No — not with the confidence you'd want to rely on for a real decision.
Short answer
No. Unlike images, audio and video — where a tool such as Google’s SynthID embeds a watermark you can check for — AI-generated text has no reliable, universal fingerprint. Detection tools exist, but no Canadian regulator has published an accuracy figure for any of them, and a confidently written passage from a person and one from a well-prompted model can be genuinely hard to tell apart.
Google’s own description of how its watermarking actually works explains the gap. SynthID “embeds digital watermarks directly into AI-generated images, audio, text or video,” and for images and video the mark is “imperceptible to humans – but can be detected by SynthID’s technology.” Text works differently: “Large language models generate text one word (token) at a time. Each word is assigned a probability score… SynthID adjusts these probability scores to generate a watermark,” per DeepMind’s own page. The catch is built into the description — a token-probability watermark rides on the model’s own word choices, so paraphrasing, heavy editing, or running the text through a second model disturbs the pattern far more easily than cropping or compressing disturbs a pixel-level image watermark.
There’s also a coverage problem. SynthID is “embedded across Google’s generative AI consumer products” — by Google’s own account — and detection runs through Google’s own tools. Text from any other model carries no SynthID mark at all, and there is no independent, government-run detector that checks every model a Canadian business might actually be using.
Canada’s federal, provincial and territorial privacy commissioners have jointly told organizations deploying generative AI that they must “evaluate the validity and reliability of the generative AI tool for the intended purpose”, because “tools must be accurate throughout the intended lifecycle of the tool and across the variety of circumstances in which they are used.”
That principle cuts against relying on an unverified detector, not for it: if you use a text-detection tool to make a decision about a person — flagging a student’s submission, screening a job applicant’s writing sample — the same reasoning that requires you to validate any AI tool before trusting its output applies just as much to the detector as to the thing it’s checking.
Treat a detector’s flag as the start of a conversation, not a verdict, because nothing in this sheet vouches for any detector’s accuracy. If the stakes are real — academic discipline, an employment decision — get corroborating evidence before you act on a score alone.
The provenance problem looks different for images and audio, where a deepfake raises its own set of questions — see whether making a deepfake is illegal in Canada. And disclosure, not detection, is the tool most businesses actually reach for — see whether AI content needs a label. Treadstone’s AI Operations hub covers keeping an AI system accurate once it’s running, which is the harder problem behind this one.
See how ongoing accuracy gets checked once an AI system is live.
