The AI Transparency Survey Asks Canadians the Wrong Questions
I completed the federal government’s AI transparency consultation a few days ago. Today, on the last day it’s open, I sent this piece to Innovation, Science and Economic Development Canada as a second, free-form submission — because the things I most wanted to say weren’t in the boxes.[1]
Somewhere around the third question in the second section, I realized I wasn’t answering as a citizen anymore. I was answering as someone who reads AI policy for interest. Most Canadians don’t, and the survey was not built for them.
Here’s the pattern. The consultation covers five areas: AI-generated content, knowing when you’re talking to an AI, information about what AI systems can and can’t do, incident reporting, and AI agents.[2] Each area opens with a question almost anyone could answer. Would it help you trust what you see online if you could tell whether it was made by AI or by a person? When does it matter most to know you’re talking to a machine? Those are good questions. They’re about lived experience, and ordinary Canadians are the only people who can answer them.
Then come the next three.
Who in the AI value chain — developers, deployers, or others — should bear responsibility for disclosure? How should a serious AI incident be defined, and how should reporting be structured to encourage proactive disclosure while addressing concerns around confidentiality and liability? Are existing frameworks sufficient, and if not, which instrument should government reach for: regulatory measures, codes of conduct, standards, research and development, literacy initiatives, or procurement requirements?[3]
Read that last one again. To choose between those instruments, you’d need to know what a procurement requirement does that a voluntary code doesn’t — but the policy vocabulary is the smaller barrier. The larger one is everything the question assumes you already know about the technology itself: how these systems get built and by whom, where in that chain a disclosure obligation could even attach, and how anyone currently checks whether a company’s claims about its own model are true.
That last part is the real gap. For most Canadians, a model card is unlikely to be a familiar object. They have no reason to know that the documentation companies publish about their systems is largely self-reported, inconsistent between firms, and rarely verified by anyone independent — a point the government’s own paper concedes.[4] If you don’t know how AI is vetted today, you have little basis for saying whether the vetting is sufficient, and you certainly can’t pick the right tool to fix it.
This isn’t a shortfall in civic engagement. It’s an information gap that only the companies themselves can close, and they have not, while the survey asks Canadians to answer over top of it. That’s not a question about your life. It’s an exam, and the passing grade is having worked in or near this field.
The framing problem runs deeper than vocabulary. To say whether existing practices are “sufficient,” you’d need to know what’s currently happening inside these companies, what safety testing found, what got disclosed and what didn’t, how often a safeguard failed before a product shipped. Canadians can’t see any of that. We experience AI systems from the outside. Asking us to certify the adequacy of frameworks we can’t observe isn’t consultation; it’s asking us to guess and then treating the guesses as evidence.
What makes this frustrating is that the discussion paper itself proves better questions were available. Its case studies are excellent and completely legible: a Toronto man negotiates a car buyback with “Quinn,” only to learn afterward that Quinn was a chatbot and the offer wasn’t real. A shopping agent misreads an instruction and orders twelve packages instead of one package of twelve. A retiree loses thousands to a deepfake video. A review of decisions published on CanLII identified 132 Canadian decisions between January 2024 and March 2026 in which a party cited a fictitious case as authority, with AI identified or inferred as the source in 96 of them.[5] Every one of those is a scenario a person can reason about from experience. The paper tells those stories and then asks about value-chain liability allocation.
The questions also lean
Not in a conspiratorial way — I don’t think anyone set out to rig this. But the asymmetries are consistent enough to notice, and once you see them you can’t unsee them.
The final section asks respondents where action would be premature. There is no companion question asking where action is overdue. One of those framings is available to you and the other isn’t, and that’s a choice someone made. On AI agents, the paper doesn’t even wait for your answer — it offers that whether government action would add value beyond emerging market and standards activity remains an open question, which is a conclusion wearing the illusion of a question.[6]
The sequencing does similar work. Each of the five sections walks you through pages of voluntary industry effort, such as invisible watermarks, provenance metadata, model cards, agent observability platforms, third-party attestation services, and only then asks whether existing practices are sufficient.[7] By the time you reach the question, you’ve been briefed by the defence.
And the costs only run one direction. You’re asked repeatedly how to weigh transparency against confidential business information, compliance burden on smaller firms, and the risk that too many disclosures leave people numb to them.[8] Those are legitimate concerns. But you are never once asked what non-disclosure costs the public — not in money, not in harm, not in the years it takes for someone to find out an algorithm was involved in a decision about their job or their credit.
To be fair, this isn’t a document built purely to justify doing nothing. It admits that text watermarking is brittle and not mature enough to rely on. It admits that system documentation is inconsistent, largely self-reported, and rarely verified by anyone independent. It admits there’s poor visibility into serious incidents.[9]
So I’d put it this way: the tilt is structural, not deceptive. This consultation was written inside Innovation, Science and Economic Development Canada — a department whose stated mandate is to help Canadian businesses grow, innovate and expand, and whose responsible minister also holds a regional economic development portfolio.[10] That is the vantage point from which the questions were drafted, and it shows in what they cover and what they leave out. Nobody has to intend a predictable answer for the instrument to produce one.
There’s a self-defeating loop in the survey as well. Buried in the menu of policy options it asks you to choose among is “literacy initiatives.”[11] The consultation is, in part, asking Canadians whether the government should help Canadians understand AI — in a format that already requires understanding AI. The people most affected by that gap are the least equipped to vote on closing it.
We have a recent precedent for why this matters. Last October’s AI strategy consultation ran from 1 to 31 October 2025 and drew more than 11,300 participants, who provided more than 64,600 responses to its 26 questions.[12] Over 150 individuals and organizations signed an open letter asking for more time and a rewritten survey.[13] Some signatories declined to take part at all and ran a parallel “people’s consultation” of their own.[14] ISED then used four AI models — Cohere Command A, GPT-5 nano, Claude Haiku and Gemini Flash — to analyze the submissions, with human reviewers validating and refining the AI-generated analyses, and described the result as thorough, unbiased reporting delivered months faster than traditional methods.[15] Michael Geist’s review of that report nevertheless found that it consistently softened the experts’ sharpest warnings into balanced policy language, creating what he called an illusion of consensus.[16] Whatever you think of that episode, the lesson is that response volume becomes the headline and “what we heard” becomes the mandate. If the instrument filters for people who already speak policy, the record will faithfully reflect the views of people who already speak policy — and it will be cited as the voice of Canadians.
The fix isn’t complicated. Split the instrument in two. One plain-language public track built on scenarios: here’s what happened to the man with the car, what should the dealership have had to tell him? One technical track for the people equipped to design reporting thresholds and liability regimes. Report the two separately, so nobody can blur a practitioner submission into a public mandate. Consultation as an accessibility problem, not just a communications one.
The window has now closed, so here’s what to watch instead. When the “what we heard” report appears, check three things. Does it report plain-language public responses separately from practitioner and industry submissions, or does it merge them into a single undifferentiated “Canadians told us”? Does it disclose how the analysis was done, and by what? And does it surface the disagreements, or does everything arrive pre-balanced, with every pillar a priority and no trade-offs named?
If the answers come back reassuring, I’ll say so. I’ve put this on the record beforehand precisely so that it’s checkable either way.
[1]Innovation, Science and Economic Development Canada, ‘Have Your Say on Advancing AI Transparency in Canada’ (ISED, 23 July 2026) <https://ised-isde.canada.ca/site/ised/en/have-your-say-advancing-ai-transparency-canada> accessed 23 September 2026. The consultation ran from 23 July to 23 September 2026, accepting responses through an online questionnaire or by email, with submissions treated as public documents.
[2]Innovation, Science and Economic Development Canada, Enhancing Trust in Artificial Intelligence Through Increased Transparency (ISED 2026) <https://ised-isde.canada.ca/site/ised/en/have-your-say-advancing-ai-transparency-canada/enhancing-trust-artificial-intelligence-through-increased-transparency> accessed 23 September 2026. The five issue areas, the consultation questions quoted or paraphrased throughout, the case studies, and the passages conceding the limits of current market practice are all drawn from this paper.
[3]ISED, Enhancing Trust (n 2).
[4]ISED, Enhancing Trust (n 2).
[5]Tom Macintosh Zheng, ‘The Rise of AI-Hallucinated Case Law in Canadian Courts and Tribunals’ (2026) 738 CanLIIDocs <https://www.canlii.org/en/commentary/doc/2026CanLIIDocs738> accessed 23 September 2026, cited in ISED, Enhancing Trust (n 2).
[6]ISED, Enhancing Trust (n 2).
[7]ISED, Enhancing Trust (n 2).
[8]ISED, Enhancing Trust (n 2).
[9]ISED, Enhancing Trust (n 2).
[10]Innovation, Science and Economic Development Canada, ‘Mandate’ (ISED) <https://ised-isde.canada.ca/site/ised/en/our-organization/mandate> accessed 23 September 2026 (‘ISED helps Canadian businesses grow, innovate and expand so they can create good-quality jobs and wealth for Canadians’). On the ministerial portfolio, see ISED, ‘Have Your Say’ (n 1), quoting the Honourable Evan Solomon, Minister of Artificial Intelligence and Digital Innovation and Minister responsible for the Federal Economic Development Agency for Southern Ontario.
[11]ISED, Enhancing Trust (n 2).
[12]Innovation, Science and Economic Development Canada, ‘Engagements on Canada’s Next AI Strategy: Summary of Inputs’ (ISED, 5 February 2026) <https://ised-isde.canada.ca/site/ised/en/public-consultations/engagements-canadas-next-ai-strategy-summary-inputs> accessed 23 September 2026.
[13]‘Canada’s New AI Strategy Is Off to a Bad Start’ BetaKit (10 June 2026) <https://betakit.com/canadas-new-ai-strategy-is-off-to-a-bad-start/> accessed 23 September 2026.
[14]Teresa Scassa, ‘Canada’s AI Strategy Consultation – Some Reflections’ (Teresa Scassa, 9 February 2026) <https://teresascassa.substack.com/p/canadas-ai-strategy-consultation> accessed 23 September 2026; ‘Human Rights Groups Launch “People’s Consultation” on AI after Criticism of Government Effort’ CBC News (25 January 2026) <https://www.cbc.ca/news/canada/british-columbia/human-rights-groups-artificial-intelligence-consultation-9.7057467> accessed 23 September 2026.
[15]ISED, ‘Summary of Inputs’ (n 12).
[16]Michael Geist, ‘An Illusion of Consensus: What the Government Isn’t Saying About the Results of Its AI Consultation’ (Michael Geist, 4 February 2026) <https://www.michaelgeist.ca/2026/02/aiconsultresults/> accessed 23 September 2026.