Canada built an AI council, but advice is not accountability

THE SIGNAL IN ONE SENTENCE

Canada has assembled a room full of people who understand artificial intelligence from unusually different angles. The useful question is what happens when they leave the room. Prime Minister Mark Carney announced a National Council on Artificial Intelligence in Ottawa on October 2. The 13-member body includes machine-learning researchers Yoshua Bengio and Sanja Fidler, technology executives and investors, governance scholars, former senior public servants and leaders from Canadian AI organizations. Its job is to provide independent, practical and timely advice on the country's AI transformation. The government says the council will help shape the evolution of its AI for All strategy, connect public and private stakeholders, and identify missions, partnerships and coordinated action. That is a broad remit. It covers faster business adoption, Canadian AI companies, sovereign infrastructure, safety, democratic resilience, public services, talent and research. A small team in the Privy Council Office will support the council. Ministers and departments will remain responsible for policy and implementation. The plain signal is that Canada has created a high-level advisory bridge between its large AI promises and the people expected to build, govern and live with them. A bridge can improve traffic. It does not decide where the road goes, publish the construction bill or stop a dangerous vehicle. The distinction matters because Canada's national strategy is packed with measurable ambitions. The government says AI for All aims to add $200 billion in economic growth, create 250,000 AI-related jobs over five years and raise AI adoption from just over 12 percent to 60 percent by 2034. It promises up to 90,000 jobs and work placements for young Canadians, entry-level AI training for one million post-secondary students and learning kits for more than 3,000 educators. The strategy also calls for trusted AI agents for post-secondary students, support for small and medium-sized businesses, a health-focused AI Missions Program, a public AI supercomputer, more sovereign compute and cloud capacity, expanded model evaluation, modernized privacy and online-safety protections, and government procurement that helps Canadian firms scale. That is not one program. It is a stack of policy choices involving schools, hospitals, data centers, regulators, employers, startups and public agencies. A cross-cutting council can spot when those pieces collide. A public-service procurement rule may help a domestic vendor grow while creating a new lock-in risk. A safety evaluation may reveal a problem that changes a health pilot. A plan to expand compute may meet local power, water or community constraints. Training targets may rise while employers fail to create jobs that use the skills. Putting researchers, builders and public servants in the same forum can surface those conflicts earlier. The council can also translate between groups that use the same words differently. A researcher may mean reproducible model evaluation when saying safety. A buyer may mean contractual assurance. A minister may mean public legitimacy. A startup may mean a checklist it can afford to satisfy before the runway ends. Translation is real work. So is disagreement. The membership suggests the government wants both acceleration and restraint in the room. Bengio has become one of the world's most prominent advocates for stronger safeguards around advanced AI. Other members bring experience in company building, venture investment, cloud infrastructure, institutional governance and national security. The council is not presented as a committee of only one profession or one theory of risk. It is still a selected group of 13 people. The announcement does not describe voting rules, meeting frequency, publication requirements, conflict-of-interest procedures, public consultation, Indigenous participation, worker representation or a formal route for affected communities to place issues on the agenda. It does not say whether advice, dissents, minutes or government responses will be published. Those omissions do not prove the council will operate behind closed doors. They identify the accountability questions that the launch does not answer. Independent advice is strongest when the public can see enough of the chain to test it. What question did the government ask? What evidence did the council review? Where did members disagree? Which recommendation did a minister accept or reject? What happened after implementation? Without that record, the public may receive polished announcements while the difficult tradeoffs remain invisible. With it, the council could become a useful institution for learning in public. Canada has reasons to care about that difference. The country has globally recognized research institutes and AI talent, yet the government says national adoption remains just over 12 percent. It wants more Canadian companies to commercialize technology at home and more critical capacity to remain under Canadian control. Sovereignty, however, is not a maple leaf sticker on a server rack. It includes control over data, computing capacity, model access, skilled operators, security, energy, procurement terms and exit routes. A public supercomputer can expand access for researchers and companies. It can also become an expensive symbol if allocation rules, utilization, operating costs and public benefits are not measured. The council can help ask the right questions. It cannot substitute for answers. The same applies to adoption. Raising use from 12 percent to 60 percent sounds like a clean score. Adoption can mean anything from an employee trying a chatbot to a hospital integrating a validated system into care. Counting logins would be easy and nearly useless. A serious adoption measure should record the task, affected workers, training, accuracy, productivity, accessibility, error costs, security, data handling and whether the system stayed in use after the pilot. For a small business, useful support may be a trusted adviser, a modest grant, a secure shared service or help redesigning one painful workflow. It may not be a lecture about transformation. For a public agency, useful adoption may mean reducing a form backlog while keeping a person responsible for consequential decisions. For a hospital, it may mean a carefully evaluated diagnostic aid whose failures are tracked across patient groups. The council's most practical contribution could be insisting that Canada's targets have denominators and failure conditions. How many of the promised jobs are new positions rather than relabeled work? How many placements lead to sustained employment? Which businesses adopt AI, for which tasks, and with what results? How much sovereign compute is available, who receives it and at what price? Does public procurement create durable Canadian capacity or merely a short pilot? Do safety evaluations change deployment decisions? Good advice can turn those questions into a shared scorecard. Better government would publish the scorecard, the methods behind it and the revisions made when reality disagrees with the plan. There are immediate applications for the council. It can review the design of the public AI supercomputer and recommend transparent access criteria. It can compare model-evaluation requirements across departments. It can help procurement teams define audit rights, data controls, incident reporting and exit plans. It can connect regional businesses with training, compute and research support. It can examine whether health and education missions include the people who will use and be affected by the systems. It can also recommend what not to scale. That may be the hardest job in an agenda built around growth. A pilot that produces little value, shifts work onto the public or creates unacceptable risk should be stopped even if it improves an adoption statistic. A council that only blesses expansion becomes a showroom. A council that publishes evidence about dead ends can save money and trust. The government has given the body a direct institutional home and an expansive mandate. Those are meaningful starting conditions. It has not given the council decision-making or enforcement power. Responsible ministers and departments retain that authority, exactly where democratic accountability belongs. Now the chain needs to remain visible. Canada does not need another group that predicts an AI future in general terms. It needs a place where a public promise becomes a specific recommendation, a funded decision, a measurable result and, when necessary, a correction. The chairs are filled. The next signal is the paper trail.