How we use, and do not use, AI

What AI may do here, on either path

• Draft a page structure from facts already recorded as source records
• Rewrite our own text for clarity
• Suggest questions readers may still have
• Flag sentences in a draft that are not backed by an attached source

What AI may never do here, on either path

• Introduce a measurement, a price, a material composition or a test result
• Describe an experience of wearing a product
• Rewrite another publication’s article into ours
• Generate a review, a rating, a tester or a quote

A generated draft is checked, sentence by sentence, for any number that is not one of the figures it was explicitly given. A draft with even one unsupported number does not publish on either path - it is held for review instead.

Path one: human-edited

Almost everything on this site follows this path.

1. Facts are collected as source records with a URL and an access date
2. Facts are normalised into fields on a product or a guide
3. A draft may be generated from those approved facts only
4. The draft is checked sentence by sentence for claims with no source
5. An editor revises it
6. Fact check
7. Reviewer sign-off, by a named person
8. Publication, by a named person

Our AI integration is optional in a technical sense as well: the site and the admin tool work fully without any AI provider configured, and this manual editorial workflow is the default path.

Path two: autonomous, mechanically gated

A small number of guides are proposed, drafted and published by a pipeline that runs on a timer, with a daily cap set in our admin tool (default: a handful a day, off by default). It does not have more freedom than the human path - it has less, and what replaces a person’s judgement is a mechanical gate instead of a human one:

1. The model is given only the exact figures one already-published, source-backed topic is allowed to state - not general knowledge, not the internet, not a number it computed
2. It drafts a new guide using only those figures, on a reader question we have not answered yet
3. The same fact-check pass every human-edited draft goes through runs again, unattended: any number it cannot trace back to that fixed list, any body-shaming phrasing, any of a list of restricted claim phrases, and the draft is left unpublished for a person to look at
4. Only a draft with zero findings publishes automatically

Guides published this way record JEGINS Fit Automated QA as author and reviewer in our database and audit trail - not a person’s name, because no person reviewed the draft before it went live. We do not print a byline on the article itself: a name at the top of a page is a claim about an individual, and this disclosure belongs to the whole site rather than to any one guide. The pipeline that runs it, and the setting that turns it on or off, live at scripts/pipeline/gpt-propose-topics.ts and are controlled from /admin/pipeline/automation.

The server this pipeline runs on also pulls and deploys its own code changes on a timer, under the same admin-controlled switch. Neither switch bypasses fact-check, and either can be turned off from the admin tool at any time.

Why we do not mass-generate pages

We keep a keyword backlog of several hundred questions. It is a backlog, not a publishing queue. Publishing hundreds of automatically generated pages would degrade the thing this site is for — being able to trust a measurement — and would make the search results worse for the people we are trying to help. Similar questions are answered on one page rather than split across several near-duplicates, and the autonomous path is capped low for the same reason.

Disclosure

If AI was used in drafting a page, that does not change who is accountable for it: the named author and reviewer are, whether that is a person on the human path or JEGINS Fit Automated QA on the autonomous one. If you find a statement on this site that no source supports, tell us via the contact form and we will correct it and record the correction in the page’s update history.