How to check AI content before publishing

STEP 01 OF 10

Separate “is this true” from “can we actually publish this”

These are two different failure modes and need two different checks. An accurate AI-drafted claim can still be something you have no right to publish; a perfectly rights-clear piece of content can still be wrong. Run both checks, and do not let a pass on one stand in for the other.

For the accuracy half specifically, see how to sanity-check an answer from AI. This guide covers the rights and compliance half, which a purely factual review will not catch.

STEP 02 OF 10

Understand why AI output ownership is genuinely unresolved in Canada

Canada’s Copyright Act, current text machine-checked, contains no provision addressing computer-generated works and no text-and-data-mining exception. Its ownership structure instead runs on authorship: s.13(1) states, “the author of a work shall be the first owner of the copyright therein.” Where a work has no identifiable human author — because it was generated by a machine with no meaningful human creative contribution — the Act’s own architecture does not clearly answer whether copyright subsists in it at all, let alone who owns it.

The honest position is that this is unresolved, not settled in either direction. Do not publish AI-generated material on the assumption that your business automatically owns exclusive rights to it — and do not assume the opposite either. Where it matters commercially, treat the question as open and plan accordingly, including through the contract terms in Step 6.

STEP 03 OF 10

Check for material that looks like it came from somewhere specific

Generative AI systems learned from large volumes of existing content, and can reproduce something close to a specific existing work, style, or a real person’s likeness without you asking for that outcome directly. Before publishing, look specifically for anything that resembles a particular known source, brand, or identifiable person — a distinctive phrase, an unusually specific image style, a name that should have been generic.

A near-copy is a materially different problem from an accuracy error, and it is not caught by a fact-check — it needs a dedicated look, especially for anything visual or anything closely paraphrasing a specific source.

This matters because Canadian copyright liability does not require a full copy. The Copyright Act defines the right itself as “the sole right to produce or reproduce the work or any substantial part thereof in any material form whatever,” so an AI output that lifts a substantial part of one specific source — without copying it wholesale — can still infringe. Copyright Act, s.3(1)

STEP 04 OF 10

Do not reach for fair dealing to justify commercial publication

Section 29 of the Copyright Act is exact and complete: “Fair dealing for the purpose of research, private study, education, parody or satire does not infringe copyright.” That is a closed list. It does not include commercial use, and it does not include “AI training” or content creation for a business’s own marketing or publication. See Treadstone Law’s guide to fair dealing for Canadian businesses for how this defence is actually assessed.

If a piece of content’s justification for using someone else’s material rests on the claim that it was only used for inspiration, or that it was transformed enough, that reasoning does not map onto any of the five enumerated purposes in s.29 for ordinary commercial publishing. If genuine research, private study, education, parody, or satire is not what this actually is, do not rely on fair dealing to clear it.

STEP 05 OF 10

Check what the AI tool’s own terms say about the output

Before publishing content from a specific AI tool at any real scale, read what that tool’s terms of use actually say about who owns the output and what rights you receive. See what to look for in AI terms of use for the specific clauses that matter here — the terms, not general assumptions about how AI tools usually work, govern what you are actually allowed to do with what the tool produced.

This matters because the ownership uncertainty in Step 2 means the tool’s contractual terms may be doing more practical work than copyright law is, for now.

STEP 06 OF 10

Check ordinary advertising law separately — it applies whether or not AI was involved

Federal and provincial advertising and consumer-protection rules do not carve out an exception for AI-produced content. See advertising and marketing law for Ontario businesses for the general framework — it addresses the underlying rules on misleading claims and consumer protection, not any AI-specific issue, and the same standard applies regardless of how the content was drafted.

A misleading claim is equally a problem whether a person or an AI tool wrote the sentence. Run the same advertising-compliance review on AI-drafted marketing copy that you would run on anything else, without treating the drafting method as a mitigating factor.

STEP 07 OF 10

Decide, deliberately, whether to disclose that AI was involved

No general Canadian law requires disclosing that content was AI-assisted or AI-generated. ISED’s Voluntary Code does ask signatories managing public-facing systems, in the Code’s own words, to: “Ensure that systems that could be mistaken for humans are clearly and prominently identified as AI systems” — see the Voluntary Code of Conduct, relevant specifically where content is presented as if a person produced it in real time, such as a chat interface, rather than a byline on a drafted piece.

Even where nothing requires it, decide this on purpose rather than by default. A business that has thought through when it discloses AI involvement and when it does not has a defensible position; one that never considered the question does not. Where an image or video is involved rather than text, see how to verify a suspicious image or video for the provenance tools available to attach your own disclosure to outbound content, not only to check content coming in.

STEP 08 OF 10

Have a named person actually review the final version

Before anything goes out, a specific person — not a general review step nobody owns — should confirm the content passes the checks above. See how to decide what a human must approve for how to design that checkpoint properly, including what the reviewer should actually be checking rather than skimming.

This is the step that turns a checklist into an actual practice. A checklist that exists on paper but is not attached to a named reviewer tends to get skipped exactly when it matters most — under time pressure, on the piece that was rushed.

STEP 09 OF 10

Keep a dated record for anything with real exposure

For content with meaningful reputational, legal, or commercial stakes, keep a short note of what was checked, by whom, and when — the same discipline this hub recommends for AI documentation generally, applied specifically to a publishing decision. This record is what lets you show, later, that the process was followed rather than assumed.

This does not need to be elaborate. A dated line noting that the copyright, fair-dealing, and advertising checks were run, and by whom, is enough to turn “we have a process” into something you can actually point to.

STEP 10 OF 10

Revisit the checklist when the tool or the use case changes

A checklist built around one AI tool’s behaviour and one kind of content can miss real issues when the tool changes or the content moves into a new use — a new tool with different terms, a new format like video or voice, a new market with different advertising rules. Treat this checklist as a starting point tied to your current tools and current publishing habits, not a permanent, tool-agnostic standard.

Revisit it whenever either side of the equation changes meaningfully, in the same spirit as the review habit recommended for AI documentation and regulatory tracking elsewhere on this hub.