← All insights
Insights · Responsible AI

Canada's privacy regulators set out six principles for AI in health. Here is where each one lives inside a health app.

On 29 October, Canada's federal, provincial and territorial privacy commissioners and ombuds published a resolution on the responsible use of artificial intelligence in health and health research. It is addressed to governments, AI developers and service providers, health care providers and researchers. Linnea is a consumer health app, not a clinical system, and the resolution does not name products like it. But it puts AI to work on health information every day, so we read the six principles as a checklist: where each lives in the product, and where it does not.

Published · 5 min read · By Arctura Technologies

Editor's note, 6 November 2026, updated 9 November 2026: Version 3.3 of Linnea's privacy policy, which took effect on 9 November 2026, adds one qualification to the "never sent" list described under Privacy protective below. A new Family AI section allows two kinds of family information to be sent to an AI provider, each only after the members concerned agree: the taste tags members set themselves and the number of people at the table, for the Family Menu; and, for a weekly family summary, the week's numbers, such as meals eaten together and the weekly average score, steps and sleep of members who agreed and share them, each under a code such as "member A" rather than a name. A managed member, who has no account of their own, is covered by a narrower exception: only their tastes and the number of days something was logged for them, and only when every person who manages the profile agrees. Records and medications are never sent, and Family AI can be turned off in You › AI features and data use. The rest of what this article describes is unchanged. Version 3.2 is kept at arcturatech.com/linnea/privacy-3.2.html; the current policy is at arcturatech.com/linnea/privacy.html.

What the regulators asked for

Adopted at the regulators' annual meeting in Ottawa in September, the resolution is a call to action, not a law. It asks governments to review or adopt legislation so that AI systems in health are built, bought, used and retired in a way that meets six core principles: "privacy protective, valid and reliable, safe, human rights affirming, transparent, and accountable". It asks developers and service providers to design for privacy from the start, minimize data, keep the most protective settings as the default, test in the real world and keep validating, and give people a way to challenge an output. It asks health care providers to explain plainly how information is used and to name a person people can bring concerns to.

The harms it names are concrete: unauthorized collection, use and disclosure of personal health information, inaccurate inferences, systemic discrimination, and patient harm from outputs that were wrong. None is a hospital-only risk. An app that recognizes a meal from a photo is making an inference from health information too.

Six principles, six places in the product

Everything below comes from Linnea's privacy policy (version 3.2, effective 28 October 2026), the AI Notices page, the "Linnea, measured" page and the App Store listing.

Privacy protective. The app asks first: before an AI feature sends anything, it shows what that feature sends and to whom, and nothing is sent until you agree. It sends only what the feature needs, such as the photo you submitted or your question with the related entries; your name, email address and date of birth are not added, though anything inside your own text or photo goes with it. Raw Apple Health and Health Connect data, health records, entries for managed members and what family members share with you are never sent. The policy names the providers, says they may not use requests to train their general models, and caps retention at 30 days for abuse monitoring.

Valid and reliable. The resolution asks for real-world testing and a way to catch mistakes. We publish our own error rates, with method and date under each number: a median calorie error of 9.5% for a meal described in words or by voice, on 480 test meals in 8 languages in October 2026; 5.7% for a dish computed from its ingredients and weights, on 400 weighed Nutrition5k dishes; and 22.4% for a day estimated from meal photos, on 60 weighed dishes never used to tune Linnea. Photos are the least accurate way to log, so every AI result is a list you check before saving and can adjust. Nutrition values are estimates for logging. The numbers are measured by us and audited by nobody.

Safe. Linnea is not a medical device and does not give medical advice. It does not diagnose or recommend treatment, and it never suggests changing a medication or dose. A scanned lab report is read on your device and you check each value before saving; Linnea shows it against the reference range printed on the report and does not interpret it. Ask Linnea answers from your entries only.

Human rights affirming. This is the principle we can claim least. The resolution's worry is bias against people under-represented in training data, and we cannot prove Linnea is free of it. Meal recognition from text or voice is measured in 8 languages, against official food tables from 8 countries and regions. The score leaves out any area with no data and says how many areas it used, rather than guessing. Linnea never infers a taste from records or medications. And AI is optional: turn it off and the rest of Linnea keeps working, so nobody has to accept AI processing to keep a health record.

Transparent. The Privacy Commissioner's office made a related point in its submission to the federal AI transparency consultation, published on 23 October: information about an AI system has to be understandable, and has to reach people at the point where they are deciding. In Linnea that point comes twice. Before anything is sent, when the app asks. And beside every result, as a version-stamped notice stating its limits; the Today score's reads: "Today score is a behavioral wellness observation, not medical guidance. Discuss persistent changes with your clinician." A notice changes only together with its version stamp.

Accountable. The resolution asks for a named contact and a practical way to challenge an output. Every AI output can be challenged in the most direct way, by correcting it before it is saved. Privacy questions, including for our privacy officer, go to privacy@arcturatech.com. Earlier versions of the privacy policy stay online with their dates, and when a change touches something we have written here, we add a dated editor's note instead of rewriting the article.

What we do not claim

We do not claim compliance with the resolution. Whether anything meets it is a regulator's judgment, not ours, and several of its expectations have no counterpart here: we have no third-party audit of the error rates we publish, and Linnea is not an AI system in health care in the resolution's sense. None of the mechanisms above makes the models correct. They make mistakes visible and correctable, which is a different and more honest promise. Our policy says the same about security: no system is completely secure, so we do not guarantee it.

Four questions for any app that puts AI on your health data

The six principles reduce to four questions a health app should answer in plain words. What does it send, to whom, and did it ask before the first time? What does each result say it is not, and is that sentence beside the result or buried in a policy? How accurate is it, measured how and when, and can you correct a result before it is saved? Can you switch the AI off and keep your record? An app that answers all four is doing most of what the regulators asked for, whether or not they had it in mind.

The six principles fit on one page. Our policy and our AI notices fit on two more. Read them side by side, and tell us where a principle still has no counterpart.

Linnea provides organizational and informational tools only; it does not provide medical advice, diagnosis, or treatment. Always consult a qualified health professional for medical decisions. Nutrition values in Linnea are estimates for logging. The resolution discussed here is a statement by Canada's federal, provincial and territorial privacy regulators addressed to governments, health organizations, AI developers and researchers; nothing in this article claims compliance with it, and this article is not legal advice. Apple Health and App Store are trademarks or service marks of Apple Inc.; Google Play and Health Connect are trademarks of Google LLC. They are referenced for identification only, and Arctura Technologies Inc. is not affiliated with or endorsed by Apple, Google, or any privacy regulator.

Sources

  1. Office of the Privacy Commissioner of Canada, news release, 29 October 2026 — "Canada's federal, provincial, and territorial privacy commissioners and ombuds call for the responsible use of artificial intelligence in health and health research": priv.gc.ca/en/opc-news/news-and-announcements/2026/nr-c_261029/
  2. Federal, Provincial and Territorial Information and Privacy Commissioners and Ombuds — Resolution, "Responsible use of artificial intelligence in health and health research", adopted in Ottawa on 15 September 2026: priv.gc.ca/en/about-the-opc/what-we-do/provincial-and-territorial-collaboration/joint-resolutions-with-provinces-and-territories/res_20260915_aihealth/
  3. Office of the Privacy Commissioner of Canada — Submission on "Enhancing trust in artificial intelligence through increased transparency", dated 23 September 2026 and published 23 October 2026: priv.gc.ca/en/opc-actions-and-decisions/submissions-to-consultations/sub_ised_20260923/
  4. Linnea Privacy Policy, version 3.2 (effective 28 October 2026), sections 3, 4, 5, 9 and 13, kept for reference: arcturatech.com/linnea/privacy-3.2.html
  5. Linnea Privacy Policy, version 3.3 (effective 9 November 2026), current, section 3 (Family AI): arcturatech.com/linnea/privacy.html
  6. Linnea AI Notices, version 1.2 (effective 20 October 2026): arcturatech.com/linnea/ai-notices.html
  7. "Linnea, measured", recognition accuracy and how it was measured (October 2026): arcturatech.com/linnea/strengths.html
  8. Linnea on the App Store, listing description as of 30 October 2026: apps.apple.com/app/id6765480037