Research Note · AI, Language, and Knowledge

When AI Criticism Changes the Question

The judgment gap between producing better papers and advancing scientific knowledge

Published: October 15, 2026
Author: Arthur Palmer

A research project can improve with every revision and still lose the question that made it worth pursuing.

I encountered this possibility while developing an argument about large language models and the production of knowledge. Several AI systems reviewed successive drafts. Their criticisms were often useful. They identified unsupported claims, asked for sharper definitions, distinguished theoretical propositions from empirical hypotheses, and helped remove weaknesses that should have been removed. Yet there was a second movement. As each objection was addressed, the research question became narrower, easier to defend, and less consequential. The manuscript was becoming more difficult to criticize, but that did not necessarily mean the inquiry was becoming more important.

This was an experience within one research project, not a controlled experiment about AI reviewers. It nevertheless exposed a problem that deserves systematic study. A correct criticism does not, by itself, determine what a researcher should investigate next. The ability to find weaknesses in a proposed answer is different from the ability to judge whether its underlying question should be abandoned, reformulated, or investigated more deeply.

The distinction matters because AI has sharply reduced the cost of scholarly criticism. We can now generate more arguments, more alternatives, and more objections than most researchers can independently investigate. The resulting bottleneck may no longer be the supply of plausible research. It may be the judgment needed to decide which possibilities deserve scarce attention, and what evidence would establish them.

The judgment accumulated before the question

Science has never been a collection of isolated individuals reasoning from the beginning. Researchers inherit theories, methods, instruments, disputed findings, and standards of evidence. They also inherit judgments about which questions are promising and which approaches have failed. As Michael Polanyi argued in The Republic of Science, the direction of inquiry emerges partly from investigators making decentralized judgments in relation to one another's work.

Not all of this experience appears in published articles. A paper records a result and some reasons for accepting it, but it rarely preserves every abandoned approach, failed experiment, institutional obstacle, or intuition that guided the investigation. The scientific literature is a substantial but selected record of knowledge and judgment.

Language models learn from parts of that record, together with other materials and training signals. They can make established scholarly practice available to someone who lacks access to a specialist colleague. They can also discover useful connections that no individual researcher has assembled. But what a model expresses is not identical to the collective judgment of a living field. It is a learned and context-dependent evaluation formed from selected material.

The important comparison is therefore specific. A model may be a better critic of statistical code than the researcher using it, while that researcher may better understand a local institution or a neglected historical source. A model may identify a missing control variable, yet be less effective at deciding which new observation would distinguish two rival explanations. The ranking can reverse from one decision to the next.

I call the relevant capability research judgment: the ability to assess what an objection establishes, what remains unknown, and which next investigation has the greatest prospective intellectual value. It is not an exclusively human faculty. Models can exercise it, and people can improve it through working with models. Nor is persistence in one's original theory evidence of good judgment. Sometimes the right decision is to stop. The issue is whether the decision rests on an adequate appraisal of the problem, rather than on the convenience of the most readily defensible answer.

A correct criticism can lead to the wrong destination

Consider a political-economy study asking why local governments accumulate debt. One explanation emphasizes fiscal incentives: local officials may receive near-term benefits from investment while later administrations and taxpayers bear much of the cost. Other explanations emphasize genuine public-service needs, credit conditions, or shifts in revenue.

A researcher could begin with accessible fiscal data and find an association between investment and borrowing. An AI reviewer might correctly observe that correlation does not establish the proposed incentive mechanism. It might then recommend a narrower paper documenting the association. That revision could be entirely appropriate if the available evidence supports nothing more.

But it is not the only response. The researcher could instead ask which administrative records, policy changes, or institutional comparisons would separate the incentive explanation from its rivals. This route may be expensive, uncertain, or ultimately impossible. It could also answer the question that motivated the project.

Here three judgments must remain distinct. First, is the criticism valid? The causal inference may indeed be unsupported. Second, what does the current evidence warrant? Perhaps only a descriptive finding. Third, what inquiry is now worth pursuing? That decision concerns possible evidence, research costs, and the value of resolving an uncertainty that has not disappeared merely because the manuscript has narrowed.

A careful descriptive paper can be a genuine contribution. So can the decision not to pursue an infeasible causal question. The mistake is to treat the first two judgments as if they have automatically answered the third.

When more reviewers do not mean more independent judgment

AI makes it easy to convene several apparent referees, ask them to criticize each other's reports, and repeat until the manuscript stabilizes. Multiple evaluations may catch more defects. They may also produce an exaggerated impression of consensus.

The benefit of additional critics depends not just on their accuracy but on whether they notice different mistakes. Models trained on overlapping materials or prompted to apply the same conventions can share blind spots. Even when their diagnoses differ, they may favor similar remedies: qualify the claim, reduce the scope, choose the conventional method, or replace the original question with something that can be checked now.

There is a deeper asymmetry. It is often easier to demonstrate that an ambitious explanation has not yet been established than to design the research that could establish it. A reviewer may be excellent at finding reasons a claim cannot stand and have no practical answer to what should be investigated instead. In a repeated review process, this can create a one-way pressure toward smaller claims. Each step may be locally justified. The cumulative change in research direction may never be explicitly considered.

Human peer review has always presented versions of this problem. What changes with AI is the speed, cost, and apparent abundance of criticism. An author can move through many rounds of evaluation before encountering another source of evidence or another investigator's independent experience. Agreement among those reviewers does not necessarily mean that the original question has been tested and found wanting.

A useful review therefore does more than identify defects. It distinguishes a correction to a claim from a change to the research agenda and asks what observation, proof, comparison, or source would genuinely resolve the objection.

The economics of an easier publication

Individual judgment is only half the story. Scientific institutions also determine what researchers can afford to pursue.

AI can greatly reduce the effort needed to survey an established literature, extend a familiar model, run routine analyses, and prepare a polished manuscript. Yet it may do much less to reduce the cost of acquiring inaccessible records, conducting a long experiment, building a new instrument, or testing an unfamiliar mechanism. When these costs fall at different rates, the relative attractiveness of research projects changes.

Suppose a scholar faces two possible projects. One offers a well-defined extension of established work, with a high probability of publication. The other addresses a consequential question but needs years of uncertain investigation. If AI makes the first project much cheaper while barely changing the second, a publication system rewarding reliable short-term output will make the first increasingly attractive. That choice can be individually rational even when the researcher understands why the second question matters more.

The consequences accumulate. Many researchers can independently produce sound, incremental results while collectively underinvestigating important uncertainties. No paper need be false. The loss lies in the questions that are not pursued and the evidence that is never gathered.

This is not a new dilemma created by AI. Research on scientific careers, novelty, funding horizons, and the division of cognitive labor has long examined the tension between dependable output and uncertain exploration. AI may intensify or alleviate it by changing the costs of different activities. If it makes decisive experiments cheaper, it can redirect effort toward difficult questions rather than away from them. The direction of change is an empirical question, not a predetermined verdict.

When computation expands discovery

There are circumstances in which the relationship looks quite different. In mathematics, a proposed construction may sometimes be tested quickly by computation or formal checking. In experimental science, a model-generated hypothesis may be linked to an informative assay. Where the test is both affordable and adequate to the question, a larger menu of possibilities can support more discovery instead of more unfinished proposals.

The crucial word is adequate. Reproducing a regression coefficient checks a calculation; it does not by itself identify a causal mechanism. Confirming a quotation checks textual accuracy; it does not settle an interpretation. Verifying a theorem establishes the conclusion under specified assumptions; it does not demonstrate that those assumptions describe an actual economy. The value of faster checking depends on what the check is capable of establishing.

A fruitful human–AI partnership can therefore assign authority differently across tasks. Let the model find coding errors and unfamiliar literature. Let competing evaluators expose different assumptions. Use their criticism to design discriminating tests where possible. Keep the choice of the research question open to revision, but require that revision to be an explicit intellectual decision rather than the unnoticed result of repeated editing.

This is not a ceremonial demand to keep a human in the loop. A researcher should defer to a model where its evaluation is better and properly evidenced. In other settings, an independently informed researcher may have the stronger judgment. What matters is the quality of the decision and the match between the evidence obtained and the uncertainty the project set out to resolve.

What would tell us whether the gap is growing?

The account developed here is a hypothesis about how research decisions change, not a finding that AI has already caused a general decline in scientific originality. Testing it requires looking beyond manuscript counts and the smoothness of AI-assisted writing.

One study could begin with the same unresolved question and ask different researchers to select a research plan before and after model criticism. It could compare a single review, several sequential reviews, and criticism paired with a requirement to propose an informative test. The outcomes would include not only defects corrected, but also whether the original question was abandoned, which new evidence was sought, and whether the eventual claim answered the motivating uncertainty. Human and model performance would need to be judged on specific decisions rather than ranked by broad reputation or confidence.

A second study could compare fields or projects where AI has mainly reduced writing and routine-analysis costs with those where it has materially lowered the cost of substantive verification. A third could examine whether longer-horizon, failure-tolerant research support changes which projects investigators choose when using the same models.

These tests would not supply a single mechanical measure of scientific importance. Some discoveries acquire significance only much later. They could, however, establish whether AI-assisted workflows systematically redirect inquiry toward easily certified claims, and under what conditions that pattern reverses.

The questions science continues to ask

It is tempting to judge the impact of AI on scholarship by the papers it helps produce. Papers are visible. Abandoned questions are much harder to count.

The more consequential change may occur earlier, when criticism forces a decision about direction. Does an objection lead to a better investigation, a justified retreat, or an easier manuscript that leaves the essential uncertainty untouched? Does an apparently convergent group of AI reviewers provide independent insight, or repeatedly apply the same familiar standards? Do academic rewards encourage the resolution of important questions or simply accelerate the production of answers that can already be certified?

I use the judgment gap to describe the distance between the research possibilities computational systems can generate and the capacity of researchers and institutions to select, investigate, and substantiate the ones worth pursuing. AI can narrow that gap by making good evaluation and decisive tests more accessible. It can also widen it when criticism and publication become easier much faster than substantive inquiry.

The ultimate question is not whether AI can write a convincing paper, or even whether it can make that paper correct within a narrowed scope. It is whether the growing capacity to produce defensible answers also improves our judgment about which questions deserve to remain open—and our willingness to pursue them.