Trust Without Recourse: The Gap at the Centre of Canada’s AI Strategy

In May, a Toronto man decided to sell his car back to the dealership he had bought it from. He submitted an online inquiry. Someone named Quinn texted back, offered him $27,162.79, agreed to take his counteroffer to the manager, and booked a time to close the deal. Then a human sales consultant called informing him that the offer was void. Quinn was a chatbot which made the offer an error, and the real number was closer to $20,000. He got his original price in the end after the media picked up his story. This is exactly the problem, as we have no idea how many Canadians have similar stories, but either their stories did not get picked up by the media, or they did not have an avenue or resources to seek recourse.

What the strategy builds, and what it leaves out

AI for All, released on June 4, is a serious document, and its instruments are real: substantial spending, a strengthened AI Safety Institute, renewed standards work, privacy reform in Bill C-36, online safety and chatbot rules in Bill C-34. Trust is named as the north star of this strategy and a major challenge for AI adoption. But the strategy aims to build trust through training and literacy and is operationalized almost entirely as information: ensure Canadians are aware when AI is in use, label synthetic content, and establish a voluntary certification program for trustworthy systems.

In the government’s fall 2025 consultation report, Canadians called for strict liability laws, human-in-the-loop systems and robust policies to safeguard Canadian interests. Canadians asked for trust, and the strategy adopted nearly all of it. Certification, independent audits, disclosure, oversight bodies, literacy- all six pillars carry them. The one instrument that allocates loss rather than information, liability, appeared in the consultation summary only narrowly, under security, and did not survive into the strategy at all. The word does not appear in it, and it leaves the trust architecture load-bearing on one side only.

The omission of liability is not an oversight; it is a choice, and it is not obviously the wrong one, as is declining to revive AIDA. Horizontal AI statutes are slow, and Canada has already watched AIDA fail.

But the question of what happens to a harmed Canadian is not being decided by the AI strategy.

What we know, and how we know it

We have no idea how often this happens in Canada, because no one counts it. There is no registry of AI harms, no intake point, no reporting threshold, and no one to complain to. Canada has a precedent on AI misrepresentation, Moffatt v. Air Canada, because one passenger pursued a claim worth a few hundred dollars through British Columbia’s small claims tribunal. The Quinn case got resolved because a journalist made it expensive not to resolve it. So currently, Persistence, press attention, and professional scrutiny are Canada’s three functioning accountability mechanisms for AI harm. Anyone without the time to litigate a small claim, without a story a reporter wants, is simply absorbing the loss.

The people on the receiving end

The strategy makes a case for increasing adoption by businesses from 12% to 60%. Statistics Canada put adoption at 19.2 percent days before the launch, as Michael Geist has noted, and the government cited 12. The strategy identifies low trust and literacy as the binding constraints on adoption, but a Statistics Canada survey finds that 2 in 5 businesses cite a lack of relevance for using AI. The finding concerns businesses deciding whether to adopt but says nothing about the Canadians on whom adopted systems act, and adoption is climbing faster than the government’s own framing admits, which also means exposure to AI harms increases. Recourse will not persuade a business to adopt AI; a use case will. But it determines whether the Canadians on the other side of these systems keep accepting them. Adoption that survives contact with its own failures is the only kind that reaches 2034.

That exposure is least escapable in the public sector. A customer can walk away from a dealership. Nobody can opt out of a benefits determination, an immigration triage, or a tax assessment. The strategy commits to AI across public services, and it does not say what a person does when one of those systems is wrong about them. Similarly, the private businesses are not secure either. In January, insurers began writing AI exclusions into standard business liability policies, which means the small firms Canada is urging to adopt may be carrying losses they cannot transfer, discharge, or price.

This does not mean that Canadians will refuse AI. It shows something more fragile: adoption without recourse holds only until the first failure that people notice. A target that depends on nobody noticing is not a foundation, but a system designed to break before it can be fixed.

One thing Canada should do

The government has recently announced a consultation on AI transparency, running to September 23, and it asks the right questions and takes up the gaps the strategy left, including how Canada should track serious AI incidents. The Quinn story appears in the federal government’s own discussion paper on AI transparency, where it makes a reasonable point: people should be told when they are dealing with a machine.

The AI transparency consultation concedes the difficulty: reporting creates a record that exposes a firm to liability, so firms stay silent. That is an honest admission by the government, and it also tells us that without mandatory reporting, regulators and policymakers cannot create rules and laws to address problems arising from increasing AI adoption. Transparency is not the answer to accountability. It is the precondition for it, and Canada has been building the precondition alone.

What is missing is a route. Canada should establish an accessible complaint and dispute-resolution path for AI harms; one that a person without a lawyer, a reporter, or a winnable small claim can actually use to seek redress when harmed. Redress should sit with an independent body that specializes in harm rather than content, and that is not the same body setting the standards it would adjudicate against.

For AI-assisted decisions in the delivery of public services, as nobody can opt out of it, the role should rest with the Office of the Privacy Commissioner, which retains public-sector jurisdiction and needs enforcement powers to provide redress. The private-sector equivalent is a question C-36 is currently answering by default, with far less scrutiny than it deserves.

Two things would make that route work. First, the government should publish a standard of care: an annually updated statement of what reasonable safeguards look like for defined categories of AI deployment, based on the CAISI’s evaluation work and SCCs standards. It would give an adjudicator something to assess claims against, give firms a target they can hit, and give the voluntary certification program content it currently lacks. Second, a human alternative in essential services and public-sector decisions, so people can seek redress so that no one has to make enough noise to get the option the man in Toronto got.

This would also help in AI for All strategy’s goal of increasing adoption by Canadian businesses. When Canadians are assured that there is an avenue to complain and seek redress, it increases trust along with transparency, and thus increases AI adoption.

Trust is the right north star. But trust is not the belief that nothing will go wrong. It is knowing what happens when it does. He found out what happens: you hope a reporter picks up.

Note: AI was used to check structure and refine grammar; the substance and reasoning are my own.

Views are my own and do not represent the positions of the Canadian Science Policy Centre.