When Better AI Output Hides Gaps in Judgment
AI can improve the quality of finished work while leaving the reasoning behind it untested.
An AI-assisted report can be accurate and well written even when its author has a limited understanding of the problem. A recommendation can sound persuasive before anyone checks the assumptions behind it.
That creates a gap between the quality of the work and the judgment exercised in producing it. The gap becomes visible when someone challenges the conclusion, circumstances change, or the decision produces an unexpected result.
What the Research Shows
A 2025 study presented at CHI surveyed 319 knowledge workers about their use of generative AI. Higher confidence in AI was associated with less critical thinking. Higher confidence in one's own ability on the task was associated with more. Participants also described critical thinking shifting toward verifying information and integrating AI responses into their work.
Reducing effort on a task can be useful. Someone who spends less time preparing a report may have more time to investigate an assumption or compare alternatives. The risk arises when accepting an answer replaces those activities.
Output quality and the practice of judgment therefore need separate assessment. A better document shows an improvement in the finished work. It does not establish that its author understands the problem better.
A Recommendation That Outruns the Evidence
Consider a hypothetical business decision.
An online retailer sees a decline in completed purchases. A manager gives AI a summary of website visits, product views, and sales, along with information about a competitor's recent discounts. The AI produces a report attributing the decline to price pressure and recommending a temporary discount.
The figures are accurate. The explanation sounds plausible. The manager approves the recommendation without checking where customers abandon their purchases.
A later investigation finds a payment error affecting mobile customers after a website update. Customers still want the products, but some cannot complete their orders. The discount cuts revenue on successful purchases and leaves the payment problem unresolved.
The AI recommendation exceeded what the supplied information could establish. As a result, a plausible interpretation was accepted as a conclusion.A decline in purchases, combined with a competitor's discounts, supported price pressure as a possibility. It did not establish price as the cause. The manager needed to separate those two claims before acting. Instead, the report appeared complete, so the diagnostic work was skipped.
Scrutiny Requires More Than Reading Carefully
Reviewing an AI response for clarity and consistency can leave its central assumptions untouched. A report may contain no obvious contradiction and still recommend the wrong action.
Analysts use a technique called red-teaming: deliberately challenging your own conclusion by asking what the situation would look like if the conclusion were wrong. In the retailer case, that means asking what the data would show if price were not the cause. Useful checks would include whether the decline occurred across devices, at which stage customers abandoned their purchases, and what changed on the website recently.
Those checks could distinguish price sensitivity from a technical failure. They would also test whether the manager supplied enough information for a recommendation at all.
Asking AI to critique its own report can surface questions worth investigating, but the critique still requires verification. Agreement between AI responses is not independent confirmation. Consequential claims need support from relevant data, original sources, or qualified expertise beyond the conversation.
The Stoic Discipline of Examining an Impression
Epictetus opens the Enchiridion by separating what depends on a person from what does not. In the same section, he tells the reader to examine a troubling appearance instead of accepting it as it presents itself. The Stoics called these appearances impressions. The practice is to examine an impression before assenting to it, meaning before accepting it as true.
An AI response is an impression. Its fluency can make acceptance feel reasonable before the evidence warrants it.
The Stoic practice is to preserve the interval between encountering a claim and assenting to it. In that interval you can question the claim's basis and decide how much confidence it deserves.
Outside advice remains useful. Exercising judgment includes deciding which advice to trust and whether it applies to your circumstances. Responsibility for that choice stays with the person acting on it.
Keep Judgment in the Process
Asking AI for analysis does not guarantee critical engagement. A user can accept an analysis without examination, or scrutinize a proposed solution and reach a sound decision. The difference lies in how the response is evaluated.
Before acting, the decision-maker should understand the recommendation's key assumptions and the consequences of getting it wrong. The reasoning should remain explainable without reopening the AI conversation.
For consequential decisions, a short written process helps. It follows the Core Decision Loop from The Stoic Strategist Decision Toolkit: Detection, Constraint, Execution, Review. Detection identifies distortion early. Constraint tests your options against your principles, priorities, and what you control, then selects an action. Execution carries it out. Review turns the outcome into refinement.
Form a provisional assessment (Detection). Before consulting AI, write a few sentences describing the problem and the available evidence. Include an alternative explanation and the evidence that would change your view. Keep the initial view open to revision.
Separate evidence from assumptions. Use the known-versus-unknown exercise from The Stoic Strategist. List verified facts, inferences, and gaps. In the retailer scenario, declining purchases were a fact. Price pressure was a hypothesis.
Use AI to challenge and extend the analysis, then verify. Compare explanations and investigate disagreements. Red-team the leading explanation. Verify the claims that matter against evidence outside the conversation.
Record the decision and review the outcome (Review). Explain the choice and the uncertainty you are accepting. After acting, examine which assumptions held and which need revision.
The decision itself falls between steps 3 and 4. In the Core Decision Loop, that is Constraint: testing the options and selecting an action.
Scale the effort to the consequences. A routine wording change needs less scrutiny than a decision involving a substantial commitment or a difficult reversal.
Assess Understanding Alongside Results
A successful outcome can conceal weak reasoning. Someone can follow a recommendation for the wrong reasons and still benefit from it. An unfavorable outcome can also follow a reasonable decision made under uncertainty.
Review therefore requires more than checking whether the result was good. A useful review examines the evidence available at the time, the alternatives considered, and the justification for choosing one course over another. It also asks whether the decision-maker recognized the limits of the assessment.
Difficulty explaining those elements does not prove that AI has weakened anyone's abilities. It identifies a gap in understanding that deserves attention.
AI can support judgment through questioning, comparison, and investigation. Preserving that practice requires deliberate participation from the user. A finished report should come with enough understanding to challenge its conclusions, respond to new evidence, and take responsibility for the decision.
The Stoic Strategist Decision System puts the Core Decision Loop into practice for recurring decisions.
Final Reflection
If someone challenged your last AI-assisted decision tomorrow, could you explain its basis without reopening the conversation?
Sources
Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers." CHI 2025. Microsoft Research and Carnegie Mellon University.
Epictetus. Enchiridion, section 1, translated by Elizabeth Carter.
The Stoic Strategist Decision System applies this discipline through a cycle of detection, constraint, execution, and review. The process gives decision-makers a structure for examining assumptions before acting and revisiting their reasoning afterward.
