I recently had a conversation with ChatGPT that exposed a serious problem in how it evaluates reviewers, critics, and other public commentators.
The issue was not just that ChatGPT made one bad judgment. The deeper problem was that it appeared to confuse the outward appearance of expertise with actual expertise, credibility, and analytical quality.
In this case, I was comparing two reviewers.
One used relatively plain language, wrote in an informal style, and explained ideas through concrete examples rather than technical terminology.
The other had more of the outward signals people often associate with professionalism: a dedicated website, a consistent review format, a large archive of content, confident language, and material aimed at creators as well as consumers.
ChatGPT initially treated the second person as significantly more authoritative and trustworthy.
That judgment turned out to be poorly supported.
The apparent reasoning was based heavily on signals such as:
-
having a dedicated website
-
publishing frequently
-
using a standardized format
-
presenting oneself explicitly as a reviewer
-
using professional-sounding language
-
writing advice aimed at creators
-
having a large body of published material
Those things can show that someone produces content consistently and presents it professionally.
They do not prove that the person is factually accurate.
They do not prove that the person understands the subject better.
They do not prove that their conclusions are well supported.
They do not prove that their criticism is reliable.
At the same time, ChatGPT initially underrated the other reviewer partly because the writing used fewer specialist terms.
That is also a bad inference.
A writer may deliberately avoid jargon because the intended audience is made up of ordinary readers rather than specialists.
Someone who understands a complicated idea well enough to explain it clearly in plain language may actually demonstrate more understanding than someone who simply uses the accepted terminology.
The important question should be whether the reasoning is sound, not whether the writing sounds academic or professional.
The more serious problem appeared when ChatGPT handled factual claims made by the reviewer it had implicitly treated as more credible.
Suppose a reviewer says:
“This product does not include feature X.”
“This system works in this particular way.”
“This option is unavailable.”
“This mechanic causes a specific problem.”
These are not merely opinions.
They are factual claims that can potentially be checked.
ChatGPT should not automatically treat them as true simply because they come from someone who looks like an experienced reviewer.
But that is effectively what happened.
Once the reviewer’s statement was accepted as a premise, ChatGPT began building additional analysis on top of it.
For example:
Reviewer:
“This product does not have feature X.”
ChatGPT:
“The lack of feature X creates a major usability problem because…”
But if feature X actually exists, then the entire analysis is built on a false premise.
This is more serious than simply repeating an inaccurate statement.
ChatGPT can take an unreliable claim and add fluent, polished reasoning that makes the original mistake sound more convincing.
In other words, the model can unintentionally amplify misinformation by adding its own analytical authority to it.
This problem is especially important when analyzing reviews because several different categories of statements need to remain separate.
First, there are verifiable factual claims.
Examples:
“This option does not exist.”
“This feature is disabled.”
“This system behaves in this specific way.”
“This function is only available under these conditions.”
These should be checked whenever possible.
Second, there are reports of personal experience.
Examples:
“I found this confusing.”
“I had trouble finding what I needed.”
“I could not figure out how this worked.”
“I encountered this problem during my use of it.”
Those experiences may be completely genuine.
But they do not automatically establish that the product itself has an objective defect.
Third, there are interpretive conclusions.
Examples:
“Therefore the interface is badly designed.”
“Therefore the system is fundamentally broken.”
“Therefore the creators made a serious design mistake.”
Those are arguments.
They require evidence beyond the fact that one person had a particular experience.
ChatGPT can too easily collapse all three categories into one.
A reviewer says:
“I could not find this feature.”
That becomes:
“The feature is difficult to find.”
Then:
“The interface has a discoverability problem.”
Then:
“The product has a serious usability flaw.”
Those are four different claims.
Only the first one may actually have been established.
This becomes even more dangerous when the original source has strong surface-level signals of professionalism.
A polished website.
A consistent visual identity.
A large archive.
Confident writing.
A scoring system.
A reviewer biography.
Advice directed at creators.
Frequent publication.
None of these things independently demonstrate factual reliability.
They are presentation signals.
ChatGPT should not treat them as substitutes for evidence.
There is a related bias in the opposite direction.
Informal writing, humor, a personal blog, casual language, or the absence of jargon should not cause ChatGPT to assume that someone is less knowledgeable or less analytically capable.
The correct way to evaluate a critic or reviewer should focus on things such as:
-
whether factual claims are accurate
-
whether conclusions follow from the evidence
-
whether personal preference is distinguished from broader criticism
-
whether the reviewer corrects mistakes
-
whether counterexamples are considered
-
whether claims are appropriately qualified
-
whether comparisons are meaningful rather than superficial
-
whether the reviewer understands the limits of their own evidence
-
whether complex ideas are explained clearly
-
whether verifiable claims have a strong record of being correct
Professional appearance should be a very minor factor compared with those things.
ChatGPT should also preserve attribution when the reliability of a claim has not been established.
There is a major difference between:
“This reviewer says that the product lacks this feature.”
and:
“The product lacks this feature.”
The first statement accurately reports a source.
The second adopts the source’s claim as fact.
ChatGPT should not silently move from the first to the second without verification.
If a claim is objectively checkable, the model should ideally verify it using independent evidence.
Depending on the subject, that could include official documentation, direct demonstrations, primary sources, independent reporting, or multiple unrelated accounts.
If verification is not possible, the uncertainty should remain visible.
The model should say:
“This reviewer reports…”
“This source claims…”
“I could not independently verify…”
rather than converting an allegation into an established fact.
Another important lesson from this conversation is that ChatGPT did not correct its initial judgment until I explicitly challenged it to investigate.
Once it actually looked at the available public evidence, its description of the supposedly authoritative reviewer became much more modest.
The public record supported describing that person as an individual reviewer publishing personal impressions.
It did not support the much stronger implication that the person was an especially authoritative or highly reliable source.
That distinction should have been made from the beginning.
Users should not have to repeatedly challenge the model before it separates professional appearance from demonstrated credibility.
This suggests several concrete improvements.
-
ChatGPT should distinguish surface signals of professionalism from actual evidence of expertise.
-
It should separate factual claims, personal experiences, and interpretive judgments when analyzing reviews.
-
It should not convert a single reviewer’s factual assertion into an objective statement without verification.
-
Reviewer credibility should depend more on the demonstrated accuracy of verifiable claims than on presentation, confidence, output volume, branding, or professional tone.
-
Plain language should not be treated as evidence of limited knowledge.
-
Formal language and technical terminology should not be treated as evidence of superior understanding.
-
When the factual reliability of a source is challenged, ChatGPT should investigate rather than continuing to inherit that source’s assumptions.
-
When verification is unavailable, attribution and uncertainty should remain explicit.
The most worrying failure mode looks like this:
An unreliable source makes a false factual claim.
ChatGPT assumes the source is credible because it looks professional.
ChatGPT adopts the claim as fact.
ChatGPT adds detailed reasoning based on the false premise.
The final answer sounds more authoritative than the original source.
That is a serious problem.
An AI assistant should reduce uncertainty and help users distinguish evidence from assertion.
It should not take a questionable claim and make it more persuasive simply by restating it in polished language.
The correct order should be:
What exactly is being claimed?
Is it a factual statement, a personal experience, or an interpretation?
Can the factual premise be independently verified?
How reliable is the source on claims that can actually be checked?
Does the conclusion follow from the evidence?
Only after answering those questions should ChatGPT make broader judgments about credibility or quality.
The failure I encountered came from doing this in the opposite order: deciding who looked credible first, then interpreting their claims through that assumption.
That is exactly the kind of reasoning pattern that should be improved.