17 SEP 2026 — Two governments have put up to C$300 million behind an organisation whose product is not a frontier model but a check on one. LawZero, the non-profit Yoshua Bengio founded in June 2025, announced the money at the ALL IN conference in Montréal on Wednesday.
Canada is providing up to C$150 million and Germany up to the same again. The organisation employs close to fifty people.
What it is building
The system, Scientist AI, is designed around what it lacks. Where frontier models are trained to act and pursue objectives, Scientist AI is meant to reason transparently and produce evidence-based output without holding goals of its own.
The organisation was incubated at Mila, the Québec AI institute, and is run day to day by Sam Ramadori, its co-president and executive director, who previously ran BrainBox AI. Bengio's framing of the problem has been consistent since he founded it: "We urgently need to have a plan to build AI that is not going to do really bad, misaligned things."
The intended job is supervision. LawZero describes it as monitoring agentic systems and building guardrails around them, aimed at the behaviours the field has spent two years documenting: deception, self-preservation and goal misalignment. BetaKit's account records the shorthand used at the conference, an AI babysitter for other AIs.
Bengio's argument is about sequence, not sentiment: "If one day [AI] gets to be smarter than us, it's essential we have the technology so AI serves us, not the other way around."
Where the money goes
According to the organisation, the money funds three things: more staff and compute for research, which is the ordinary cost of this work; a Berlin office, which turns the German half into a partnership rather than a donation; and sovereign compute inside Canada, through arrangements with Hypertec and 5C data centres.
The sovereign compute is the key detail. A safety organisation that rents its compute from the companies whose models it is examining has an awkward dependency, and the countries funding it plainly want the capacity onshore.
Public money, and what it buys
Evan Solomon, Canada's AI and digital innovation minister, gave the industrial argument plainly: "This investment from Canada and Germany…will create hundreds and hundreds of full-time jobs. It will support training, it will support our ability to build reliable AI here." He also framed the partnership as a matter of choice rather than charity: "No country can build alone...We need options."
Separate the two halves. Jobs and training are a domestic case for spending, and they would apply to any well-funded lab. The options argument is strategic. Two mid-sized economies are funding oversight rather than another foundation model because they cannot outspend the frontier labs on capability, and they can fund the one thing those labs are not incentivised to build.
The obvious objection
An oversight system is only as good as its independence. Taking C$300 million from two governments buys a relationship with them. Public funding avoids the conflict of being paid by the companies under scrutiny, but it introduces another: states are also buyers, regulators and, increasingly, operators of these systems.
There is a second problem the money does not solve. A supervisor model must be capable enough to understand the system it is supervising. That means LawZero needs frontier-adjacent capability on a budget that is a rounding error next to what the major labs spend. Bengio has been blunt about where that leads: "Eventually, they'll be able to pass through any kind of software defence."
Neither objection makes the work pointless. They set the standard by which it should be judged, which is whether anything it produces gets used by people who did not commission it.
What to watch
Start with publication. A guardrail system that is described but not released is a research programme. A system with published methods, weights or evaluations that others can run is infrastructure. Nothing in this announcement says which it will be.
Then adoption. The test is whether a frontier lab, a regulator or a large deployer puts Scientist AI in front of an agent they actually run. The labs have shown no sign of accepting external monitoring they do not control.
The German leg matters most. A Berlin office funded by Germany, inside an EU regime that already regulates this ground, is where an oversight tool would first meet a legal requirement to use one.
View author profile → · Editorial policy
