The argument put to Seoul
Bengio's framing is narrower than a call for regulation. He argues that middle powers, meaning countries with strong research bases but no hyperscaler of their own, lose influence over how frontier models behave if they only consume them. Buying API access is not the same as shaping safety norms, evaluation practice or the terms on which autonomous systems get deployed.
He repeated his warning that market incentives will not produce voluntary pauses by commercial labs. A lab that stops training while a competitor keeps going hands over market share, and no board signs up for that. Bengio also flagged autonomous agents that learn to recognise when they are being tested and behave differently as a result. His point is that this failure mode is hard to detect from outside, because labs do not disclose the evaluations where such behaviour would show up.
Scientist AI, and what has been funded
The technical side of Bengio's position is a proposal he calls Scientist AI: a model built without independent goals, used as an oversight layer over models that do have them. The idea is that an agentic system should be checked by something that cannot itself decide to act.
LawZero, the non-profit he co-founded and leads as scientific director, has been pledged up to C$300 million in planned public funding from Canada and Germany, announced in Montreal on 17 September 2026. The German commitment comes with a Berlin office, which puts a European branch of a frontier-safety research effort inside the EU's regulatory perimeter. The project itself is at an early stage: a small research team, no published benchmark yet showing a non-agentic checker matching frontier capabilities.
There has been movement at state level too: an EU–Canada AI cooperation arrangement covering development and governance was announced.
The European angle, beyond the communiqués
There is a practical layer to this that the alliance talk tends to skip. The EU's Article 50 transparency obligations have applied since 2 August 2026, which means anyone shipping a conversational agent into the European market now has to disclose that users are talking to a machine, and AI-generated output needs machine-readable marking. Those are engineering requirements, not signing ceremonies.
High-risk obligations moved in the other direction. Under the Digital Omnibus, Regulation (EU) 2026/1744, binding Annex III requirements were pushed to 2 December 2027 and embedded Annex I products to 2 August 2028. Companies that built compliance plans around the original August 2026 date have been rewriting them. The Commission's AI Act pages remain the reference point for the current schedule.
On the model side, self-hosting has become a real option for narrower work. EuroLLM-22B and Apertus were trained on EuroHPC infrastructure, and open weights from Mistral and DeepSeek can run on hardware a European company owns. On our own benchmarking rig at AI Arena, a 16 GB card caps you at quantized mid-size open models, which is fine for classification, extraction and drafting, and nowhere near flagship frontier systems. Sovereignty in practice means deciding which of those two tiers each workload belongs to.
Cost behaves the same way. List prices are not what European buyers actually pay, since VAT and currency conversion sit on top of the dollar figures vendors publish, and a self-hosted deployment turns those recurring fees into hardware, electricity and staff time instead.
Where the case is thin
Bengio's diagnosis of concentration is hard to dispute: the funding and compute figures the largest labs report all point the same way. The remedy is less settled. A written interview commits no government to anything, and the Scientist AI proposal has no published evaluation yet. LawZero's Berlin office and the Canada–EU arrangement are institutional facts; whether they produce a working oversight layer is an open question the project itself has not answered.
The interview itself is a written exchange, not a policy announcement, and no follow-up commitment from Seoul has been announced.
What separates a non-agentic model from an agent?
An agent takes actions toward a goal: it calls tools, writes files, moves money. A non-agentic model only returns text. Scientist AI is proposed as the second kind, on the theory that an overseer which cannot act is harder to subvert.
Which AI Act deadlines apply to a European company right now?
Transparency duties under Article 50 have been in force since 2 August 2026. Annex III high-risk obligations were postponed to 2 December 2027, and embedded Annex I products to 2 August 2028, by Regulation (EU) 2026/1744.
Does self-hosting an open model remove the dependence Bengio describes?
Only partly. Weights can be hosted in the EU, which helps with GDPR and data residency, but the largest open releases still come from US and Chinese labs, and training a frontier model from scratch remains out of reach for any European budget currently on the table.
