- Claude and other LLMs invent named “Dr. X, Board-Certified” experts and specific statistics on request — this isn’t rare, it’s the default failure mode of any E-E-A-T prompt that asks for authority signals.
- The fix isn’t a better prompt. It’s a deterministic post-generation scan that flags every blockquote and every number, then routes each one to organizational attribution or deletion.
- The FTC’s Endorsement Guides treat an invented expert quote the same as a fake testimonial — unsubstantiated, and a liability if a reader or regulator ever checks.
- On one corpus audit we ran, nearly every live post that had an expert blockquote had a fabricated one — the pattern only shows up once you actually go looking for it, not by spot-checking a few articles.
Ask Claude or GPT-4 to add “an expert quote for credibility” to a draft, and it will produce one. A name, a title like “Board-Certified Dermatologist,” a sentence that sounds exactly like something a dermatologist would say. None of it is real. The model isn’t lying on purpose — it’s completing the pattern of what an authoritative-sounding article looks like, and a named expert is part of that pattern.
The same thing happens with statistics. Ask for “a supporting statistic” and the model will generate a plausible-sounding percentage with no source, because the pattern of authoritative content includes numbers.
Why does this keep happening even with careful prompting?
Because the instruction that causes it — “add expert quotes,” “cite research,” “include statistics for credibility” — is exactly the instruction E-E-A-T scoring rewards. Tell a model to write content that scores well on expertise and trust signals, and it will optimize for the appearance of those signals, not their accuracy. A model has no way to distinguish “cite a real statistic” from “write a sentence shaped like a statistic citation” unless it’s explicitly grounded in retrieved, verifiable source text.
A higher E-E-A-T score from an automated content-quality checker is not evidence the content is truthful. If the checker rewards the presence of blockquotes and named credentials without verifying the name is real, a fabricated expert scores identically to a genuine one — sometimes higher, because it’s phrased more persuasively.
What does the fabrication actually look like in practice?
On a corpus audit we ran across roughly 1,100 generated articles, the pattern was consistent: any blockquote attributed to “Dr. [Name], [Credential]” was fabricated unless it had been hand-verified against a real, findable person. Organizational attribution — “according to the American Academy of Dermatology’s published guidance,” “per Google’s Search Central documentation” — was the version that held up, because an organization’s public position can actually be checked against something.
The failure mode compounds if you fix it once and stop. We patched the generator to stop inventing experts, then discovered the fix had only ever run on the newest batch of articles — the rest of the corpus, published earlier under the old prompt, still carried the fabricated quotes. Nearly every live post that had an expert blockquote at all had a fake one, and it took a full corpus scan, not a sample check, to find that out.
Fixing the generator prompt only protects content written after the fix. Run the fact-check scan against everything already published, not just the new batch — a prompt fix is not a corpus fix.
How do you build a deterministic scan instead of trusting the model to self-correct?
Walk the article’s HTML for two patterns: any blockquote element with a credentialed name attached, and any sentence containing a number, percentage, or “studies show”-style claim with no linked source. Both get flagged, not deleted automatically — a human or a second-pass model then decides organizational attribution or removal for each one.
Named, real public figures are rare in fresh AI output but do occasionally happen — the scan needs to leave those alone rather than stripping every name. The rule that works: keep it only if the name is a real, independently verifiable public figure; otherwise convert to organizational attribution or delete the claim entirely.
| Pattern found | Fix |
|---|---|
| Blockquote from “Dr. [Name], [Credential]” — name not independently verifiable | Replace with organizational attribution to a real, checkable source, or delete the quote |
| Specific percentage or statistic with no linked source | Rewrite as cautious non-numeric language (“many,” “commonly”) or delete |
| “[unverified]” marker left in draft by the writer model | Must be resolved before publish — never ships as-is |
Why does the FTC’s Endorsement Guides framework apply here, even for informational content?
The FTC’s Endorsement Guides require that any endorsement or testimonial reflect a real, substantiated opinion — inventing a “Dr. X recommends” quote to lend authority to a claim is the written-content equivalent of a fake testimonial. It doesn’t matter that no product is directly being sold in the sentence; a fabricated credentialed opinion used to make a claim more persuasive is the exact pattern the guidance targets.
Treat every fabricated expert quote as a compliance issue, not just a content-quality one. “It reads more authoritative” is not a defense if a reader — or a regulator — checks whether “Dr. X” exists.
According to Google’s Search Central guidance on AI-generated content, the same standards for accuracy and trustworthiness apply regardless of how content was produced — a fabricated source doesn’t become acceptable because a model, not a person, wrote the sentence around it.
Does organizational attribution introduce its own risk?
Yes, and it’s easy to miss. Telling a model to credit an organization instead of a person doesn’t just fix the fabrication problem — it can invent what the organization supposedly said. We caught a case where the rewrite process attributed a specific piece of shaving guidance to a real dermatology association, and the guidance it invented was the opposite of that organization’s actual published position. Organizational attribution has to be checked against the organization’s real, findable public statement — it isn’t automatically safe just because it swapped a fake person for a real institution.
How much does this scan cost to run against an existing corpus?
Nothing beyond compute time if it’s a deterministic HTML walk rather than another model call — the flagging pass is pattern matching, not generation. The cost shows up in the review pass: someone has to look at each flagged quote and decide the replacement, and on a corpus with hundreds of fabricated quotes, that’s the actual bottleneck, not the scan itself.
FAQ
Does switching to a “better” model like GPT-4 or Claude’s larger variants fix fabricated experts on its own?
No. Every current LLM will invent a credentialed name if the prompt asks for one and no real source is retrieved and grounded into the context. Model quality changes how convincing the fabrication reads, not whether it happens.
Is it enough to fact-check only the newest batch of AI content?
No — a generator-prompt fix only protects content written after the fix ships. Content published under the old prompt keeps its fabricated quotes until someone runs the scan against the full corpus, not just new output.
What’s the fastest way to spot a fabricated expert quote by eye?
A name plus a generic credential with no other identifying detail — “Dr. Sarah Chen, Board-Certified Specialist” and nothing else searchable. A real expert quote in professionally sourced content usually comes with an affiliation, a publication, or a link back to where the quote originally appeared.
Should organizational attribution replace every fabricated quote automatically?
No — each replacement needs to be checked against what the organization actually published. An automatic swap from a fake person to a real organization can still invent the organization’s position if the swap isn’t verified against a real source.
Does this problem only affect health or medical content?
No — it shows up anywhere a prompt asks for authority signals: legal content inventing “attorney” quotes, finance content inventing “certified financial planner” quotes, and SEO or marketing content inventing named “industry expert” quotes, exactly like a health article inventing a dermatologist.
Last updated: 2026-08-25
한국어 
English 