A reviewer performs a final check on a promotional material draft generated by AI. The output is polished, the structure is logical, and the sense of unease is faint. The certainty felt at that moment could be confidence built on thorough verification, or overconfidence born from skipped verification. How can you tell the difference from the inside? The difference is not in the accuracy of the judgment but in the process: whether counter-evidence was actively sought.
01The more polished the AI output, the less anxiety you feel
AI-generated drafts are formally well-constructed. Grammar errors are rare. The structure appears logical. Headings align with body text. In terms of surface-level completeness, they often exceed rough human drafts.
That polish lowers the reviewer's guard. A rough draft triggers the assumption that "something must be wrong here." A polished document invites the assumption that "this is probably fine."
Here lies the trap. As document quality improves, verification quality can decline in inverse proportion. And the reviewer does not notice the decline. Whether "nothing was found" or "nothing was looked for" cannot be distinguished from the outcome alone.
02Confidence and overconfidence feel identical from the inside
Daniel Kahneman demonstrated that the subjective certainty of intuitive judgments often has no correlation with their actual accuracy. When people feel confident in a judgment, they cannot determine from the inside whether that confidence is grounded in evidence.
Philip Tetlock's prediction research yielded the same finding. There was no difference in subjective certainty between highly accurate and highly inaccurate forecasters. The feeling of "I'm right" is not an indicator of being right.
Applied to material review: the certainty that "this material is fine" could be confidence earned through thorough verification or overconfidence resulting from skipped verification. The internal sensation alone cannot tell you which.
03The difference appears in the process
What separates confidence from overconfidence? Not the outcome. Outcomes are only known after the fact.
The separator is process. Specifically, whether the reviewer took the action of searching for counter-evidence. Before concluding "there are no problems," did the reviewer ask "if there were a problem, where would it be?"
Karl Popper identified falsifiability as the criterion for scientific propositions. A hypothesis can only be validated by attempting to falsify it. Review judgments share this structure. The conclusion "this is fine" deserves to be called confidence only when it survives an attempt at falsification. Without that attempt, the same subjective feeling is more accurately called overconfidence.
Structure of confidence
Counter-evidence was sought and not found, or found issues were resolved. The certainty that remains afterward has a basis.
Structure of overconfidence
Counter-evidence was not sought, or only surface-level checks were performed. The certainty arises with the same intensity. It has no basis.
04As AI accuracy improves, the conditions for overconfidence accumulate
When AI output quality improves, two things happen simultaneously.
First, actual errors decrease. Better AI means fewer obvious mistakes. This is a good thing.
Second, errors become invisible. As obvious mistakes disappear, reviewers predict that "looking won't find anything" and lose the motivation to look. This is a problem, because AI errors never reach zero, and the remaining errors are precisely the kind that are hard to spot from the surface.
| Dimension | Low AI accuracy | High AI accuracy |
|---|---|---|
| Obvious error count | High | Low |
| Reviewer alertness | High | Tends to decline |
| Nature of remaining errors | Surface-level, easy to spot | Structural, hard to detect |
| Overconfidence risk | Low (errors are visible) | High (errors are invisible) |
05Three triggers for overconfidence in review
Three situations in material review practice where overconfidence is most likely to occur.
First, time pressure. When deadlines approach, the capacity to search for counter-evidence shrinks, and reviewers shift to reading under the assumption that "there should be no problems."
Second, repetition. After reviewing dozens of the same type of material, past experience dominates current judgment. "It has always been fine" is not evidence that it is fine this time.
Third, familiarity with AI. After repeatedly checking AI-generated drafts and finding no issues, the granularity of verification coarsens. This is the scenario where AI accuracy improvement and reviewer alertness decline advance in parallel, creating the maximum conditions for overconfidence.
06Embed counter-evidence seeking into behavior
Distinguishing confidence from overconfidence through internal sensation alone is difficult. Therefore, rather than relying on sensation, embed counter-evidence seeking into behavioral routines.
Gary Klein's "premortem" is a technique that asks, before a decision is finalized, "if this decision turned out to be wrong, what would the cause be?" Instead of analyzing failure after the fact, it imagines failure in advance.
Applied to material review, this means posing one question before final approval: "If there is a problem with this material, where would it be?" If the question is asked and no problem is found, the resulting judgment is confidence earned through an attempt at falsification. If the question is not asked, the judgment may happen to be correct, but as a process, it is overconfidence.
As Amos Tversky demonstrated, humans tend to collect information that confirms what they already believe and ignore information that contradicts it. Confirmation bias cannot be overcome by willpower alone, but it can be mitigated by procedure. Making counter-evidence seeking part of the routine removes dependence on the strength of individual will.
The premortem question
Before final approval, ask: "If there is a problem with this material, where would it be?" This single question triggers a falsification attempt.
Verification records
Write down what you checked and what you did not. The act of recording forces conscious awareness of the verification process.
Accumulated evidence
Over time, records of past judgments build a foundation for genuine confidence. Repeated recording improves judgment accuracy itself.
07Records turn invisible processes visible
There is one more practical method: recording the process of judgment.
When you judge "no problems," write briefly what you checked and what you did not check. The act of writing forces conscious awareness of the verification process. If you try to write and cannot recall what you checked, that is a signal that verification was insufficient.
One trait shared by the people Philip Tetlock called "superforecasters" was that they recorded the basis for their predictions and reviewed them afterward. Recording not only prevents overconfidence but also accumulates the foundation for genuine confidence. In work that involves repeated judgments of the same kind, the accumulation of records eventually improves the accuracy of judgment itself.
There is no need to fear overconfidence. What deserves fear is making repeated judgments without realizing you are overconfident. The most reliable means of awareness is embedding counter-evidence seeking and recording into the judgment process itself.
- Confidence and overconfidence are indistinguishable from the inside. Subjective certainty is not an indicator of accuracy. The difference lies in whether counter-evidence was actively sought: a process distinction, not an outcome distinction.
- As AI output accuracy improves, reviewer alertness tends to decline and overconfidence conditions accumulate. The errors that remain are precisely the kind that are hard to detect from the surface.
- Rather than relying on sensation, prevent overconfidence through behavior: ask "if there is a problem, where would it be?" and record the verification process as part of the routine.
The feeling of "this is fine" is not evidence of confidence. It is merely a report that certainty has occurred.
I return to this question myself. The sense of "this is good" when I finish writing a piece is usually just the point where I can no longer see what needs to be changed. Whether I can no longer see it or have stopped looking: the difference only becomes clear when I read it again the next morning.
Searching for counter-evidence is tedious. But only after accepting that tedium does the word "confidence" become earned.
- Daniel Kahneman. Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011. (The disconnect between subjective certainty and accuracy in intuitive judgment. System 1 and System 2 distinction.)
- Philip Tetlock. Superforecasting: The Art and Science of Prediction. Crown, 2015. (The absence of correlation between prediction accuracy and subjective certainty. The recording habits of superforecasters.)
- Karl Popper. The Logic of Scientific Discovery. Routledge, 1959. (The principle of falsifiability. A hypothesis can only be judged through attempted falsification.)
- Gary Klein. The Power of Intuition. Currency, 2004. (The premortem technique. Imagining failure in advance to improve decision quality.)
- Amos Tversky, Daniel Kahneman. Judgment under Uncertainty: Heuristics and Biases. Science, 185(4157), 1974. (Systematic study of cognitive biases including confirmation bias.)
- Justin Kruger, David Dunning. Unskilled and Unaware of It. Journal of Personality and Social Psychology, 77(6), 1999. (Low-ability individuals tend to have higher self-assessments. The psychological basis of overconfidence.)
- Don A. Moore, Paul J. Healy. The Trouble with Overconfidence. Psychological Review, 115(2), 2008. (Three forms of overconfidence: calibration, precision, and placement.)