusters emerged. These included large-scale manipulation and psychological exploitation, particularly where children are concerned; loss-of-control risks in advanced AI systems; chemical, biological, radiological, and nuclear (CBRN) threats; cyber-offensive capabilities; autonomous control of critical infrastructure; lethal autonomous weapons and escalation dynamics; and broader risks to human rights and justice.
Participants agreed that identifying and agreeing upon red lines is not straightforward, and is only the beginning. A central challenge is to make them operational, which involves translating high-level norms into clear, technically grounded thresholds that can be evaluated and verified across jurisdictions. This would require sustained multilateral processes, shared technical standards, and robust methodologies for assessing and validating safety claims. Systems that cannot demonstrate compliance with agreed safety limits should not be developed or deployed, and prohibitions should be accompanied with credible verification mechanisms and institutions.
There was also strong support for grounding this work in existing international law and the rule of law, rather than constructing entirely new legal architectures from scratch. Treaties governing weapons, human rights, and international security already define categories of unacceptable harm. Aligning AI governance with these frameworks strengthens legitimacy, accelerates coordination, and reinforces a simple principle: advanced AI systems must remain subject to the rule of law and to shared standards of human safety.
Read a summary of the Athens Roundtable session on AI red lines:
The session on red lines was part of The Athens Roundtable 2025, aimed at moving from shared concerns to joint action.
- Read the 2025 Athens Roundtable Report for a summary and takeaways from the full event.
- Read the AI Incident Prevention & Preparedness Report for insights from the closed-door dialogue session on cross-border AI incident infrastructure.
