Fluency Is Not Competence: Canada’s AI strategy measures AI use. It should measure judgment

Hugh Laurie has never diagnosed a patient. Neither has George Clooney, nor Ellen Pompeo. They have played doctors so convincingly that a generation associates their faces with medical authority, yet you would not hand one of them your chart. The performance is flawless; the competence behind it does not exist. We grasp this instantly when the white coat is a costume. We are far worse at seeing it when the costume is fluent text on a screen.

That is the thing to understand about a large language model. It is a performance of expertise. It produces the cadence, the vocabulary, and the unhesitating confidence of someone who knows, because it has learned, from an enormous corpus, the pattern of how knowing sounds. Whether the answer is correct is a separate question, and not one the fluency can answer for you.

Canada’s new national AI strategy, “AI for All,” is built to put this technology into as many hands as possible. It promises free AI literacy training, tools for roughly a million post-secondary students, and tens of thousands of job placements, all aimed at lifting Canada off the bottom of the international rankings, where we currently sit 44th of 47 countries on AI literacy and 42nd of 47 on trust in AI systems. The ambition is right. But the strategy talks about literacy at length without ever measuring the thing that matters. To its credit, it calls for Canadians who can recognize AI and judge its outputs. Yet every target beneath that language counts reach: a million students trained, three thousand teachers equipped, tens of thousands of placements, an AI agent in every student’s hand, business adoption pushed from one in eight firms toward most of them by 2034. Not one asks whether a single Canadian got better at telling a confident wrong answer from a right one. Judgment is named in the prose and missing from the metrics, and what a strategy does not measure, it does not build. Trust is named as the strategy’s north star, but it is pointed almost entirely at institutions: modern rules, a national safety institute, content watermarking. Very little is pointed at the judgment of the individual Canadian staring at a confident answer and deciding whether to believe it.

We use fluency as a proxy for competence because, in people, it usually works. Someone who explains a problem clearly, in the right register, without fumbling, has typically earned that fluency by actually understanding the thing. The heuristic is fast and mostly reliable, which is why it is so deeply wired into how we judge each other. Large language models break it. They deliver maximal fluency with variable accuracy, and they deliver it identically whether they are reasoning soundly or inventing. The single signal we have always trusted has been quietly decoupled from the thing it used to signal.

This matters most precisely where the strategy invests most. Government is being positioned as an anchor customer for Canadian AI. The priority sectors are high-stakes by design: health, energy, the systems that ordinary people cannot easily check for themselves. Every one of those settings is a place where a fluent wrong answer costs more than a hesitant right one, and where the person on the receiving end is busy, trusting, and primed to read confidence as correctness. In medicine, I watch how readily a clear, well-phrased answer earns trust it has not earned on the merits; as an educator, I see the same in trainees who will trust a confident chatbot over their teacher or their peers, because the machine never sounds unsure -the same reflex runs in a courtroom, a benefits office, and a boardroom. The more fluent the output, the more it disarms the very scrutiny it most needs.

Real literacy is not knowing how to write a prompt. It is calibration: the habit of separating how good an answer sounds from how likely it is to be true, knowing where these systems characteristically fail, and verifying before relying. It is the discipline to grow more skeptical of the polished answer, not less. The strategy’s checklist, spotting bias, misinformation, and privacy risks, comes down to one instruction: check the output. That is the easy part. The hard part it never names is that we only check what looks doubtful, and AI never looks doubtful. The confident, well-written answer is the one that slips through unchecked. A program that raises confidence without raising discernment does not close the gap the strategy is worried about; it widens it, because it produces people who use AI more and question it less. That is not a literate population. It is a more efficiently misled one.

So the measure of success cannot be headcount. A million Canadians trained to operate AI tools, with no change in their ability to catch the tool being wrong, is not the outcome the strategy says it wants. If trust is genuinely the north star, then the thing to build, and the thing to measure, is calibrated trust: whether people can tell, under real stakes, when to lean on the machine and when to override it. That is harder to count than licences and placements, which is exactly why it will be tempting to leave out, and exactly why it has to be left in. Literacy that does not change behaviour under pressure is not literacy; it is exposure with a certificate.

Hugh Laurie reading the script of Gregory House is reassuring and useless at the same time. If you desperately needed the diagnosis, you would want House in the room, not the actor who plays him, however persuasive. The danger of this technology is not that it is a bad actor. It is that it is a superb one, available to everyone, at every hour, and never once breaking character. AI for All is the right goal. It will only work if that “all” includes Canadians who can sit through the performance and still ask whether anyone here actually knows the answer.