May 2026

This monthly newsletter covers recent developments and upcoming events in AI safety, ethics, and governance in Montréal.

Events

AI Safety Papers We Love #1: Multi-Agent Risks from Advanced AI
Wednesday, May 6, 6:30–8:30 PM. Ω Labs.
Reading group convened by Orpheus on “Multi-Agent Risks from Advanced AI” by Hammond et al. Discussion of failure modes and risk factors in advanced multi-agent AI systems.

IVADO Workshop: Statistics in Trustworthy AI
May 11–15, 2026, 9:30 AM–5 PM.
Five-day workshop on statistical methods underpinning trustworthy AI, part of IVADO’s thematic semester on Statistical Foundations of AI. Registration open on a first-come basis with limited capacity.

Understanding and Addressing Fairwashing in Machine Learning
Tuesday, May 19, 7–9 PM. UQAM Pavillon Président-Kennedy, PK-1140, Montréal.
Sébastien Gambs (UQAM, Canada Research Chair in Privacy-preserving and Ethical Analysis of Big Data) on fairwashing: how unfair black-box models can be made to look fair through post-hoc explanation manipulation, why fairwashing attacks transfer across models, and what limits detection.

Mila’s Community of Practice: AI Governance
Thursday, May 21, 9 AM–12 PM. Agora de Mila, 6666 Rue Saint-Urbain #200, Montréal.
Third session of Mila’s AI Governance Community of Practice, an ongoing series convening leaders, experts, and practitioners to share real-world challenges across sectors. This session is organized around four axes: evaluation protocols subtle enough that autonomous AIs cannot detect the test and bias their behavior, secure testing environments for agentic AI, support for employees in supervisory roles over autonomous agents, and policy adaptation for legal responsibility as AI agents grow more autonomous and interactive.

Obvia: Les risques éthiques de l’intelligence artificielle pour la démocratie
Thursday, June 4, 2026, 1–6:30 PM. HEC Montréal, Édifice Côte-Sainte-Catherine, Amphithéâtre Banque Nationale.
Public launch of an Obvia ethics-commission advisory naming nine ethical risks generative AI poses to democratic processes, with twenty-one recommendations. Panels with experts on electoral integrity and the quality of public discourse.

Research from Montréal

Obvia’s État de la situation 2026
Obvia, April 16, 2026. Third annual stocktake on AI’s societal impacts in Québec, with the 2026 edition focused on agentic AI and warning of a weakening governance capacity, deepening power concentration, and widening inequalities.

SHAPO: Sharpness-Aware Policy Optimization for Safe Exploration
Mani, Pequignot, Mai, Paull. ICLR 2026. A sharpness-aware policy-update rule for safe RL: evaluating gradients at perturbed parameters makes updates pessimistic about the actor’s epistemic uncertainty, expanding the safety–performance Pareto frontier on continuous-control tasks.

Generative Adversarial Post-Training Mitigates Reward Hacking in Live Human-AI Music Interaction
Wu, Brade, Ma, Fowler, Yang, Banar, Courville, Jaques, Huang. ICLR 2026. Live human-AI music jamming under coherence rewards collapses to reward hacking; a co-evolving discriminator pulls the policy back toward the data distribution, restoring output diversity in real-time interaction with expert musicians.

Benefits and Limitations of Communication in Multi-Agent Reasoning
Rizvi-Martel, Bhattamishra, Rathi, Rabusseau, Hahn. ICLR 2026. A theoretical framework for the expressivity of multi-agent reasoning systems, with bounds on agent count, inter-agent communication, and achievable speedups for state tracking, recall, and k-hop reasoning.

A Benchmark for Evaluating Outcome-Driven Constraint Violations in Autonomous AI Agents
Li, Fung, Weiss, Xiong, Al-Hussaeni, Fachkha. NeurIPS 2025. 40 scenarios test whether autonomous LLM agents violate ethical, legal, or safety constraints under KPI pressure. Across 12 frontier LLMs, violation rates run 11.5% to 66.7% (most above 30%) and do not reliably improve across model generations.

Have an event, opportunity, or article for next month? Send us a note at team@horizonomega.org, or just reply to this email.

This newsletter is by HΩ, researched and written with the assistance of AI.