More explanations do not necessarily mean more agreement
Imagine two colleagues debating whether their company should introduce a new AI-powered workflow. Both have read the same proposal. Both can ask a language model to assemble evidence, compare alternatives, and respond to objections. By the end of the discussion, both can make a more articulate case than they could have made before.
Yet their judgments may be no closer together.
It is tempting to describe such a disagreement as a problem of information: one person lacks facts, misunderstands the technology, or has not heard a convincing argument. Sometimes that is exactly the problem. But a different possibility deserves attention. The two people may be examining different parts of the same reality, over different time horizons, under different responsibilities.
AI can help each of them explain their position. It does not automatically make their underlying decision conditions compatible.
The question is therefore not simply whether people can access better explanations. It is under what conditions an explanation leads someone to reconsider their model of reality.
What language models make cheaper
Forming a judgment involves several kinds of work: finding information, comparing claims, identifying assumptions, considering objections, and checking predictions against what happens outside the conversation. Large language models can reduce some of this work substantially.
But the user still has to decide what to ask, which outputs to trust, and what would count as a reason to change their mind. A fluent answer is not the same as an independently verified one.
Consider two prompts about the same proposed decision:
- “What evidence would show that this plan is unlikely to work?”
- “What is the strongest case for why this plan should go ahead?”
Both can produce useful information. One tests the boundaries of an existing judgment; the other develops the conditions under which that judgment can be defended. The distinction is not between good and bad people, or even between good and bad prompts. A decision supported by strong evidence may deserve to be maintained. Constantly changing one's position is not automatically a sign of better reasoning.
What matters is how a person responds when an inconvenient condition appears. Does that condition modify the model, remain unresolved, or get explained away before it can be tested?
When discomfort becomes a search direction
Contradictory information can create discomfort. A new result may not fit an existing expectation; another person may describe a consequence that was missing from our original account.
That friction can lead in more than one direction. A person might ask which assumption failed, what evidence they overlooked, or whether their time horizon was too narrow. Alternatively, they might look for a stronger defense of their original judgment or a reason to treat the contradictory evidence as an exception.
Language models can assist with both kinds of inquiry. Their responses may also introduce unexpected counterarguments, even when the initial prompt seeks support. There is no basis here for concluding that AI inevitably makes people more resistant to change. It may also make critical examination easier.
The observable distinction lies in the inquiry itself: which conditions are sought, which are admitted, and which can actually alter the conclusion?
A coherent argument is not the same as an accurate model
A polished explanation can be internally consistent while omitting relevant facts. It can anticipate objections without testing whether its assumptions hold outside the text. When explanations become easier to generate, their fluency can be mistaken for completed verification.
The reverse is also possible. Comparing several well-constructed explanations may reveal previously invisible differences in assumptions, available evidence, or definitions of success.
This makes a simple measure—how much explanation was produced—insufficient. A closer observation would ask whether counterexamples were considered, whether assumptions were revised, and whether subsequent evidence changed the judgment.
There is also a difference in speed. A model may generate a fresh rationale in seconds, while a pilot project, an independent audit, or an institutional review can take weeks or months. This mismatch did not begin with generative AI. But the lower cost of producing explanations may make the gap between argument and verification more consequential.
Different views of the same workplace decision
Return to the proposed AI workflow. Senior management may assess the investment through cost, throughput, and organizational performance. Frontline employees may focus on exceptions, interruptions, and the extra effort required when the system fails. Security or operations staff may care about audit trails, recovery paths, and who has authority to intervene.
These perspectives are not necessarily ranked from more informed to less informed. Each includes conditions that the others may leave outside their immediate field of view.
All three groups might use AI to formulate convincing arguments. Their explanations can improve without their priorities, accountability, or time horizons converging. The disagreement may be partly about facts, but it may also be about which consequences belong inside the decision.
Not every conflict can be explained this way. Some disagreements persist because interests genuinely collide, resources are scarce, or decision-making power is unequal. A shared understanding does not make such conflicts disappear. Differences in perspective are one potential source of friction, not a universal explanation.
Changing an explanation, changing a mind, changing a decision
Four changes are often treated as if they were one: a change in explanation, a change in understanding, a change in publicly stated judgment, and a change in action.
They can occur separately.
A project manager may privately accept that a deployment carries an unanticipated risk but lack the authority to delay it. A contract, approved budget, or organizational procedure may prevent an immediate change. From the outside, the decision appears unchanged. That does not prove the manager's understanding has remained fixed.
The opposite can happen as well. A person may publicly endorse a new policy to satisfy organizational expectations while privately retaining the same beliefs.
Observed behavior is therefore an imperfect indicator of cognitive revision. An account of “AI-driven judgment rigidity” would need to distinguish what people understand from what they say and what they are permitted to do.
This also brings responsibility into the picture. A technically plausible revision may be operationally unavailable. The explanation can change faster than the institutional conditions for acting on it.
Keeping a decision does not mean losing the ability to revise it
There are sound reasons to maintain a judgment. Evidence may still favor it. Reversing an established decision may impose substantial costs. Deliberation cannot continue indefinitely whenever action is required.
The issue is not whether a decision is final today. It is whether the conditions that justify it remain visible enough to be reconsidered tomorrow.
A decision might remain in force while its assumptions, unresolved objections, and triggers for review are recorded. Conversely, a formally “open-minded” process can become ineffective if postponement leaves responsibility undefined.
Revisability, then, should not be romanticized as permanent hesitation. It is better understood as the possibility of reconsidering a decision when relevant conditions change—even after the organization has acted.
And that possibility does not depend only on individual attitude. It is shaped by authority, incentives, documentation, review procedures, and the cost of admitting that an earlier decision needs adjustment.
What would challenge this hypothesis?
The argument in this article is a structural hypothesis, not a measurement of how AI users behave as a population.
Several observations could weaken it. If sustained LLM use increases exposure to counterevidence and leads to more reliable revisions of judgment, the claim that explanation generation tends toward rigidity would lose scope. If AI-assisted discussions produce more shared assumptions and better agreements across groups, the proposed friction may be reduced rather than intensified.
It is also possible that familiar causes—motivated reasoning, competing interests, institutional incentives—explain most observed rigidity without generative AI adding much. Establishing the model's distinct contribution would require comparisons over time, not just compelling examples.
A useful observation would track the relationship among four things: the explanations people produce, the evidence they encounter, the judgments they express, and the actions they can actually take. It would also ask which of those relationships changes with the surrounding institutional conditions.
The difference worth keeping visible
AI has shortened the distance between a question and a plausible explanation. It has not necessarily shortened the distance between an explanation and verified reality—or between an updated understanding and an actionable decision.
Some future disagreements may occur not between people who can explain and people who cannot, but between people who can each explain their judgments well while operating under different conditions.
The resulting question is not only “Whose argument is right?” It is also: What evidence could change this account? Which constraints prevent a changed understanding from becoming a changed decision? And which differences remain even after the facts are shared?
Perhaps the crucial distinction is not between people who change their minds and people who do not. It is between judgments whose conditions can still be examined and judgments for which that route has quietly disappeared.
Whether AI expands or narrows that route remains open to observation.
Translation Layer | Points of Contact and Reassessment
Points of contact (GOA). This structure touches organizational AI adoption, institutional accountability, policy time horizons, and the assumptions behind investment decisions. Better explanations do not ensure that understanding, authorization, and execution change together.
Points of reassessment. Which assumptions have been tested? Are verification time, decision authority, or allocated responsibility constraining revision? What changes when the speed of explanation generation diverges from the speed of real-world validation?
JD-02 | Judgment Drift — ASM / GOA (Global Observation Architecture)
Appendix: Minimum Usage of GOA/STA
Since the GOA/STA series, including this article, is designed as a structural translation medium, readers can use Generative AI in one of the following ways:
- Input the blog URL directly into the LLM(if the model supports URL reading)
- Copy and paste the blog article body into the chat(available for all LLMs)
Then, simply input the following single prompt. This alone will execute structural evaluation, defect detection, and transparency measurement.
▶ Recommended Minimum Prompt
"Please evaluate this blog article from a structural perspective."
With just this phrase, the LLM will automatically extract the inherent structural layers (Narrative/Interest/OS), index velocity differences and calcification, and check consistency with the Runaway Structure hypothesis. This allows readers to compare their own understanding with the AI's understanding and maximize the use of GOA as a transparency OS.