Using AI drafts responsibly #

This page is for the owner and editor-in-chief of a team that uses the writing assistant (Exponential Intelligence). The software makes sure an AI draft cannot reach visitors without a person: it writes drafts only, with the editor's own permissions, labeled "Exponential Intelligence" (README.md). What it cannot do is check that a draft is true, fair and in your voice. That is the team's job, and this page gives you rules to start from.

Contents #

1. The principle
2. Checking a draft
3. What not to ask for
4. Images and alternative texts
5. Translations
6. Telling readers
7. A monthly review
8. Policy template

The principle #

The person who publishes is responsible for every word, as if they had written it. An AI draft is a starting point, not a source. Keep your approval workflow where it was; AI drafts go through it like any other version (Approvals and review).

Checking a draft #

• To do: Every fact, name, date, number and quote checked against a source you trust.
• To do: Nothing promised that your organization does not offer (prices, delivery times, opening hours, legal terms).
• To do: Links lead where the text says.
• To do: Tone and words match your house style; jargon removed.
• To do: Headings, lists and link texts follow the Accessibility checklist for editors.
• To do: For a change to an existing page: compare the draft with the published version on the versions page, so you see exactly what changed (Versions of an object).
• To do: For bulk work in Vector: review each change per field before anything is written, and use What changed with revert afterwards (expai).

What not to ask for #

• Content about real, named private persons beyond what your organization may publish anyway.
• Medical, legal or financial advice published without review by a qualified person.
• Text on topics or sections that your rules exclude from outside providers. Your administrator can enforce this: rules per section and content class decide which content may go to which provider (README.md).
• Personal data from form submissions, orders or user accounts pasted into a request.

Images and alternative texts #

The assistant can suggest alternative texts for images from a vision model, and a "Missing alt text" report can queue suggestions for a subtree (expai). Treat a suggestion as a draft: does it say what matters in this context, and does it avoid guessing who a person in the photo is?

Translations #

The assistant can prepare translations as drafts and keep existing translations in step after the source changes, patching only changed passages (Assisted translation with Exponential Intelligence). A language owner who reads the language reviews before publishing, through per-language approval where you have it (Translation routine).

Telling readers #

Whether and how to tell readers that a text was prepared with AI help is your organization's decision, and in some places a legal question; this page is not legal advice. Decide once, write it in the policy below, and apply it consistently.

A monthly review #

Policy template #

Template.

AI DRAFTS POLICY for ______________________   Agreed on: __________   Owner: ______________
1. A person reviews and is responsible for everything published.
2. Allowed uses: ______________________________________________________
3. Not allowed: _______________________________________________________
4. Sections/classes never sent to an outside provider: ___________________
5. Who may use it (policy expai/write): ________________________________
6. Approval required for AI drafts in: _________________________________
7. Disclosure to readers: ______________________________________________
8. Monthly limit: ________   Review by: ______________ every month

Related pages #

• Exponential Intelligence
• Drafts, versions and undo
• Personal data in Exponential
• Free documentation: this page is free to copy, translate and improve under the GNU FDL 1.3.