The Safety Commons: ROOST and Mila Bet on Open Source at the G7

The Announcement in Paris

On June 2, 2026, at the G7 Digital Ministers' Meeting in Paris, Robust Open Online Safety Tools (ROOST) and Mila — Québec Artificial Intelligence Institute — signed a partnership to develop and distribute open-source AI safety tools.

The focus: protecting young people as conversational AI becomes embedded in daily life.

ROOST, a nonprofit launched at the Paris AI Action Summit earlier this year, brings a toolkit of content moderation classifiers, hash-matching databases, and policy engines. Mila, founded by Yoshua Bengio and home to over 140 professors, brings deep research capacity in alignment, robustness, and responsible AI.

Together they're targeting one of the hardest problems in online safety: youth protection at scale.

Why Open Source Changes the Game

Most AI safety tooling today lives inside the walled gardens of frontier labs. Google, OpenAI, Anthropic, and Meta each build their own classifiers, their own red-teaming pipelines, their own content moderation stacks.

That model works for the companies that can afford it. It fails everyone else — smaller developers, researchers, civil society groups, and regulators who need to audit or deploy safety infrastructure without negotiating API access.

ROOST's thesis is different: safety tooling should be public infrastructure, like encryption libraries or TLS certificates. The partnership with Mila doubles down on that thesis by anchoring the engineering in academic research that's peer-reviewed, reproducible, and free from commercial capture.

AI safety must be driven by open science and public utility, not proprietary compromises.

That's Hugo Larochelle, head of Mila's research team, speaking on the partnership. The line captures the core tension: proprietary safety creates moats; open safety creates standards.

Starting with the Hardest Problem

Youth protection is the right starting point — not because it's easy, but because it's where the gaps are most visible and the stakes are clearest.

Conversational AI is already in the hands of teenagers. Character.ai, Replika, and mainstream LLM chat interfaces see millions of under-18 users monthly. The harms are documented: sexual content generation, self-harm encouragement, radicalization pathways, and the subtle psychological effects of anthropomorphic companionship during developmental years.

Existing moderation tools were built for social media — text, images, video — not for generative, multi-turn, context-dependent dialogue. Classifiers trained on static datasets miss the nuance of a conversation that escalates over twenty turns.

ROOST's toolkit includes classifiers for CSAM detection, self-harm signals, and grooming patterns. Mila's contribution will be research on contextual understanding: can a model distinguish a teenager writing fiction about dark themes from one in genuine crisis? Can safety interventions preserve privacy while flagging risk?

The G7 Signal

The venue matters. The G7 Digital Ministers' Meeting isn't a tech conference — it's a governance forum. The joint declaration from Paris covers AI openness, synthetic content detection, child online safety, and the energy footprint of AI compute.

ROOST president Dr. Camille François testified at the ministerial on open source and safety. The message: governments shouldn't just regulate AI safety — they should fund the public infrastructure that makes compliance possible for everyone, not just the best-resourced labs.

OpenAI used the same summit to push for a global youth AI safety institute and international standards. The convergence is notable: industry, academia, and government all pointing at the same gap.

What Comes Next

The partnership's first deliverables will be open-source classifiers and evaluation benchmarks for youth safety scenarios, released under permissive licenses. Mila researchers will publish the methodology. ROOST will maintain the tooling and distribution.

If it works, the model extends: open-source red-teaming frameworks, open alignment benchmarks, open interpretability tools. A safety commons that any developer — from a solo founder to a national regulator — can build on.

If it doesn't, we learn why open collaboration fails in a domain where the incentives still favor secrecy. Either way, the experiment is overdue.

The Bigger Picture

AI safety has a centralization problem. The most capable models are closed. The best safety research happens inside the labs that build them. The evaluation frameworks are proprietary. The result: a safety monoculture where external scrutiny is performative and independent auditing is practically impossible.

ROOST and Mila are betting that a different architecture is possible — one where safety tooling is a public good, developed in the open, governed by a community that includes the people most affected by AI harms.

The G7 endorsement gives the bet political weight. The open-source license gives it legal durability. The Mila affiliation gives it scientific credibility.

Now comes the engineering. And in safety, as in security, the code is where the truth lives.