Stale sources turn AI content into expensive cleanup

An AI-assisted article can be fluent, on-brand, and wrong because the source behind one sentence changed six weeks ago.

The cost is rarely the sentence itself. It is the work that follows: a subject expert traces the claim, a writer repairs the section, a reviewer checks every similar statement, and the launch moves while the team argues about which version is current.

Freshness is part of content cost. If a source cannot be dated and owned, it will be rediscovered during review.

Not every source needs the same clock#

Treat source freshness as a rule based on the claim, not a single expiration date for every document.

These are starting points, not universal policy. The important part is that the team can answer why a source is considered current and who can renew it.

Preserve the source set with the request#

A link to a folder is not enough. Keep the source identity at the content boundary:

When the asset is reviewed, the person should be able to see what the request used, not what happens to be in the folder today. That distinction matters when a page changed after drafting.

A source failure has a recognizable shape#

You can spot stale-source work in the production data:

• the same claim is returned by multiple reviewers;
• a draft uses a version that is older than the campaign brief;
• product marketing edits a sentence that appeared in several assets;
• the content owner cannot identify which source supported a claim;
• a reviewer asks for a screenshot or link after the draft is complete;
• the final asset has a different fact than the original request.

Track the return reason as source_missing, source_stale, source_scope_unclear, or claim_not_supported. “Needs edits” is too broad to improve the source process.

Example: the cheap draft that cost a week#

An AI tool drafts a product comparison page from a help center export. The export is 45 days old. During review, product marketing discovers that an integration limit changed. The page has 18 related claims, so the writer checks each one. Legal asks for a new citation. SEO rewrites two sections after the comparison changes.

The model request cost $3.40. The repair used:

At a blended $95 hourly cost, the stale source added roughly $546 and delayed the page. The provider bill did not move.

Make freshness a decision before generation#

Before a request starts, the process should choose one of four outcomes:

1. Use the source. It is current, scoped, and owned.
2. Refresh the source. A named person updates or revalidates it.
3. Narrow the claim. The source supports a smaller statement.
4. Hold the work. The asset cannot responsibly proceed.

Do not silently use an old source because the request is urgent. Urgency can justify a smaller claim or a clearly labeled draft. It does not make an expired fact current.

The audit-row guide explains the evidence fields needed to reconstruct an automated action. The same discipline helps content: source IDs, versions, owners, and outcomes turn a late fact check into a visible process decision.

Put source freshness in the cost review#

At the end of the month, ask:

• How many accepted assets used expired or unscoped sources?
• Which source owner caused the most returns?
• How many review minutes were spent finding evidence?
• Which claims appear across the most assets?
• What source should be updated once instead of repaired many times?

This is a better improvement list than “write better prompts.” A current source set gives the model less room to invent, gives the writer less to repair, and gives the reviewer a smaller surface to inspect.