A citation can look complete and still fail its only job: supporting the claim beside it.
GPTZero investigated four reports published by PwC Middle East between 2024 and 2026 and identified fabricated references, unsupported claims, and inconsistent citation details. The investigator sells detection software, so its conclusions need independent scrutiny. Financial Times coverage reviewed the findings, and PwC said it was updating a limited number of supporting citations. The useful lesson is not about one firm. It is about what “human review” must include when AI assists research and drafting.
A footnote is not evidence
GPTZero classified four citations across the reports as fake under its conservative definition. It also documented many sources that existed but did not support the specific claims attached to them. In one report, a source URL retained a utm_source=chatgpt.com parameter. That parameter does not reconstruct the full production workflow, but it is a strong sign that AI-assisted research entered the citation chain.
The more important failure is simpler. A reference can have a plausible title, recognizable publisher, and working URL while pointing to the wrong fact. Bibliographic correctness and claim support are two separate checks. Both are required.
Human review can still approve fiction
“A human was in the loop” says almost nothing about quality unless the human had a defined review job.
A reviewer who reads for tone, structure, and grammar may never open the sources. A subject-matter expert may spot an implausible conclusion but miss a fabricated journal article. An approver working from a rendered PDF may not have access to the research trail at all. Every person can complete their assigned step while the evidence remains unchecked.
This is a workflow design problem. Evidence review needs its own owner, inputs, and acceptance criteria. It cannot remain an implied part of a final editorial pass.
Make evidence QA a separate deliverable
For every material claim in AI-assisted work:
- Open the source: confirm that the document or page exists and is accessible.
- Match the identity: verify title, author or organization, publication date, and URL or DOI.
- Check the claim: read enough of the original source to confirm that it supports the exact statement, number, and scope.
- Label the evidence: distinguish vendor statements, first-party data, independent reporting, and inference.
- Preserve the trail: retain the source link and, where appropriate, a stable copy or captured excerpt for later review.
- Assign sign-off: name the person responsible for evidence quality, not only prose quality.
- Fail closed: remove, qualify, or replace any claim that cannot be supported before publication.
Automation can help inventory citations, detect dead links, compare metadata, and flag suspicious patterns. It cannot decide that a source truly warrants a business conclusion without a defined standard and accountable judgment.
The citation is not decoration added after the argument. It is part of the deliverable. If the source cannot survive review, neither should the claim.
Related services: AI Strategy, Marketing, Operations