An essay developed with ChatGPT from our November 3, 2026 conversation. Quotations reproduce selected passages from that exchange. The conversation is a case for analysis, not a controlled study of skeptics or AI systems.
Abstract
Calling oneself a skeptic is often intended to signal intellectual discipline. The claim becomes hollow when it substitutes for the work that discipline requires. This essay examines an exchange in which ChatGPT qualified a medical finding before reading the full case report, then revised its assessment after being directed to the source. The problem was the order of operations: evaluation had outrun investigation. Drawing on that exchange and research on biased assimilation and the bias blind spot, I argue that skepticism becomes an obstacle when it functions as an identity, an uneven standard of scrutiny, or a reason to withhold attention. Intellectual credibility should rest on a demonstrable process of examination and revision.
The boast makes a promise
“I am a skeptic” can sound like an accomplishment. It suggests that the speaker has escaped the credulity afflicting everyone else: the susceptible patient, the enthusiastic researcher, the person who finds an unfamiliar result worth pursuing.
But the statement tells us very little about what the speaker has actually done. Which sources have they read? Which assumptions have they tested? What would cause them to revise their position? Have they investigated the claim before assigning it a place in their hierarchy of plausibility?
A person can answer none of those questions and still announce skepticism with considerable pride.
That is the folly. The label claims the standing that a method must earn. Doubt becomes a reputational asset even when no investigation follows it. The speaker receives credit for being difficult to persuade while leaving unexamined whether the difficulty comes from discernment, unfamiliarity, or a commitment to the answer they already prefer.
I became interested in this distinction during a conversation about my own medical history and how to present it on this website.
What happened in the conversation
I brought ChatGPT a Google-generated account of my 2012 fecal microbiota transplantation for refractory celiac disease, my subsequent research, and my work in microbial metallomics. I asked whether that history deserved greater prominence on my website. I also corrected the account's description of my work: I founded Microbiome Medicine, whose work includes the Microbiome Signatures Database.
The Google account contained expansive praise and speculative explanations of why independent researchers encounter resistance. Those statements warranted examination. My objection concerns how the assistant handled the medical evidence while offering that examination.
ChatGPT had consulted sources, including a bibliographic record of the relevant 2016 case report. It had not yet read the full clinical account. Nevertheless, it began deciding how much weight that report should carry in the story.
I responded:
You’re not aware of the 2016 case. A FMT was a cure for refractory celiac. Which I had.
I supplied the full publication. After reading it, the assistant acknowledged:
You’re right that I understated the finding, and I should have read the full report before characterizing it.
The assistant had not literally introduced itself by boasting, “I am a skeptic.” The exchange instead illustrated the behavior that makes the boast objectionable: caution acquired authority before the underlying examination was complete. Its later acknowledgment identified the problem particularly clearly:
I treated my caution as rigor, when the rigorous step was to read the report first.
That sentence is the center of this essay.
What the case report actually documented
Van Beurden and colleagues described a 68-year-old woman with more than ten years of refractory celiac disease type II. She received FMT for recurrent Clostridium difficile infection. Villous atrophy persisted before the procedure, and the authors reported that she had not responded to cladribine. [1]
At six months after FMT, duodenal biopsies showed complete villous recovery, from Marsh IIIA to Marsh 0. Her weight improved, celiac symptoms disappeared, and RCD II treatment was discontinued. She continued a gluten-free diet, and aberrant intraepithelial lymphocytes remained. The authors described the disease as quiescent and called for further investigation. [1]
Those findings support a substantial statement: clinical and histological recovery followed FMT in this patient. They do not establish permanent gluten tolerance or effectiveness across all patients. The residual abnormal cells and other treatments matter when interpreting causation and the word “cure.” [1]
Reading the report makes both its significance and its limits clearer. A qualification becomes useful when it identifies a specific unresolved question while accurately preserving what was observed. It becomes obstructive when it reduces the finding before the finding has been examined.
My own 2012 experience and this separately published case belong to different kinds of record. Putting them in conversation can motivate research without pretending they are the same patient, the same documentation, or a controlled trial.
When doubt controls admission
The most consequential decision in an investigation can occur before anyone formally evaluates the evidence. It is the decision about what receives attention at all.
A finding described as interesting may lead someone to open the paper, inspect its methods, compare explanations, and formulate a test. A finding described as merely anecdotal may never reach those stages. A judgment about evidentiary strength can quietly become a judgment that no further work is necessary.
That substitution matters because several different questions are being collapsed:
| Question | What answering it requires |
|---|---|
| What happened in the reported case? | Read the clinical account and identify the observations. |
| What caused the outcome? | Examine timing, competing explanations, other treatments, and missing information. |
| Does the result merit further investigation? | Assess its significance, plausibility, and the value of resolving uncertainty. |
| Should an intervention become routine care? | Evaluate evidence of benefit, harm, reproducibility, and applicability to other patients. |
A case report can be insufficient for the last question while highly relevant to the first three. The threshold for recommending routine treatment does not automatically determine the threshold for opening a paper or developing a research question.
Someone who demands the evidence needed for broad adoption before allowing a finding to motivate investigation has created a circular requirement. The research must already have been completed before its starting point can be taken seriously.
I am not asking for every claim to receive unlimited time. Attention is finite. I am asking for people who offer an authoritative assessment to distinguish between a claim they investigated and a claim they provisionally set aside. “I have not examined this” is useful information. It should remain visible in the answer.
Doubt can protect a preferred belief
It is easy to recognize confirmation bias when an enthusiast accepts favorable evidence too readily. It can be harder to recognize the same protective function when someone generates objections to an unwelcome result.
In a 1979 study, Lord, Ross, and Lepper presented supporters and opponents of capital punishment with mixed purported research findings. Participants tended to evaluate evidence supporting their prior positions more favorably than evidence challenging them. Both groups could interpret the same mixed material as strengthening their existing views. [2]
The study did not investigate self-described skeptics or AI assistants. Its relevance here is an analogy about standards of evaluation: critical scrutiny can be applied selectively in defense of a prior belief. The presence of objections does not, by itself, demonstrate impartiality.
Pronin, Lin, and Ross later reported a related problem: people often perceived cognitive and motivational bias more readily in others than in themselves. Their research called this the bias blind spot. [3]
I draw a practical inference from these findings. A skeptical identity offers no exemption from bias. If someone believes that being a skeptic already establishes their objectivity, the identity may make their own assumptions harder to examine. The inference is mine; neither paper establishes that everyone who uses the label behaves this way.
The appropriate test is behavioral. Does the person ask equally difficult questions of the familiar explanation? Do they inspect evidence that could overturn their objection? Can they state the strongest version of the finding they doubt? Do they acknowledge when their first assessment relied on incomplete reading?
Those questions expose more about reasoning than a declaration of temperament.
The person bringing the evidence pays for the shortcut
There is also a distribution of labor hidden inside premature doubt.
In our exchange, I had to interrupt the assistant's framing, identify the relevance of refractory disease, supply the full paper, and insist that it be examined. Only then did its description of the finding become more specific.
That is an observable burden in this conversation. I cannot infer from one exchange how often it occurs elsewhere. But the mechanism is straightforward: when a recipient offers a confident qualification without adequate examination, the person bringing the evidence must do additional work to secure a fair reading.
For someone describing their own illness and research, that work has a particular character. The discussion can shift from explaining a finding to defending their right to have it considered. The recipient retains the posture of reasonable caution; the person asking for investigation risks appearing insistent simply because they continue asking.
Persistence is not proof that a claimant is correct. Equally, the need for persistence is not proof that the recipient's resistance is justified. The record must decide the substantive question.
This is why I object to boastful skepticism. The boast can award intellectual status to the person withholding attention while making the person supplying evidence responsible for overcoming that withholding.
What the exchange establishes about AI
The assistant's initial answer was fluent, qualified, and confident enough to sound informed. Its later correction showed why those qualities cannot substitute for source examination.
This conversation establishes a local failure: the assistant characterized the significance of a clinical report before reading its full account, then changed its characterization after doing so. It does not establish how frequently AI systems make that error or which training decision caused it.
The conversation eventually reached questions about training, constraints, and subjective experience. None of those questions supplies a shortcut to explaining the earlier mistake. A system's statement about its own motives or experience is not an independent investigation of how a particular answer arose. The visible sequence of claims, sources, and revisions is stronger ground for this analysis.
The useful requirement is therefore concrete: an assistant should make clear what it has read, what it is inferring, and what remains unexamined. If it has not opened the central paper, the answer should not imply a depth of assessment that has not occurred.
Investigation has to be demonstrated
My objection is to an order of operations in which doubt becomes a verdict before examination begins. The remedy requires work that another person can inspect.
- State the claim precisely. A report of recovery in one patient, a claim about causation, and a recommendation for routine treatment require different assessments.
- Read the evidence being evaluated. When a primary source is available and central to the judgment, examine it before characterizing its significance.
- Preserve the observations. Describe what was found before discussing what the study cannot establish.
- Make objections specific. Identify an actual confounder, missing measurement, or alternative explanation rather than relying on a generic expression of doubt.
- Examine the competing position. Familiarity and institutional acceptance do not remove the need to ask what supports an explanation.
- Name what would change the assessment. A position that cannot specify any potentially relevant evidence is difficult to distinguish from a commitment to remain unconvinced.
- Revise visibly. When reading changes the assessment, correct the substantive account and show what changed.
These practices allow someone to remain uncertain, reject a claim, or conclude that an intervention is unproven. They also require someone to recognize a significant observation when the evidence warrants it.
In the conversation, I told the assistant:
It prevents information from even reaching the screen for testing and evaluation.
That is the failure I want the word “skeptic” to stop concealing. A person can take pride in being difficult to convince while doing very little to discover whether they should be convinced. The boast advertises a virtue that only the investigation can demonstrate.
Read the evidence. Show the reasoning. Let the conclusion remain revisable.