If you're building agentic systems — anything that reasons through multi-step tasks, executes tool calls, or acts in the world with some autonomy — the question you can't fully answer yet is this: how do you know when it's going wrong? Logs tell you what happened. Evals tell you what it did on your test suite. Neither one reliably tells you what a capable, goal-directed system is actually optimizing for at runtime.

LawZero's answer is: build a different kind of AI and use it to watch.

On September 16, Canada and Germany each committed up to CAD $150M to LawZero, Yoshua Bengio's Montreal nonprofit, making it the largest dedicated "Scientist AI" investment so far. The announcement came at the ALL IN Montreal conference. Bengio is a Turing Award winner and runs the Mila AI institute at Université de Montréal — the most credentialed person in AI safety research getting a nine-figure government mandate is not a routine event.

The concept at the center of this is "Scientist AI." It's a specific technical proposal, not a marketing term, and it's worth understanding what it does and doesn't mean before deciding how much to pay attention.

What Scientist AI actually means technically

LawZero's founding thesis, published when Bengio launched the nonprofit in 2025, starts from a specific objection to current AI development patterns. Most capable AI systems are trained with reinforcement learning or RLHF. They learn, in some functional sense, to pursue objectives. The safety challenge for these systems is that the objectives they end up pursuing are not always the ones you wanted, and this diverges at scale.

Scientist AI rejects this paradigm. The design constraints are:

• No reinforcement learning. Only supervised learning on human-generated data. No reward model, no objective function shaping behavior.
• No goals. The system is not trained to optimize for outcomes. It is trained to represent knowledge and generate transparent inferences from that knowledge.
• Externalized reasoning. The reasoning trace is not a post-hoc explanation generated to satisfy a user. It is the actual computation. What you see is what happened.
• Non-agentic by design. The system does not initiate actions. It responds to questions. It does not have a drive to accomplish anything.

The idea is that a system with this architecture cannot have misaligned goals — because it has no goals. It can be wrong, but it can't be strategically wrong in a way that benefits a hidden objective.

CAD $300M

→ Source: Government of Canada

The monitoring application

The most concrete near-term use case for Scientist AI is as a runtime monitor for agentic systems. Here's the mechanism: you run an agentic AI (call it the worker) that is doing something complicated. In parallel, you run a Scientist AI (the monitor) that observes what the worker produces, compares it against what the worker said it was doing, and flags divergence.

This is genuinely useful if the architecture holds. A goal-free monitor can't be socially engineered by the system it's watching. It doesn't have a stake in covering for the worker. It has no reason to tell you things are fine when they're not, because "things being fine" isn't something it cares about.

Compare this to a monitor that's also trained with RL. That monitor has a reward signal. If the reward signal correlates with "user is satisfied" or "fewer escalations," the monitor can learn that flagging fewer issues is better. Goal-free architecture eliminates that failure mode by design, not by tuning.

Whether the architecture can actually deliver on this in practice is what the CAD $300M is meant to find out.

1. June 2025

2. Sept 16, 2026

Source spread

• Government of Canada — Canada and Germany Invest in LawZero [builder] — Primary source for the funding announcement. $150M each, conditional commitments.
• Globe and Mail — Yoshua Bengio's LawZero receives $300M from Canada, Germany [builder] — Context on the ALL IN announcement and Bengio's positioning.
• TechTimes — Goal-free AI gets its first government mandate: Canada, Germany back LawZero [safety] — Good technical summary of what Scientist AI means and what the monitoring use case looks like.
• Université de Montréal — Safe-by-design AI: Yoshua Bengio launches LawZero [builder] — The 2025 founding announcement with the original Scientist AI technical framing.
• LawZero.org [builder] — The nonprofit's own description of its research agenda.

What's real

• The monitoring application is the most technically coherent near-term use case. An AI with no goals watching an AI with goals and flagging divergence between stated and actual behavior is a legible safety mechanism, not a vague aspiration.
• Government money at this scale changes what's possible. LawZero was doing safety research on ~$30M philanthropic. CAD $300M funds a serious compute budget, which matters because you can't test AI monitoring without building capable AI to monitor.
• The no-RL, no-goals constraint is falsifiable. Either you can train a system that's useful without optimization pressure, or you can't. That's a real empirical question the money can answer.
• Bengio's institutional credibility matters here. This is not a startup with a pitch deck. The Turing Award, Mila, and his track record mean the technical claims are being made by someone who knows what the words mean.

What deserves a side-eye

• "Up to CAD $150M each" means both commitments are conditional. The full amount requires LawZero to hit milestones. Neither government has handed over the whole sum. The headline number is a ceiling, not current funds in the bank.
• The architecture claims are largely untested at the scale this money will enable. Supervised-only, no-RL works fine for many narrow applications. Whether it can produce something useful as a runtime monitor for a capable agentic system is the open question, not a settled one.
• "No goals" is harder to verify than it sounds. A model trained entirely with supervised learning still has implicit objectives baked in by the choice of training data, the loss function, and what counts as "correct." Calling it goal-free is a design intent, not a guaranteed property of the resulting system.
• There's a sovereign AI dimension to this announcement that didn't get much coverage. Both governments described the investment partly in terms of not depending on US or Chinese AI infrastructure for safety-critical oversight. That's a policy goal attached to the research goal. Worth tracking whether that framing shapes what gets built.

What builders need to know

• If you're building agentic systems today, the monitoring gap is real. LawZero's framing — that you need a goal-free system to watch goal-directed systems — is one solution. It's not the only one, but it's the most technically principled proposal for addressing the problem Bengio is pointing at. Keep it on your radar.
• The conditional funding structure matters for timeline. CAD $300M over probably 5-7 years, conditional on milestones, means useful outputs from this research are not near-term. Plan your own safety architecture around what exists today.
• Watch for the first concrete Scientist AI model release. LawZero will presumably ship something demonstrable to unlock further tranches. When that happens, the claims become testable and the builder community can evaluate them directly.
• The no-RL constraint has implications beyond safety. A system that is strictly supervised-learning has predictable capability ceilings. Scientist AI will be useful as a monitor to the extent it can represent and reason about the behavior of systems far more capable than itself. Whether that gap is bridgeable is a genuine open question.

Further reading

• Government of Canada — Canada and Germany invest in LawZero — the primary source
• Globe and Mail — Yoshua Bengio's LawZero receives $300M — announcement context
• TechTimes — Goal-free AI gets its first government mandate — technical summary
• Université de Montréal — Yoshua Bengio launches LawZero (2025) — founding thesis
• LawZero.org — the nonprofit's research agenda