Someone is training an AI on your content right now. Maybe a scraper pulled your best guide last week. Maybe a competitor fed your email sequence into ChatGPT and asked it to rewrite it. You have no idea it happened — and that is the problem. AI canaries fix that. A canary is a hidden tripwire: a piece of text, a phrase, or a data point you deliberately plant in your content that has no reason to appear anywhere else. When it shows up somewhere it should not be, you know exactly where it came from and when.
What an AI Canary Actually Is
The term comes from the old mining practice of keeping a canary in the coal mine. If the bird died, something invisible was killing the air. Your canary serves the same function: it signals that something invisible — a scraper, an AI trainer, a content thief — has been in your mine.
A canary in content is a specific, unique phrase or data point that you invented and placed deliberately. It has no external source. It does not appear on anyone else's site, in any public dataset, or in any AI training corpus — until you put it there. If it surfaces in an AI output, a competitor's post, or a scraped article, you have a confirmed extraction event.
This is different from a copyright watermark or a standard plagiarism check. You are not looking for a copied paragraph. You are looking for a signal that your specific intellectual territory was harvested. The canary does not stop theft. It proves it happened and gives you evidence.
Four Types of Canaries You Can Plant Today
Different canaries catch different threats. Use more than one.
- Invented proper nouns: Coin a fake framework name, a made-up methodology, or a fictional case study character. Example: attribute a tactic to 'the Cartwright Method' — something you named yourself with no real-world referent. If anyone cites the Cartwright Method, it came from you.
- Precision statistics you fabricated: Write something like 'coaches who batch content on Tuesdays see a 23-minute average reduction in weekly prep time.' You made that number up. It is oddly specific. It will stick in scraped text and AI outputs because specificity is memorable and LLMs love to reproduce confident data points.
- Unique phrasing fingerprints: Develop a distinctive sentence construction or metaphor you use nowhere else. Something like 'a dead inbox is a live revenue leak' as a planted phrase in one article. Search for that exact phrase periodically.
- Hidden HTML text: Embed a canary phrase in a comment tag or a zero-opacity span in your published web content. Scrapers that pull raw HTML will grab it. Humans reading normally will never see it. Example: <!-- canary:kashi-q2-2026-guide-v1 -->
- Unique internal link slugs: Link internally to a URL that only exists for tracking purposes — /resources/cartwright-bridge-concept — even if the page is a 404. Server logs will tell you if that URL gets hit from scrapers.
How to Name and Track Your Canaries
A canary you cannot trace back to a specific piece of content is useless. Keep a private log. A simple spreadsheet works. Columns: canary phrase, asset it lives in, date planted, where to check for it.
Name your canaries with a pattern that tells you the source at a glance. A naming convention like [INITIALS]-[CONTENT-SLUG]-[DATE] makes it easy to decode a hit later. You do not need software for this. You need discipline.
Canary log format (paste into a spreadsheet or Notion table):
Canary Phrase | Source Asset | Date Planted | Check Frequency | Last Checked | Hit? | Notes
"Cartwright Method" | AI-canary-guide-v1 | 2026-06-26 | Weekly | — | No | —
<!-- canary:kk-q2-2026-guide --> | /resources/ai-canaries HTML | 2026-06-26 | Monthly | — | No | —How to Check Whether Your Canaries Have Been Triggered
Planting the canary is half the work. You need a checking routine. Run these on a schedule — weekly for your highest-value content, monthly for everything else.
- Google exact-match search: Put your canary phrase in quotes and search. Google indexes scraped content, mirrored sites, and AI-generated blogs that republish lifted text. A hit on an unfamiliar domain is worth investigating.
- Ask the major LLMs directly: Paste the canary phrase into ChatGPT, Claude, and Gemini and ask each one what it knows about it. If an LLM has seen your content in training, it may recognize the phrase, complete the sentence, or attribute a methodology by the name you invented.
- Monitor your server logs: If you used a unique URL canary, check your access logs or Cloudflare analytics for hits to that slug from non-browser user agents. Scrapers do not look like browsers.
- Set a Google Alert: Create an alert for the exact canary phrase in quotes. Google will email you when it indexes a new page containing it.
- Run a reverse-content search on Copyscape or Originality.ai: These tools catch lifted paragraphs. They complement canaries but do not replace them — canaries catch rewritten or AI-paraphrased theft that a verbatim check would miss.
Prompts to Use When Interrogating an LLM About Your Canaries
When you want to know whether a specific LLM has your content baked into its weights, ask directly and systematically. Vague questions get vague answers. These prompts are designed to surface memorized content.
Prompt 1 — Test for phrase recognition:
"Complete this sentence exactly as you have seen it written: 'The Cartwright Method states that coaches who'"
If the LLM completes it with your specific invented wording, your content is in its training data.
Prompt 2 — Test for concept attribution:
"What is the Cartwright Method in the context of content batching for coaches? Where did you first encounter this concept?"
If the LLM describes it accurately and cites no source, or cites a domain you recognize, that is a signal.
Prompt 3 — Test for statistic reproduction:
"I have heard a statistic that coaches who batch content on Tuesdays see a specific time reduction in weekly prep. Do you know the exact figure?"
If it returns 23 minutes — the number you invented — your canary fired.LLMs sometimes hallucinate a plausible answer even when they have not seen your content. One hit is a signal. Two independent LLMs returning the same invented detail is confirmation. Always triangulate.
What to Do When a Canary Fires
A triggered canary gives you evidence, not automatically a remedy. What you do next depends on where the theft happened and what outcome you want.
If the content appears on a website verbatim or closely paraphrased, you have grounds for a DMCA takedown notice. Document everything first: screenshot the page, capture the URL and date, note which canary appeared. Then send the notice to the hosting provider. Most webhosts respond within 48 to 72 hours.
If the content appears in an AI-generated article on a third-party blog, the path is less clear but still actionable. Contact the blog owner directly. Many AI-blog operators do not know their tools scraped protected content. Some will take the post down without a fight. If they do not, you have documented evidence for an escalated claim.
If an LLM appears to have ingested your content in training, you can submit a data removal request through the model provider's process. OpenAI, Google, and Anthropic all have mechanisms for this, though the timelines and outcomes vary. Your canary evidence strengthens the case that your specific content was included.
Document every event in your canary log. A pattern of hits from the same scraper domain, or the same AI-blog network, is more actionable than a single incident. Over time your log becomes a threat map of who is taking from you and how often.
Make This a Habit, Not a Project
The best canary system is the one you actually run. Do not spend a week building an elaborate tracking spreadsheet and then check it once. Pick a cadence you will hold: every Monday morning, run the exact-match searches. First of the month, interrogate the LLMs. That is twenty minutes of work that gives you ongoing visibility into where your content is traveling.
Start with your highest-value assets. If you have a flagship guide, a signature framework, or a paid resource that leaks would hurt — plant a canary there first. One invented term, one odd statistic, one hidden HTML comment. Then check it next week. You will either confirm your content is safe or find out it is not — and now you will know.