Methodology and editorial rules
AI Policy Log records what official vendor pages say about training, retention, zero data retention (ZDR), data-processing agreements, and related exceptions, as of a stated check date. It is a quotation log, not a legal opinion.
Where the facts come from
- Every policy field on a tool page is taken from the vendor’s official documentation, help center, legal, or trust pages listed on that tool page.
- We include a direct quote, a source URL, and the date the page was checked (America/Toronto).
- If a field is not clearly stated on the pages we read, we write Not stated / not verified. We do not guess, infer from blogs, or fill gaps from memory.
- Accuracy comes before completeness. A partial page that is correct is better than a complete page that invents a rule.
What we will not write
- We do not say a product is “safe,” “not safe,” “HIPAA compliant,” or “GDPR compliant.”
- We do not treat training, retention, and ZDR as the same thing.
- We do not collapse consumer and business (or API) plans into one row when the vendor describes them differently.
- We do not put affiliate or referral links in policy tables.
- We do not auto-publish a policy change. The monitor script may draft a “what changed” note; a person has to approve it before the site is rebuilt with that change.
How plans are listed
One row per plan or product tier the vendor actually names on the official pages we read. If the vendor does not currently describe a commonly used informal name (for example a “Team” plan), we do not invent that row.
Exception cells are for cases the vendor itself flags: feedback that can still be used after an opt-out, temporary chats, Trusted Access or similar programs, features carved out of a BAA, aggregator/provider-specific rules, and so on.
How monitoring works
A script fetches each monitored official URL, stores a dated text snapshot, and diffs it against the previous snapshot. Trivial noise (cookie banners, asset hashes, “updated 2 days ago” relative timestamps without a policy change) is ignored. Meaningful diffs go into a review queue as pending human approval. They are never written into the published tool pages automatically.
Corrections
If a quote is wrong or incomplete, the fix is to re-read the official page, update the quote and date, and record the correction in the changelog.