Last week, AI researcher Jacob Coxon resigned from Anthropic, accusing the industry of “gambling with our lives.” A few days later, in a public statement, Dario Amodei, Anthropic’s chief executive, did more than agree with his former employee — he issued a warning of his own.

The industry, Amodei argued, must slow the development of increasingly capable models long enough for safeguards to catch up. Otherwise, AI systems capable of improving themselves could advance beyond our ability to understand or control them.

The predictions may sound extreme. But when the head of the company joins departing employees in calling for a slowdown, the warning can no longer be dismissed as the unease of a few disgruntled staff. It reveals that the people who understand these technologies best increasingly doubt that the industry can control what it is building.

In May 2023, Canadian computer scientist Geoffrey Hinton, now a Nobel laureate, joined dozens of experts in placing the risk of extinction from AI alongside pandemics and nuclear war.

Three months later, I wrote in the Ottawa Citizen that, despite the warning, governments displayed little sense of urgency.

Since then, AI capabilities have advanced dramatically. The warnings have multiplied. Yet the political response remains familiar: an announcement, a consultation, a few voluntary commitments — and then public attention moves on.

An answer to the wrong question

Ottawa’s response offers a case in point.

Asked about the recent resignations, federal AI Minister Evan Solomon acknowledged that the warnings were “not new” and insisted that regulatory and technical solutions exist.

He pointed to privacy legislation, digital-literacy programs, Canadian safety research and international collaboration.

All these initiatives may be worthwhile. But they fail to name the danger being raised: that a company could train or release a system it could no longer control.

Confronted with a warning about losing control of systems more capable than humans, Ottawa responds with data protection, public education and voluntary industry commitments. The threat is existential; the response is peripheral.

This disconnect exposes a troubling blind spot in our democracies.

The tyranny of the present

Our institutions are poorly equipped to confront dangers whose worst consequences lie beyond the next election.

The precautionary principle imposes immediate costs in exchange for future benefits that are difficult to measure. A government that slowed AI development today could lose investment and jobs. A government that prevented a catastrophe might receive credit only in history books.

The benefits of AI are immediate and championed by some of the most powerful corporations in history. Its risks are uncertain and borne largely by people who cannot yet perceive them — or who have not yet been born.

The interests profiting from the race are powerful and highly organized. Those who may bear its consequences are dispersed and often voiceless.

Financial markets have already priced decades of anticipated profits into AI-related companies. Nvidia, the world’s leading supplier of the chips that power AI systems, was valued at approximately US$5.6 trillion at the end of the previous week.

By Monday afternoon, its market value had fallen by more than US$300 billion. Analysts attributed the precipitous decline in Nvidia and other AI stocks largely to the repeated calls by Amodei and other industry leaders to slow the development of the technology.

The market’s reaction illustrates the political trap. A warning intended to protect the public can erase hundreds of billions of dollars in market value almost overnight.

Through index funds, retirement savings and pension plans, millions of people have acquired a financial stake in keeping the race going. Governments therefore face pressure not only from technology companies. They also fear that precautionary regulation could send markets tumbling, erode pension values and — as President Trump has shown — be portrayed as surrendering economic leadership to China.

One crisis feeding another

The dynamic hides a cruel irony. AI is often promoted as a tool for fighting climate change. It could indeed improve electricity grids, climate modelling and industrial efficiency.

Yet the infrastructure required to develop it is becoming not only an enormous consumer of energy and water, but also a major source of greenhouse-gas emissions.

By 2035, data centres could generate 350 million tonnes of emissions annually — nearly 60 per cent of the amount now produced by international aviation. In the United States, data centres directly consumed 66 billion litres of water in 2023. Producing their electricity required nearly 800 billion litres more.

We are exacerbating one existential crisis our institutions are failing to address in order to accelerate a technology whose risks they have not learned to control.

The climate experience should have taught us that uncertainty is not an excuse for inaction when the damage may be irreversible.

Instead, we are repeating the same mistake at much greater speed.

Naming the danger

Mark Carney brought the central banker’s practice of “forward guidance” into government. His YouTube videos have become a kind of digital-age fireside chat.

In his latest instalment, largely devoted to the Canada–U.S. trade dispute, he called AI a “crucial issue” and promised to devote a future video to it.

We will be watching. Will he speak only about digital sovereignty, investment and adoption? Or will he finally name the risk identified by the people building the most powerful systems: the real possibility that we could lose control of them?

Naming that danger must come before any credible solution.

Canada’s new national AI strategy promises safety, trustworthy systems and closer alignment with international standards. But its centre remains adoption, investment and economic growth. It aims to increase business adoption of AI from 12 per cent to 60 per cent by 2034 and to create hundreds of thousands of jobs.

Even the Canadian Artificial Intelligence Safety Institute illustrates the imbalance. It sits inside the same federal department responsible for promoting the AI industry. It can study emerging risks and evaluate models, but it is not an independent regulator created by Parliament. It has no statutory authority to compel access, impose corrective measures or stop a dangerous system from being released.

Parliament should change that. It should establish an independent AI safety authority in law, with a clear public-interest mandate and institutional independence from the departments promoting the industry. The authority should be entitled to examine advanced models and their testing data, order independent audits, impose corrective measures and suspend the training or deployment of systems that cross established danger thresholds.

Canada cannot solve this problem alone. Nor would simply assigning it to the United Nations create an enforceable system.

Ottawa should begin by working with the countries that control the most advanced chips, computing infrastructure and frontier laboratories. Together, they should establish common danger thresholds, mandatory independent evaluations and reciprocal verification. Access to advanced chips and computing capacity could then be made conditional on compliance.

That coalition could eventually provide the foundation for a broader international agreement. But credible governance must begin with the countries that possess the technical capacity and economic leverage to enforce it.

In 2023, the pioneers of artificial intelligence asked us to take seriously a risk they described as existential. In 2026, engineers are leaving their companies — and industry leaders are calling for a slowdown — because they no longer believe those companies can contain it.

Meanwhile, the machines keep learning. Our political leaders and institutions, not so much.