AI-Assisted Content Policy

Where assistance is allowed, and where it is not

• Allowed: drafting from a human outline

  A person decides the subject, the argument and the structure. Assistance turns that outline into prose, which is a writing task rather than an editorial one, and the result is reviewed like any other draft.

• Allowed: restructuring, tightening and formatting

  Reordering sections, shortening paragraphs, generating alternative headings and summaries, and applying formatting. None of these changes what the page asserts, and any that did would be a substantive edit requiring review.

• Allowed: first-pass translation support

  Machine translation may produce a first pass of French copy. A French-speaking reviewer edits it before it publishes, and a fluent mistranslation is treated as a factual error rather than a stylistic one.

• Allowed: internal consistency checks

  Flagging a page that contradicts itself, uses a term two ways, or states a figure differently in two places. The flag is a prompt for a person to check, never an authority to resolve it.

• Never: originating facts or statistics

  No factual claim and no number enters a RapidCents page because a model produced it. Every figure traces to the dataset or source that generated it. A plausible number from a model is a fabrication, whatever it happens to resemble.

• Never: selecting sources or deciding conclusions

  A model does not choose what a claim rests on and does not decide what a comparison concludes. Scoring inputs come from first-party sources under the comparison methodology and are entered by a person.

• Never: replacing subject-matter review or compliance sign-off

  Compliance, legal, security and regulatory content is reviewed and approved by the function that owns it. No assisted output is published on the strength of a model having produced something that reads correct.

• Fact-checking applies to model output more strictly

  Every factual statement in an assisted draft is verified against a source before publication, under the fact-checking policy. Confident phrasing is a reason for more scrutiny rather than less.

• A named human is accountable for every page

  Recorded internally, alongside the reviewing function. Assistance does not transfer accountability, and an error in an assisted draft is corrected under the corrections policy on the same terms as any other error.

• No fabricated people, quotes or credentials

  No invented authors, no synthesized customer quotes, no attributed statements nobody made, and no credentials that do not exist. Bylines name real accountable functions and testimonials come from real customers.

• Merchant and cardholder data is not an AI input

  Merchant transaction data and cardholder data are not used as input to general-purpose AI tools in the production of content. The research data ethics policy governs what may be analysed at all.