Earlier this year, the Government of Canada asked Canadians a series of questions meant to inform whether, and how, government should get involved in building public trust in AI. Among them was whether AI-generated content should be marked or labeled, what form that should take, and who should be responsible for it. The consultation details are available on the Government of Canada website. It closed to submissions on September 23, 2026.
This post draws on the submission I filed to the consultation. In it, I argued that regulating this area prematurely risks undermining the very public trust it seeks to build. It could also disadvantage Canadian creators, innovators, small businesses, and the individuals AI has empowered most — costs that would outweigh the harm the rule is meant to prevent.
Marking cannot be made reliable — and that helps the wrong people
Requiring every piece of AI-generated content to carry some kind of mark — an invisible or visible watermark, a disclaimer, embedded metadata — sounds like a simple fix. In practice, it cannot be reliably enforced. AI models developed in less regulated jurisdictions are unlikely to comply, and even compliant marks can be removed or tampered with by a sufficiently motivated bad actor.
The risk of false confidence
If a labeling rule exists but cannot be reliably enforced, honest creators will label their work and dishonest ones will not. Over time, people may start to assume that anything without a label was made by a person. A rule that cannot be fully enforced risks manufacturing false confidence rather than genuine transparency.
The risk of pushing people toward worse tools
There is also a risk that a labeling requirement pushes people in the wrong direction. If the public or organizations start refusing or deprioritizing AI-assisted content, more people may turn to models built to evade detection, growing the market for the bad actors behind them. We have already seen this pattern: colleges' efforts to detect AI-generated writing drove students toward paid "humanizer" tools designed to defeat detection software NBC News. The same applies to the models themselves: if watermarked content from compliant models is more readily flagged, restricted, or filtered out, creators gain a clear incentive to switch to models that apply no marks at all, including those built by bad actors or in less regulated jurisdictions. The result: good-faith developers lose users, while the very models the policy is meant to discourage gain them.
Platform enforcement would mean over-blocking
If hosting platforms are required to detect and label AI content themselves, the likely result is over-blocking. Accurate detection is costly. Wrongly blocking content costs almost nothing. This is especially true in a market as small as Canada's, where platforms have less incentive to invest in careful review. Small businesses have already lost social media accounts to wrongful flags, and their appeals rarely succeed because platforms devote few resources to reviewing them. A labeling mandate would likely make this worse, and small creators would bear the greatest cost.
Marking blurs a line that should stay clear
A label also fails to draw the distinction that actually matters. An idea can be entirely a person's own, even when AI helped express it. That is true whether AI is used to tighten a paragraph, retouch a photo, turn a rough sketch into a finished illustration, or edit and caption a video. A blanket labeling rule cannot distinguish between "I used AI to bring my own idea to life" and "AI produced this with no human input at all." Treating both cases the same penalizes the people AI has helped most: those without the time, training, or budget to write polished copy, hire a designer, or produce professional video. If enforcement is needed, it should target fully automated content, not the much larger group of people who simply use AI as a productivity tool.
A better path: public literacy
The more realistic path to public trust is investing in public literacy: helping people build a habit of healthy skepticism and lean on trusted, credible sources that do their own due diligence. This works well because it does not depend on government enforcement — credible outlets have their own reason to be transparent about their sourcing, since their reputation depends on it.
Detect harmful content, not AI-generated content
None of this means harmful content should be ignored, or that hosting platforms have no role in monitoring content. Their focus should be on flagging and removing harmful content, a narrower but far more important task. That responsibility should sit with the platforms that host and profit from it and be tied to actual, demonstrated harm. Reporting has suggested that a meaningful share of one major social media platform's 2024 advertising revenue came from ads for scams and banned goods, and that its platforms were linked to a large share of successful scams in the United States Reuters. That is where accountability belongs. The right role for government is clear legislation that makes it easier for people to hold platforms responsible in court for real harm, financial or otherwise. Clear liability rules give platforms a direct financial incentive to address harmful content on their own, and the approach could serve as a model for other major markets.
A note on timing
It is also worth watching how marking regulations for AI-generated content play out elsewhere first. The European Union has already introduced watermarking requirements, but it is too early to know whether that approach will work at scale. Canada's smaller market has less room to absorb the cost of a policy that turns out not to work, so it makes more sense to watch the EU's rollout closely and adopt a similar approach later if the evidence supports it, rather than run a second, unproven experiment now.
There is a further reason for caution. Canada is not making this decision in isolation. If domestic rules make it meaningfully more difficult for Canadian businesses and creators to build with AI, that gap will not remain unfilled — it will be filled by foreign companies operating under different, and possibly less stringent, rules.
Conclusion
Mandatory labeling is a natural first instinct, but it is unlikely to hold up. It cannot be made technically reliable, it cannot tell the difference between AI-assisted and fully automated content, and its costs would fall hardest on the people and businesses least able to absorb them. A more durable approach combines two things: real accountability for platforms where demonstrated harm occurs, and sustained investment in helping Canadians evaluate what they see online.
This is the first of three pieces based on the submission I filed to Canada's AI transparency consultation. The next piece looks at the consultation's other four issue areas, and the third raises a concern that fell outside the consultation altogether.