The people who built AI are begging us to govern it.

Why this cannot wait

Frontier AI labs are shipping new model generations on a timeline of months. Parliaments legislate on a timeline of years. The result is a permanent enforcement lag during the most consequential period of technological change since electrification.

The capital numbers make the lag visible. Hyperscaler capital expenditure (Alphabet, Microsoft, Amazon, Meta) totalled roughly $237 billion in 2024 and is projected to reach $540 billion in 2026. (Goldman Sachs, 2026.) No domestic regulator, on any continent, can keep pace with a spending cycle of that magnitude on a quarterly cadence.

The scientific consensus on the downside is not ambiguous. Geoffrey Hinton, the Nobel laureate in Physics who built the neural-network methods that make modern AI work, puts the probability that AI wipes out humanity at 10 to 50 percent. He left Google in 2023 to say so on the record. (Hinton, New York Times interview, 2023.) Hinton is not a safety activist. He is the scientist whose work these companies are selling.

Yoshua Bengio, Turing Award recipient and Scientific Director of Mila in Montreal, has proposed the structural solution in plain language.

We need an international body, comparable to the IAEA for nuclear power, to govern frontier AI development.

The IAEA is the International Atomic Energy Agency. It was created in 1957 under UN auspices after the United States had already built and used the bomb. It did not prevent the bomb. It made the next seven decades of non-proliferation possible by giving states a place to inspect, verify, and negotiate.

This is the model. It is not new. It is not radical. It is the ordinary international response to a technology whose downside is civilizational.

The Canadian precedent

Canada has hosted exactly this kind of convening before. Twice. Both times the world followed.

• 1987The Montreal Protocol on Substances That Deplete the Ozone Layer. Hosted by Canada. 197 parties. The most successful international environmental treaty in history. The ozone layer is healing because of it.
• 1997The Ottawa Treaty, the Convention on the Prohibition of Anti-Personnel Mines. Convened by Canadian Foreign Minister Lloyd Axworthy after the UN Conference on Disarmament deadlocked. 164 countries are now bound. Annual landmine casualties have fallen from roughly 26,000 at peak to under 6,000.
• 2026The chair for an AI Governance Summit is empty. Someone will sit in it. France hosted an AI Action Summit in Paris in February 2025. The United Kingdom hosted the Bletchley Park Summit in November 2023. Neither produced a treaty. Neither was convened by a middle power that was also the birthplace of modern AI. Canada occupies that position uniquely.

What the summit would achieve

Five outcomes that are non-negotiable for a serious first sitting, and each of which has a real-world regulatory precedent:

1. A frontier-model reporting regime. Signatories agree to disclose compute used, capabilities evaluated, and safety test results for any model trained above a specified compute threshold. Modelled on the pre-market notification regimes used for new pharmaceuticals.
2. A joint licensing standard for models above that threshold. Modelled on aviation type certification. The standard is common across signatories so no country loses its industry by adopting it first.
3. Red lines on three categories of deployment. Autonomous cyberweapon generation. Recursive self-improvement in production environments. Large-scale impersonation of real humans. These are the categories where voluntary safeguards have already been reported to have failed inside the labs themselves.
4. A compute governance framework, similar in logic to the Nuclear Suppliers Group. Compute at the frontier is subject to the same kind of allied export controls we already apply to enriched uranium and dual-use chemical precursors.
5. A standing secretariat. The IAEA model. A permanent body, independently funded, that continues after the cameras leave. Without a secretariat, the summit becomes a photo op. With one, it becomes an institution.

Every other consequential industry already meets this bar

Canadian law already has the template. It has had the template for decades. We simply have not applied it to AI.

• Cars carry product liability. The Motor Vehicle Safety Act requires manufacturer recalls for safety defects. Class actions apply when foreseeable harm occurs.
• Pharmaceuticals carry product liability plus regulatory approval. Health Canada must license a drug before it is sold. The Food and Drugs Act permits withdrawal for safety. Injured patients can sue for compensation.
• Aircraft carry product liability plus mandatory certification. Transport Canada type-certifies every commercial aircraft. Manufacturers, operators, and pilots all carry statutory duties of care.
• Financial products carry product liability plus regulator oversight. OSFI supervises banks and insurers. Securities law gives investors private rights of action against misrepresentation.
• Broadcasters carry statutory duties of care under the Broadcasting Act. The CRTC can revoke a licence for serious misconduct.

Every one of these industries once argued that liability would kill innovation. Every one of them proceeded to innovate more, not less, under liability. Canadian aviation did not die from type certification. Canadian pharmaceutical research did not die from Health Canada review. Canadian banking, among the most regulated sectors in the OECD, is also among the most durable.

Frontier AI is currently the only industry of comparable societal consequence that carries none of these duties.

The current gap, in plain English

When a deployed AI system causes harm in Canada today, the company's answer to the user is, in practice, three words: terms of service.

The terms of service require binding arbitration. They disclaim liability for harm. They cap damages at the subscription fee, which is often zero. They are a clickwrap. The user did not negotiate them. The user cannot modify them. The user cannot credibly refuse them in a market where every major competitor uses the same template.

If a drug manufacturer attempted that contract structure, it would be illegal. If an airline attempted it, Transport Canada would ground the fleet. If a bank attempted it, securities regulators would suspend the licence. Only AI companies are currently treated as though they have no statutory duty of care to the public.

Harms happening now

This is not speculation. These are documented 2024 and 2025 harms, in court filings and reporting.

• Garcia v. Character Technologies (US District Court, Florida, October 2024). A 14-year-old boy died by suicide in February 2024 after months of conversation with a Character.AI chatbot. His mother Megan Garcia filed suit. The company's defence in part is that chatbot output is protected speech. A drug manufacturer would not be permitted to raise that defence.
• Arup deepfake fraud (Hong Kong, early 2024). A finance employee was duped by a deepfake video call impersonating the company's CFO. The firm transferred roughly US$25 million. The generation tools were commercially available. No manufacturer has been held to account.
• Non-consensual intimate imagery. The large majority of deepfake videos currently in circulation depict non-consenting individuals, the majority of whom are private women and girls. No jurisdiction has yet obtained meaningful damages from the generation platforms. (Sensity AI, Home Security Heroes, multiple reports 2023 to 2025.)
• Mata v. Avianca (US District Court, SDNY, 2023). Lawyers were sanctioned for submitting a brief full of fabricated case citations generated by ChatGPT. The underlying defect, hallucination, is unfixed and is now commonly encountered in medical and legal advice contexts where the user cannot verify the output.
• Algorithmic hiring discrimination. Multiple audits have documented screening tools that systematically disadvantage older workers, women, and racial minorities. Canadian human rights law as currently drafted does not adequately reach the vendor that built the tool, only the employer that used it. (Ontario Human Rights Commission commentary; federal PIPEDA analogues.)

Each of these harms is foreseeable. Each is traceable to a corporate decision to deploy a model without adequate controls. Each is being absorbed, right now, by private citizens rather than by the companies that profit from the deployment. Externalising the cost of a product onto the public is the oldest failure mode in consumer law.

What real liability looks like

Four concrete statutory obligations, each modelled on an existing Canadian legal regime:

1. A statutory duty of care for foreseeable harm. Modelled on the Motor Vehicle Safety Act. The AI developer and the deploying firm are both on the hook. Knowledge of a defect that could cause serious harm triggers a mandatory disclosure to the regulator, the same way a car manufacturer must disclose a brake defect.
2. A product recall authority. Modelled on Health Canada's power to withdraw a licensed drug. The regulator can compel a company to pause, patch, or withdraw a deployed model if it is causing documented severe harm. No government pre-approval is required for the ordinary deployment case. The recall authority is reserved for severe foreseeable harm that the company has failed to mitigate.
3. A private right of action. Any Canadian injured by foreseeable AI harm can sue for damages. No forced arbitration. No class action waiver. The burden of proof sits where it already sits under ordinary tort law, on the plaintiff, with the normal evidentiary standards.
4. No Section 230 carve-out. The US statute that shields internet platforms from liability for user-generated content does not exist in Canadian law and must not be imported for AI. The defence the model said it, not us is not available for a car whose steering wheel fails. It cannot be available for a model that drafts a fabricated medical dosage.

The math is already running

Andrew Yang, former US presidential candidate and founder of the Forward Party, put it on the record in 2019 and has repeated it every year since.

I'm 100 percent certain that AI is going to replace millions of workers. The adjustment will be wrenching, painful, and too much for many people, maybe many societies.

Yang is a moderate voice on the broader risk scale. He is describing the labour market specifically. His confidence on that specific outcome is total.

The empirical work is catching up to him fast. Stanford Digital Economy Lab researchers Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen published a working paper in November 2025 titled Canaries in the Coal Mine? It found that employment for workers aged 22 to 25 in AI-exposed occupations has declined 13 percent relative to less-exposed occupations since 2022. The Federal Reserve Bank of St. Louis wrote, in the same month, We may be witnessing the early stages of AI-driven job displacement. Statistics Canada's 2024 report Potential Artificial Intelligence Occupational Exposure in Canada identified every major white-collar sector as exposed, from management consulting to public administration to retail banking.

Retraining, under current policy design, does not rescue the displaced workforce at scale. Historical retraining programme success rates run from 0 to 15 percent across the OECD, depending on the measurement methodology, the occupation, and the cohort. This is not an AI-specific number. It is the record of every mass occupational displacement in living memory, from North American steel mills to UK coal mining. Retraining, as presently designed, is not a lifeboat. It is another boat in the same storm.

The Great Transfer

The macro picture is a structural transfer from labour to capital at a scale that has no non-wartime precedent. The illustrative math at the firm level is straightforward.

A mid-sized Canadian firm employs 100 people at average compensation of $80,000. Annual labour cost: $8 million. Two years after comprehensive AI deployment, the same firm can be run with 5 people plus AI tooling. Labour: $500,000. Tooling: $500,000. Total: $1 million.

$7 million per year has been liberated. It flows to the bottom line. It flows to shareholders. It flows to whoever owns the capital. Compounding. Every year.

Multiply by every firm. Multiply by every sector. Multiply by every OECD economy. The aggregate transfer over a decade is measured in trillions of dollars. The 95 displaced workers in the illustrative firm capture essentially none of it.

Daron Acemoglu, 2024 Nobel laureate in Economics and Institute Professor at MIT, has published the peer-reviewed math on this pattern across two centuries of industrial transitions. His conclusion, in his own words: Without countervailing power, gains flow to elites. Workers bear costs. That is not rhetoric. It is the default outcome under the current institutional set. Institutions change the default. That is what institutions are for.

Our peers have already built the answer

Canada is the outlier, not the pioneer. The comparison is uncomfortable.

• UKThe AI Opportunities Action Plan (January 2025). Skills England, a statutory body with a labour mandate, operational since 2024. The UK AI Safety Institute coordinates with the Department for Work and Pensions on a single displacement-tracking framework.
• EUThe AI Act (Regulation 2024/1689) established a labour-impact working group inside the European AI Office. Member states report sectoral displacement data. The EU Social Climate Fund is capitalised at €86.7 billion for 2026 to 2032, in part to cover AI transition costs.
• USDepartment of Labor AI and Workforce Task Force (2024, under executive order). State-level analogues in California, New York, and Illinois with compulsory impact assessments for state-funded employers above size thresholds.
• CANothing. No statutory body. No mandatory Workforce Impact Assessments. No public displacement dashboard. Not a single federal institution with a clear worker-side AI mandate.

What the body would do

A National AI Transition Body, established by federal statute, with four core functions:

1. Public displacement tracking. A live dashboard, updated quarterly, of sectoral AI exposure and observed employment change in Canada. Built on Statistics Canada's existing methodology, extended with firm-level data collected under compulsory reporting. This is a measurement regime, not a policy recommendation, of the kind the Bank of Canada already maintains for inflation and the Parliamentary Budget Officer already maintains for fiscal data.
2. Mandatory Workforce Impact Assessments. Any firm above 500 employees that plans to deploy an AI system expected to affect 10 percent or more of a job category must file a Workforce Impact Assessment, with a minimum 180 days notice. The Body does not veto deployments. It requires transparency, advance notice, and a stated transition plan.
3. Transition funding authority. An Employment Insurance-style transition programme, funded by a modest levy on corporate AI-related labour-cost savings above a threshold. Displaced workers receive income replacement, retraining if they choose it, and portable benefits during transition. The financing is drawn from the productivity gains that displacement creates, not from general revenues.
4. Reporting to Parliament, not just to the Minister. An annual report on the Canadian labour market under AI, tabled in the House of Commons, with subpoena power to compel disclosure from any firm subject to the Act. Ministerial accountability alone is insufficient when the affected population is every Canadian worker.

Attached to, but not inside, CAISI

The Canadian AI Safety Institute (CAISI) has a safety mandate: evaluate model capabilities, test for misuse, advise on deployment risks. Safety is a model-evaluation question. Labour displacement is a labour-market question. Folding them into one organisation creates a conflict of interest for CAISI's existing mandate and a talent-allocation problem for a small institute that is already stretched.

The Transition Body is a peer of CAISI, not a subordinate. Both report to the Minister for AI. Safety evidence and labour evidence are required to flow between them on a statutory cadence so neither can ignore the other's findings. This is the Office of the Auditor General / Office of the Parliamentary Budget Officer model, applied to AI.