AI content review: the citation gate that stopped a confident wrong answer

The draft arrived clean: short paragraphs, a calm tone, and a clear answer to a customer question. It also contained one sentence the company could not support. That sentence was not a typo or a wild hallucination. It was a plausible claim built from an outdated policy, and it would have shipped because the draft looked finished.

This case describes a content-review gate that separates draft quality from claim support. The team still used AI to draft. It stopped treating a fluent draft as evidence that the claims were true.

A polished draft is not a supported draft

Editors are trained to notice structure, tone, repetition, and clarity. Those checks are necessary, but they do not answer a different question: is each consequential claim supported by a current, authorized source?

A draft can pass every style check and still be wrong in ways that matter:

• a refund window that changed last quarter;
• a feature that exists only in a roadmap;
• a price that varies by region;
• an eligibility rule that depends on the customer’s plan;
• a legal or medical statement that needs specialist review.

The team separated three review layers:

Layer	Question	Who checks	Output
Structure	Is the draft readable and complete?	Editor	Edit or approve
Claim support	Is each material claim backed by a current source?	Reviewer with source access	Pass, revise, or block
Authority	Are we allowed to say this?	Policy, legal, or product owner	Approve, narrow, or remove

A single reviewer can perform more than one layer, but the layers cannot be collapsed into “looks good.” The claim layer needs the source open. The authority layer needs someone who can commit the company to the statement.

Separate structure, claim support, and authority

The team first marked the claims that could change a customer’s decision or the company’s obligation. Not every sentence needs a citation. These do:

• numbers, dates, deadlines, and prices;
• eligibility, exclusions, and conditions;
• what a product can or cannot do;
• legal, medical, safety, and compliance statements;
• comparisons with a competitor;
• promises about support, refunds, or timelines.

Then it wrote a rule for each claim: no current source, no publish. A source was not just a document name. It needed a version, an effective date, an owner, and a passage that actually supported the sentence.

The reviewer did not ask the model to “check its answer.” That would have produced another fluent paragraph. The reviewer opened the source and compared the claim with the passage. The model could help locate candidates; it could not be the final witness.

Example: the refund window that changed last quarter

The draft said: “Annual plans can be refunded within 30 days of purchase.”

The cited document was a real refund policy. It was also the version from two quarters earlier. The current policy used a 14-day window for annual plans and a different rule for prorated refunds after the first month.

The model had done something subtle. It found a policy with the right words, quoted a passage that looked relevant, and did not notice that the page carried a newer effective date. The draft was internally consistent and externally wrong.

The team added three fields to the citation record:

1. Source version. A document ID or revision, not only a title.
2. Effective date at the time of review. The date the policy became active, not the date it was downloaded.
3. Supporting passage. The exact lines the reviewer read, with enough context to catch a missing exception.

The revised draft did not say “30 days.” It said the customer should check the current policy, gave the 14-day rule, and routed edge cases to a human. The model still helped write the explanation. It no longer supplied the legal conclusion.

Counterexample: the feature that never shipped

A second draft described a dashboard export: “You can export the full report to CSV and schedule a weekly email.”

The CSV export existed. The scheduled email did not. It was on a roadmap slide, and the model had blended the two into one capability.

No policy document contained the false claim, so a citation-only check would have found nothing to compare. The team added a capability source: the current product documentation, release notes, or a named product owner. If a capability has no shipped source, the claim is blocked, even if the model can describe it confidently.

The reviewer’s note was short: “CSV yes; scheduled email no; remove or mark as not available.” The draft changed before publication. The model’s explanation was useful, but the product owner was the authority.

The citation gate

The team turned the review into a table that had to be completed before publish. It was intentionally small enough to use on every article.

Claim	Source of truth	Version / effective date	Owner	Decision	Note
Refund window is 14 days	Refund policy v4.2	Effective 2026-07-01	Policy owner	Pass	Passage 2.1
Scheduled CSV email is available	Product docs	Not shipped	Product owner	Block	Roadmap only
Support replies within one business day	Support SLA	Effective 2026-01-15	Support lead	Pass	Excludes weekends
Price is $29 per month	Pricing page	Region-dependent	Pricing owner	Revise	Add region qualifier

The gate had three outcomes, not two:

• Pass. The claim is supported, current, and within the publisher’s authority.
• Revise. The claim is partly right but needs a condition, a narrower scope, or a date.
• Block. There is no current source, the source is not authorized, or the claim requires a specialist.

A blocked claim did not stop the whole article. The editor removed it, replaced it with a verified statement, or asked the owner for a decision. The important part was that “block” was a normal outcome, not an exception that required escalation every time.

Who owns a stale source

A citation gate fails if every stale source becomes the editor’s problem. The team assigned ownership before publication:

• Policy pages belong to the policy owner, who confirms the effective date and the passage.
• Product capabilities belong to product or support documentation, not to the marketing draft.
• Prices and eligibility belong to the commercial owner, who confirms the region and plan.
• Legal and medical statements belong to a qualified reviewer, not to an editor or a model.

The owner did not have to rewrite the article. They had to confirm one thing: “This passage is current and supports this claim.” That is a smaller, reviewable request than “check the whole draft.”

The team also added a freshness check to the publishing checklist. Any source older than a set interval, or any policy with a known review cycle, was re-opened before the next update. The interval was based on how often the source changed, not on a universal number.

Correction and rollback

Even a good gate will miss something. The team wrote a correction path before it needed one:

1. Locate the claim. Use the citation record to find the article, the passage, and the owner.
2. Decide the severity. A typo gets a normal edit. A wrong eligibility rule or price gets a visible correction note and a support-team notification.
3. Publish the correction. State what changed and when, without rewriting history silently.
4. Update the source record. If the source changed, mark the old version as superseded and check other articles that cite it.
5. Ask why the gate missed it. Was the source stale, the passage ambiguous, or the reviewer unauthorized? Fix the gate, not just the sentence.

The correction note is not a confession. It is evidence that the publication has a review process and can correct itself. Readers and search engines both benefit from a page that is maintained rather than frozen.

Boundary

This gate fits help-center updates, product documentation, policy explainers, and editorial content with checkable claims. It does not replace legal review, medical review, or a product owner’s approval. It is not designed for breaking news, opinion pieces, or creative writing where the standard is not factual support. It also does not fix a bad source: if the policy owner cannot confirm the passage, the article should wait. The model can draft, summarize, and locate candidates; it cannot be the authority that approves the claim.