Canada and Germany fund Yoshua Bengio's non-profit alternative: a public bet on safe AI
On 16 September 2026, at the ALL IN conference in Montréal, Canada's Minister of Artificial Intelligence and Digital Innovation Evan Solomon and Germany's Federal Minister for Digital Transformation Karsten Wildberger announced joint public backing for LawZero. The non-profit, launched in June 2025 as an organisation incubated at the Mila institute and founded by Yoshua Bengio, has around 50 employees and is taking the opposite path to the commercial giants: building 'Scientist AI', a system with no goals of its own, designed to give evidence-based answers. According to the organisation, it is meant to act as a guardrail for today's frontier models. This is the first direct move by two G7 countries in support of a non-profit alternative to frontier models.
Look at the numbers, though, and neither amount has been paid out yet. Canada plans to invest CAD 150 million through the Strategic Response Fund; the project includes sovereign compute infrastructure in Canada with the companies Hypertec and 5C. Germany plans EUR 100 million for the research and development work of LawZero's new Berlin office. LawZero speaks of a total commitment of 'up to CAD 300 million': a figure that only adds up at today's exchange rate, because Germany's official share is in euros (100 million). The German share remains conditional on notification to, and review by, the European Commission under state aid rules. Nor has either side said over how many years the money will be disbursed or what the plan's interim milestones are.
The deal aims to strengthen both countries' technological sovereignty. It fits into Canada's 'AI for All' strategy and answers the need Solomon voiced, as reported by the Globe and Mail, that 'no country can build alone'. Wildberger described the model as an alternative to 'potentially uncontrollable' autonomous agents. Yet the real weight of the announcement runs up against a lack of technical data: Scientist AI has not published any metrics or performance data that third parties can verify, and that would show how this transparent model could actually guide or constrain today's frontier systems.
Funding a non-profit alternative to commercial frontier models is a meaningful political signal, but good intentions are no substitute for technical verification. With no disbursement timeline and no empirical evidence that Scientist AI works, the risk is mistaking a desirable theoretical architecture for a result already achieved. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the Canadian government's official release (canada.ca, ISED): amounts, currencies, Strategic Response Fund, EU notification, the German office, partners Hypertec and 5C, and Minister Joly's statement. I read LawZero's release on PR Newswire: total of up to CAD 300 million, Berlin office, Wildberger's statement. Date, ALL IN conference, headcount and earlier funding are confirmed by The Globe and Mail, The Logic and BetaKit. Some outlets report '150 million each', but Germany's official figure is EUR 100 million: that is the one used. Quotes were extracted with a tool that summarises web pages, so the exact wording should be checked against the sources.
- Incertezze
- No source says over how many years the funds will arrive or what the milestones are. Canada 'plans' to invest and Germany's share awaits European Commission review under state aid rules: no money has been paid out yet. The 'up to CAD 300 million' total is LawZero's figure and only adds up at today's exchange rate. Scientist AI has no verifiable public results yet. The Globe and Mail mentions talks with other countries about further funding, without naming them.
- Perché pubblicarla
- Two governments, one of them in the EU, are putting public money into an approach to AI safety that offers an alternative to commercial autonomous agents. The news dates from 16 September 2026, touches on European digital sovereignty, EU state aid and non-profit research, and does not overlap with articles already published.