AI policy

Articles on this site are written with AI assistance — specifically, Anthropic’s Claude family of models. The workflow:

• Source-first. Every article starts from a research file with primary-source citations (21 CFR, FDA Food Code, SFCR, CDC data, named industry reports) prepared by Andrew. The AI cannot invent the regulatory framework; it works against the source pack.
• AI drafts prose; Andrew owns the claims. The model structures sections, suggests phrasings, builds out worked examples, and pulls together the supplied citations into readable paragraphs. Andrew reviews every paragraph before publication and edits anything that misrepresents a source or drifts from the operator voice.
• No claim ships unverified. Every regulation citation is checked against the primary text. Every statistic is checked against its source. If the AI confabulates a section number or misquotes a stat, the review catches it before publication.
• Andrew is the author. The byline is his, the accountability is his, and corrections (logged at /about/corrections) are tracked under his name. AI tooling does not appear in the byline because it is a writing tool, not an author.

Hero images on article pages are generated by Google’s Gemini Flash Image model. The prompts are written by Andrew, the generations are reviewed individually, and any image that contains legible text, identifiable people, or visual artifacts is regenerated or replaced.

Three things we do not do with generated imagery:

• No fake people, no fake faces.The style prompt explicitly forbids people, faces, and identifiable hands. Imagery is kitchen interiors, equipment, paperwork, and produce — not staged human subjects implying a relationship that doesn’t exist.
• No fake regulator logos, no fake credentials. Generated imagery never includes FDA, USDA, CFIA, or other agency marks; never includes fake inspection certificates; never depicts anything that could be mistaken for an official document.
• No misrepresentation as photography.When we describe an image as generated, we say so. When it’s a stock photograph or Andrew’s own photo of the Brantford facility, we say that too. (Most current hero imagery is generated.)

The HACCPlan app uses AI in two specific places, each with the guardrails described below. Neither feature is advertised as replacing your judgment as a Person In Charge (PIC), Preventive Controls Qualified Individual (PCQI), or licensed food-safety professional.

HACCP plan generator.The wizard walks you through the seven Codex principles for your specific process, then drafts plan language using Anthropic’s Claude. The output is a starting point that must be reviewed and approved by a qualified human (your PCQI, your QA manager, or you if you hold the appropriate accountability). The plan is not pre-approved by any regulator; it is a draft you customize and certify.

AI document scanner.Drop a Certificate of Analysis, a Safety Data Sheet, or a supplier spec sheet, and the scanner extracts structured fields, flags allergens and pathogens, and drafts a supplier-approval review. The scanner does not approve suppliers; it drafts a review that a qualified human signs off on. The extraction uses Anthropic’s models; uploaded documents are processed ephemerally and are not used to train any model under the API terms described in our Privacy Policy.

A short list of explicit no-go behaviors. If you ever see HACCPlan doing one of these, email corrections@haccplan.com.

• We don’t publish AI-drafted prose without human review by Andrew. There is no “ship without reading” mode.
• We don’t generate fake testimonials, fake reviews, fake case studies, or fake quotes attributed to real operators. Composite quotes are labeled “composite” with the role and segment.
• We don’t use AI to write articles whose value depends on a competitor being wrong. Comparison articles describe other tools by target market, never as wrong, behind, or bloated — see the editorial standards.
• We don’t use customer-uploaded documents (CoAs, HACCP plans, recipes) to train AI models. The API terms with Anthropic and Google specifically prohibit training on submitted content.
• We don’t present AI output as regulatory advice. The generator drafts; the human (you, or your PCQI) certifies.

Because every article on this site is AI-assisted under the policy above, we do not put a per-article AI-disclosure badge on each page — that’s what this page is for. The byline names the author (Andrew). The corrections log tracks every change to a published claim. The combination is the disclosure.

If we ever publish an article that does NOT follow this policy — for example, a guest post by another operator, or an article that isn’t AI-assisted — the article itself will say so up front.

AI tooling moves fast. Model versions change, capabilities expand, third-party API terms get updated. When we change how we use AI in any of the ways above — new model, new product feature, new image source — this page gets updated and the change appears in the corrections log (because it’s a published claim that changed).