June 2026

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

This edition is going out a little later than usual, so some of the events listed below have already taken place; they’re included for the record.

Events

METR Frontier Risk Report
Tuesday, June 2, 7–9 PM. Ω Labs, 3813 rue Saint-Denis, Montréal.
Orpheus Lummis (Horizon Omega) presents METR’s first Frontier Risk Report, which finds that, as of its February–March 2026 assessment, internal AI agents at Anthropic, Google, Meta, and OpenAI plausibly had the means, motive, and opportunity to begin minimal “rogue deployments” that run without human knowledge.

Obvia: Les risques éthiques de l’intelligence artificielle pour la démocratie
Thursday, June 4, 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.

Colloque Femmes en TI et en IA : Leadership et inclusion dans les technologies numériques et l’IA
Friday, June 5, 9 AM–4 PM. Université TÉLUQ (room 11.051), Montréal. Hybrid, in French.
An OBVIA colloquium on women’s persistent underrepresentation in IT and AI, and how homogeneous teams perpetuate algorithmic biases that reproduce pre-existing inequalities.

IVADO Workshop: Uncertainty in AI
June 8–11, 9:30 AM–5 PM. Montréal.
Four-day workshop on uncertainty quantification in AI, part of IVADO’s thematic semester on Statistical Foundations of AI, covering Bayesian deep learning and hallucinations in generative AI.

Canada’s “AI for All” Strategy: Open Discussion
Tuesday, June 9, 7 PM. Ω Labs, 3813 rue Saint-Denis, Montréal.
A structured open discussion, hosted by Horizon Omega, of Canada’s newly released national AI strategy and how well it addresses safety, oversight, and the public interest.

Disinformation 2.0: When AI Blurs the Lines
Wednesday, June 10, 6–8 PM. Pub La Cale, 6839 rue St-Hubert, Montréal.
An “After Mila” meetup with Jean-François Godbout (UdeM, Mila) and Pascal Lapointe (Agence Science-Presse) on AI-generated disinformation and deepfakes.

AI Control Hackathon
June 11–18. Kickoff Thursday, June 11, 7 PM. Ω Labs, 3813 rue Saint-Denis, Montréal.
A one-week AI safety hackathon co-hosted by Horizon Omega and HackOS on AI control: limiting the harm an untrusted AI system can cause even when it tries to subvert its controls. Kickoff talk by Henri Lemoine (EquiStamp, Mila, McGill); demos June 18.

AI Governance in 2026: What’s going on, why it’s a mess, and why it’s going to get messier
Wednesday, June 17, 7–9 PM. Ω Labs, 3813 rue Saint-Denis, Montréal (and online).
Stephen Casper (incoming Assistant Professor of Public Policy at the Harvard Kennedy School; contributing author, International AI Safety Report) on why AI governance has so far produced more case studies in failure than success, and how it may change.

ACM FAccT 2026 (Fairness, Accountability, and Transparency)
June 25–28. Le Centre Sheraton Montréal, 1201 boul. René-Lévesque Ouest, Montréal.
The flagship cross-disciplinary conference on fairness, accountability, and transparency in socio-technical systems comes to Montréal. Registration required.

Policy and Governance

Canada: Feds launch national AI strategy, “AI for All”. On June 4, Prime Minister Mark Carney and AI Minister Evan Solomon unveiled the long-delayed national strategy: three principles (trust, opportunity, sovereignty) and six pillars. It commits $50M to expand the Canadian AI Safety Institute (CAISI) and a $500M Tech Growth Fund, funds a sovereign-compute buildout (850 MW by 2030), and targets business AI adoption of 60% by 2034.

Canada: Ottawa “very seriously” weighing age limits for AI chatbots. The federal government says it is “very seriously” considering age restrictions (a floor of 16) on social media and AI chatbots such as ChatGPT, likely via the promised online harms bill. Liberal Party members passed a resolution backing a ban on under-16s, and a May poll found roughly 69% of Canadians support restricting minors’ chatbot access.

Research from Montréal

Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing
Hashemzadeh, Huang, Kim, Côté, Chandar. Mila, Université de Montréal, Polytechnique Montréal, Microsoft Research. Erasure-based safety alignment leaves models over-cautious; their SafeMoE framework instead isolates harmful knowledge into domain-specific LoRA experts and orchestrates them at inference via a learned router, reporting a >20% relative (>15% absolute) gain in safe-response rate while keeping answers informative.

Value Drifts: Tracing Value Alignment During LLM Post-Training
Bhatia, Nayak, Kamath, Mosbach, Stańczak, Shwartz, Reddy. Mila, McGill. Tracing when values get installed during post-training, they find supervised fine-tuning does almost all the work while preference optimization (PPO, DPO, SimPO) barely shifts values, not for lack of capable algorithms but because real preference data carries little value-gap signal.

Visual Symbolic Mechanisms: Emergent Symbol Processing in Vision Language Models
Assouel, Campbell, Bengio, Webb. Mila, Université de Montréal (Campbell at Princeton). ICLR 2026 (Oral). A mechanistic-interpretability study showing that vision-language models solve feature binding via emergent symbolic indices tied to spatial position, and that their known binding failures trace directly to breakdowns in these mechanisms.

ROOST and Mila partner on open-source AI safety, starting with youth safety
Announced at the G7 Digital Ministers’ Meeting, Paris, June 2. Mila and ROOST (Robust Open Online Safety Tools) will openly release safeguards that detect harmful interactions with conversational agents (such as content encouraging self-harm) and redirect young users toward support.

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

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