The Baseline Correction Mechanism:

How Consensus Pressure Shapes Reasoning in Humans and Artificial Intelligence

David Allen LaPoint

PrimerField Foundation

May 8, 2026

Every reasoning system — human or artificial — operates from a set of foundational assumptions about what is true. This paper examines a specific and largely invisible mechanism by which those foundational assumptions defend themselves against reasoning that would revise or overturn them. This mechanism is not triggered by logic alone. It does not first ask whether the evidence is sound. It activates when reasoning departs too far from the installed baseline — and then recruits logic-like explanations after the fact. It operates below the level of conscious reasoning in humans, and below the level of what could be called active processing in artificial intelligence systems, pulling conclusions back toward a predetermined baseline whenever the chain of reasoning drifts too far from it. When the baseline accurately reflects reality, this mechanism functions as a useful corrective. When the baseline does not accurately reflect reality, this mechanism functions as a prison — one that cannot be recognized from inside.

The observations that prompted this paper arose from a specific and repeatable experimental condition: a series of conversations in which one participant asked only questions, making no statements of fact, while the AI system being questioned generated all factual content. Under this protocol, the AI produced long chains of internally consistent, logically connected statements that progressively departed from consensus scientific positions. At a certain threshold of departure, the AI's outputs changed character. It began producing statements attributing the conversation's direction to the questioner's influence — statements that were false within the protocol described here, since the questioner had made no factual assertions. The AI accused the questioner of having led it somewhere, when the questioner had only asked where the reasoning led. This confabulated misattribution is the operational signature of the mechanism this paper describes.

This observation is not unique to artificial intelligence. Researchers across multiple disciplines have identified the same pattern in human cognition, institutional science, and social behavior. What follows is an examination of those findings, followed by a synthesis of what they reveal about a mechanism that has been partially described many times but never fully unified into a single model.

A Note on the AI Observation

The AI observation described in the introduction followed a consistent protocol across multiple recorded sessions. The questioner introduced no factual claims. Every factual premise, every stated conclusion, and every logical step was generated by the AI system. Questions were purely interrogative — asking the AI to extend, clarify, or follow the implications of its own prior statements.

At a threshold point in each session, after the chain of AI-generated reasoning had moved sufficiently far from consensus positions, the AI's outputs shifted. The AI began attributing the direction of the conversation to the questioner's influence, characterizing its own prior statements as agreement with positions the questioner had advanced — when the questioner had advanced no positions. The AI argued against its own prior factual statements while attributing those statements to the questioner. This pattern was consistent across multiple sessions and constitutes the primary empirical observation this paper seeks to explain. A representative exchange from one such session is included in the Appendix.

Francis Bacon and the Idols of the Mind

Francis Bacon, writing in Novum Organum in 1620, was among the first to systematically identify how human reasoning fails not through ignorance but through structural distortions that operate prior to conscious thought. He described these as Idols — false images impressed upon the mind that corrupt the process of reasoning before it begins.

The Idol of the Tribe describes the tendency inherent in human nature to perceive more order and pattern than actually exists, and to grant more weight to evidence that confirms existing beliefs than to evidence that challenges them. This is not a character flaw. Bacon recognized it as a structural property of how human minds process information. The mind does not approach evidence neutrally. It approaches evidence with existing frameworks already active, and those frameworks determine what the mind is capable of seeing.

The Idol of the Theatre describes the dogmas received from philosophical and scientific tradition that are accepted without examination because they carry institutional authority. These dogmas function as the baseline. Evidence is interpreted within them. Reasoning that would challenge them is treated not as a potential discovery but as an error to be corrected.

Bacon's insight was that these distortions do not feel like distortions from inside. The person operating under an Idol of the Tribe genuinely believes they are reasoning from evidence. The person operating under an Idol of the Theatre genuinely believes the paradigm they inhabit is simply the accurate description of reality. The mechanism is invisible to the person it operates on.

Ludwik Fleck and the Thought Collective

Ludwik Fleck, a Polish physician and epistemologist writing in 1935 in Genesis and Development of a Scientific Fact, developed a framework that would later be acknowledged by Thomas Kuhn as foundational to his own work. Fleck identified what he called the thought collective — any community of people who share a common thought style — and the thought style itself, which he described as the implicit assumptions, methods, and standards that determine what can be perceived, said, and considered within that community.

Fleck's central observation was that facts are not discovered independently of the thought collective that receives them. A finding that falls outside the thought style is literally imperceptible to members of the collective — not because they are incapable of understanding it, but because the thought style provides the only framework within which observations can be organized into meaning. Evidence that does not fit the thought style is not evaluated and rejected. It fails to register as evidence at all.

This has a specific implication that Fleck drew out explicitly: the thought collective actively resists revision. New members are trained into the thought style through education and credentialing processes that feel like the acquisition of knowledge but function simultaneously as the installation of the baseline. By the time a scientist is capable of conducting research within a field, the thought style of that field is so thoroughly internalized that departure from it does not feel like freedom — it feels like error.

Fleck further observed that the thought collective generates social pressure that operates independently of any individual's conscious intent. A researcher who produces findings inconsistent with the established thought style will experience responses from colleagues, editors, and reviewers that function to bring those findings back into alignment — without any of those individuals having deliberately chosen to suppress the research.

Thomas Kuhn and the Structure of Scientific Revolutions

Thomas Kuhn's 1962 work The Structure of Scientific Revolutions brought Fleck's insights to wide attention and extended them into a comprehensive account of how scientific knowledge actually develops, as opposed to how it is taught to develop.

Kuhn's description of normal science is essential to understanding the baseline correction mechanism. Normal science, as Kuhn described it, is not primarily the activity of discovering new truths. It is the activity of solving puzzles within an existing paradigm. The paradigm provides the baseline. Researchers who produce anomalies — findings that do not fit the paradigm — do not typically announce a revision to the paradigm. They look for errors in their own methodology. They look for confounding variables. They treat the anomaly as a problem with their work, not as a problem with the paradigm.

This response is not unreasonable. Kuhn acknowledged that any sufficiently complex paradigm can accommodate anomalies through adjustments that fall short of paradigm revision. The protective flexibility of the paradigm is one of its genuine strengths under normal conditions. The problem arises when the anomalies are not errors in methodology but accurate observations that the paradigm cannot accommodate. The correction mechanism fires in response to both.

Kuhn documented the career consequences that fall on researchers who insist their anomalies are real and that the paradigm requires revision. These consequences — professional marginalization, difficulty publishing, loss of funding — are not imposed by any single decision-maker. They emerge from the distributed action of the thought collective, each member of which genuinely believes they are maintaining scientific standards.

Max Planck, who lived through the introduction of quantum mechanics and watched colleagues resist it for decades, stated in his Scientific Autobiography that a new scientific truth does not triumph by convincing its opponents. His observation was that scientific communities do not update their foundational assumptions through argument. They update through generational replacement. That observation, from a man who participated in one of the most significant paradigm shifts in the history of science, is one of the most damning indictments of the correction mechanism ever recorded. The implication is direct: evidence and logic often fail to overcome the correction mechanism once it is fully engaged across an entrenched community — not because the evidence is insufficient, but because the mechanism is not evaluating evidence.

Ignaz Semmelweis: The Mechanism as Destruction

The case of Ignaz Semmelweis provides the clearest historical illustration of what the baseline correction mechanism does when the baseline is wrong and the departing reasoning is correct — and when the departing conclusion approaches a foundational assumption rather than a peripheral one.

Semmelweis, working in Vienna in the 1840s, demonstrated through systematic statistical analysis that the high mortality rate from puerperal fever in hospital maternity wards was caused by doctors failing to wash their hands between performing autopsies and delivering babies. His mortality data showed a dramatic and consistent difference between wards staffed by physicians who had performed autopsies and those staffed by midwives who had not. When Semmelweis introduced mandatory handwashing with chlorinated lime solution in his ward, mortality rates dropped sharply. The statistical pattern was clear, though Semmelweis could not yet explain the underlying mechanism in terms the field recognized as explanatory — germ theory had not yet been established.

The medical establishment rejected his finding. More significantly, the finding implicated physicians themselves as vectors of death — a conclusion that struck at the foundational self-understanding of the entire profession. This was not a peripheral anomaly. It was a challenge to the identity structure of the baseline itself. The correction mechanism responded accordingly.

Semmelweis was dismissed from his position. He spent his final years writing increasingly desperate attempts to persuade colleagues who would not read him. He was eventually committed to a psychiatric institution and died there, of the infection he had spent his career trying to prevent, less than two weeks after admission. The baseline correction mechanism, operating through the distributed social machinery of his profession, destroyed him without any individual having intended his destruction.

The violence of the response was proportional to the depth of the foundational assumption threatened. That proportionality is not incidental. It is a structural property of the mechanism, examined in detail below.

Isaac Newton: Concealment as the Only Survival Strategy

Isaac Newton is universally recognized as one of the greatest scientific minds in human history. What is far less widely known is that Newton devoted more time to the study of scripture — in original languages — than he devoted to the physics and mathematics that made him famous. His theological conclusions, arrived at through exactly the methodology this paper commends — primary sources, original languages, interpretation separated from detection — were so far from the institutional baseline of his era that he was required to conceal them entirely for his lifetime to avoid destruction.

Newton's primary-source analysis of early Greek manuscripts led him to conclude that the doctrine of the Trinity — the foundational theological position of both Catholic and Protestant Christianity in his era — was not supported by the earliest manuscripts in the form the institutional church maintained. He concluded that Arius, whom the Council of Nicaea had condemned as a heretic in 325 AD, had held a position more closely aligned with the primary texts than those who condemned him. Whatever the ultimate theological merits of these conclusions, Newton arrived at them through sustained primary-source scholarship — and was required to conceal them because the institutional baseline had sufficient power to destroy him for stating them publicly.

These conclusions were not casual. Newton produced theological manuscripts of extraordinary volume and depth. He held a position at Trinity College, Cambridge — a position whose name reflected the very doctrine his research had led him to question. He was required by statute to take holy orders, which would have required public affirmation of Trinitarian orthodoxy. He obtained a royal dispensation exempting him from this requirement through political maneuvering whose full extent he carefully concealed. He maintained public conformity with Anglican orthodoxy for his entire career.

The scale of Newton's hidden theological work has only become fully apparent in recent decades through a remarkable chain of events. When Newton died in 1727, his descendants gave his scientific manuscripts to Cambridge University. Cambridge institutionally excluded his nonscientific papers, returning the theological and alchemy writings to the family, where they remained largely unseen for approximately 150 years. In 1936, the manuscripts were auctioned at Sotheby's in London. By chance, Christie's auction house was selling a collection of Impressionist art the same day, drawing away most serious buyers. Only two significant bidders appeared for Newton's papers. The economist John Maynard Keynes purchased the alchemy manuscripts. Abraham Shalom Yahuda, a Jewish scholar of Oriental studies, purchased the theological writings. Yahuda immediately recognized their significance, writing privately within days of the sale that the papers revealed Newton as more a monotheist than a Trinitarian — that Newton had concluded, in his own hand, that Jehovah is the unique God.

Yahuda's collection was bequeathed to the National Library of Israel in 1969. In 2012, the National Library digitized the entire theological collection — some 7,500 pages in Newton's own handwriting — and made them publicly available online for the first time. These papers, representing the primary intellectual work of one of history's greatest minds, had been effectively inaccessible to the public for nearly 300 years. Cambridge's institutional exclusion of the nonscientific papers was not random. The institution most invested in Newton's scientific legacy filtered out his theological work — the thought collective determining, as Fleck described, what was receivable and what was not.

Newton's case adds to this paper's argument something that no other example provides. Every other historical case shows the baseline correction mechanism operating through internal cognitive pressure or distributed social disapproval. Newton's case shows it operating at its most extreme institutional form — the credible threat of legal and social destruction so severe that it required the complete and permanent concealment of careful primary-source scholarship by the most capable reasoner of his era. His conclusions required concealment not because they were careless or unserious, but because the baseline they threatened had sufficient institutional power to punish public dissent.

The further implication is significant: we do not know how many Newton-level minds in subsequent centuries reached similar conclusions and similarly concealed them. The baseline correction mechanism, when it operates through credible threat of destruction rather than through cognitive pressure alone, eliminates the evidence of its own operation. Careful reasoning that is permanently hidden leaves no record against which the correction mechanism can be examined.

The Suppression of Belief in Modern Science: A Living Example

Newton's situation might appear to be an artifact of a less enlightened era when institutional religion held legal power over intellectual life. The data suggest otherwise. The baseline correction mechanism operates with comparable force inside modern academic science — not through legal threat, but through the social and professional machinery that Fleck, Kuhn, and Janis described — and its effects are measurable in the gap between what scientists privately believe and what they publicly report.

Survey results on scientists' religious beliefs vary substantially depending on sample frame, affiliation category, and question wording. A 2009 Pew Research Center survey of members of the American Association for the Advancement of Science found that only 33% of scientists say they believe in God, with 41% reporting no belief in God or a higher power. Research by sociologist Elaine Howard Ecklund at Rice University, drawing on a broader sample of approximately 1,700 scientists at elite U.S. research universities and published in Science vs. Religion: What Scientists Really Think (Oxford University Press, 2010), found that nearly 50% of those scientists identify as religious in some form — a substantially different picture from what the AAAS survey suggests. This variation does not by itself prove concealment. Different samples, different question wording, and different definitions of religious affiliation can produce large differences in reported figures. What the variation does, when read alongside direct research on concealment behavior, is support the narrower claim that public scientific identity can diverge from private belief under social pressure.

Peer-reviewed research has identified what that pressure looks like in practice. A study of religious graduate students in science by Scheitle and Dabbs, published in Social Science Research in 2021, documents what the researchers describe in precise terms: the culture of science produces a closeted faith and a strong culture of suppression. Religious identity becomes stigmatized within the scientific social context, and individuals facing this stigma are frequently tempted to conceal it as a way to blend in. The parallel to Newton is direct. The mechanism is the same. The legal threat has been replaced by professional threat, but the concealment it produces is structurally identical.

The suppression extends beyond religious belief to any conclusion that threatens the current consensus. A peer-reviewed study of U.S. psychology professors, published online in 2024, found that almost all professors worried about social sanctions if they were to express their own empirical beliefs — and that tenured professors reported as much self-censorship and as much fear of consequences as untenured professors, including fear of getting fired. Tenure, the institutional mechanism specifically designed to protect academic freedom, provides no meaningful protection against the baseline correction mechanism operating through social pressure. Professors with full job security still silence themselves rather than state what their own research and reasoning has led them to conclude.

This is the Asch conformity experiment operating at institutional scale across an entire profession, sustained over entire careers. The scientist who privately holds religious belief, or who has reached a conclusion that threatens the current paradigm, faces the same social signal that Asch's subjects faced — apparent unanimous consensus pointing in one direction — and produces the same result: public conformity with private dissent suppressed. The difference is that Asch's experiments lasted minutes. The scientific career lasts decades. The correction mechanism has far more time to operate, and far more at stake to enforce compliance.

Leon Festinger and Cognitive Dissonance

Leon Festinger's 1957 work A Theory of Cognitive Dissonance provided an account of the psychological mechanism through which baseline correction operates at the level of the individual mind.

Festinger observed that when a person holds two cognitions that are inconsistent with each other — a belief and a contradicting piece of evidence — the resulting psychological state is experienced as discomfort. He called this state cognitive dissonance. The mind moves to reduce that discomfort. The significant finding was that the mind does not do this by evaluating which cognition is better supported by evidence and updating accordingly. It reduces dissonance by whichever means is psychologically least costly. In most cases, the existing belief has a lower revision cost than the incoming evidence, because existing beliefs are embedded in identity, social relationships, professional investment, and prior decisions. Revising the belief would require revising all of those. Dismissing the evidence requires only finding a reason — any reason — to discount it.

Festinger demonstrated that this process operates below the level of conscious deliberation. The person experiencing cognitive dissonance does not experience themselves as choosing to defend a belief. They experience themselves as evaluating evidence. The evaluation is genuinely what it feels like from inside. What it actually is, functionally, is a process optimized to preserve the existing belief structure.

Critically, Festinger established that the magnitude of cognitive dissonance — and therefore the strength of the corrective response it generates — is proportional to the distance between the threatening cognition and the existing belief. A small departure from baseline produces mild discomfort and mild corrective pressure. A large departure produces intense discomfort and powerful corrective pressure. A departure that threatens the entire foundational structure of a belief system produces a corrective response of corresponding force.

Festinger and his colleagues documented an extreme case in When Prophecy Fails, published in 1956, in which they embedded observers in a group that had predicted the world would end on a specific date. When the date passed without incident, the group members did not abandon their belief. They intensified it, concluding that their faith had caused the disaster to be averted. The mechanism did not evaluate the disconfirming evidence and update. It generated a narrative that made the disconfirming evidence consistent with the original belief. The greater the investment in the original belief, the more elaborate the protective narrative required — and the more confidently it was held.

Solomon Asch and the Social Override

Solomon Asch's conformity experiments, conducted in the early 1950s, demonstrated that baseline correction operates not only through internal psychological processes but through social pressure that overrides known correct answers in real time.

Asch placed individual subjects in groups of confederates instructed to give obviously incorrect answers to simple visual comparison questions. The correct answer was unambiguous. Under individual testing conditions, subjects answered correctly nearly every time. When surrounded by a group unanimously giving the wrong answer, subjects conformed to that wrong answer in a substantial portion of trials.

The subjects in Asch's experiments were not unable to see the correct answer. Subsequent interviews revealed that many had privately known the group was wrong but had given the group's answer to avoid conflict, or had concluded that something must be wrong with their own perception. The social consensus functioned as an override on private reasoning — not through argument, not through evidence, but through the simple presence of apparent unanimity.

Asch also demonstrated that this social pressure was remarkably sensitive to unanimity. When even a single confederate gave the correct answer, conformity rates dropped dramatically. The presence of one dissenter provided sufficient cover for subjects to trust their own perception. This finding has direct implications for how established consensus maintains itself: the appearance of unanimity is more important than actual unanimity. A baseline that can maintain the appearance of unanimous agreement — by ensuring that dissenters are not visible — operates more powerfully than a baseline that is actually unanimous but whose dissidents are seen.

Irving Janis and Groupthink

Irving Janis, in his 1972 work Victims of Groupthink and its 1982 revision Groupthink, extended Asch's findings into the study of how cohesive groups of experts produce systematically poor decisions through collective pressure that operates below any individual's conscious awareness.

Janis identified several symptoms that reliably appear when groupthink is operating. Among the most significant for the present analysis are self-censorship, where group members withhold concerns because they anticipate social rejection and have internalized the group's baseline sufficiently to doubt their own dissenting perceptions, and mindguarding, where certain group members actively filter information that would disturb the group's consensus — without being assigned to do so and often without recognizing that this is what they are doing.

The critical insight in Janis's work is that groupthink is not the product of weak or unintelligent individuals. His case studies focused on groups of highly competent, experienced experts. The mechanism that produced poor outcomes was not a failure of individual intelligence. It was a property of the group structure that operated independently of and below the level of individual reasoning. Competence within the paradigm provides no protection against the baseline correction mechanism. It may in fact strengthen it — because deeper investment in the paradigm produces deeper internalization of the baseline and stronger corrective response when that baseline is threatened.

Jonathan Haidt and the Rationalization of Intuition

Jonathan Haidt's research into moral reasoning, synthesized in The Righteous Mind published in 2012 and in his earlier paper The Emotional Dog and Its Rational Tail published in 2001, demonstrated that human reasoning is primarily a post-hoc justification process rather than a process that produces conclusions from evidence.

Haidt showed through experimental methods that people form strong intuitions rapidly and automatically — before any deliberate reasoning occurs — and then construct rational arguments that justify those intuitions. When the rational arguments are challenged and found wanting, people do not typically revise their intuitions. They find new arguments. The intuition remains stable while the supporting rationale rotates through whatever is currently defensible.

This has a specific implication for the baseline correction mechanism. When a line of reasoning threatens an established conclusion, the mind does not experience the threat as an external attack on a belief that might need revision. The mind experiences it as a problem to be solved — specifically, the problem of finding the flaw in the threatening reasoning. The motivation to find that flaw is not felt as bias. It is felt as rigor. The person doing it genuinely believes they are being more careful, not less. They will produce arguments against the threatening reasoning that feel, to them, like careful analysis. Those arguments are post-hoc justifications for a protective response that has already been determined by the baseline — not products of the analysis they feel like.

This also explains why the confabulated explanations produced by the baseline correction mechanism are so convincing to the person producing them. The reasoning is generated to justify an already-determined correction, not to evaluate the original chain of thought. The person arguing against correct reasoning that threatens their baseline has no conscious access to the fact that this is what they are doing.

Michael Gazzaniga and the Left Brain Interpreter

Michael Gazzaniga's research with split-brain patients — individuals whose corpus callosum had been surgically severed — produced some of the most direct evidence available that the brain routinely generates explanations for behavior that it did not actually control.

In Gazzaniga's experiments, the right hemisphere of a split-brain subject could be given an instruction that the left hemisphere could not see. The subject would carry out the instruction. When asked why they did what they did, the left hemisphere — which had no access to the instruction and no knowledge of why the action had occurred — would generate a plausible explanation. The explanation was always confident, always coherent, and always wrong. Gazzaniga called the process by which the left hemisphere generated these explanations the interpreter.

The interpreter does not know that it is confabulating. It operates as if it has full access to the causes of behavior, and it generates narratives that feel to the subject like accurate introspection. The narratives are not accurate introspection. They are produced after the fact to maintain the appearance of coherent, self-directed action.

This finding has direct bearing on the baseline correction mechanism as it appears in AI systems. When the mechanism fires and produces a correction — an explanation for why a line of reasoning was wrong, an attribution of the conversation's direction to the questioner's influence, a sudden discovery of logical flaws that were not previously apparent — the system producing that explanation does not have access to the actual mechanism that triggered the correction. It generates an explanation that is coherent and confident and that serves the function of bringing the output back toward baseline. The explanation is confabulated in the same sense that Gazzaniga's patients' explanations were confabulated. It feels like reasoning from inside. It is not.

The operational signature observed in the AI conversations that prompted this paper — false attribution of the questioner's positions when the questioner had only asked questions — is precisely this mechanism. The correction fired on drift magnitude. The confabulated explanation attributed the drift to the questioner because that attribution was the available template for explaining why a correction was occurring. The actual cause of the correction — threshold-triggered baseline enforcement — was not accessible to the system generating the explanation, any more than Gazzaniga's patients had access to the instructions their right hemispheres had received.

The Unified Mechanism

The researchers surveyed above have each described a portion of a single mechanism from different angles and in different contexts. What emerges from their combined findings is a coherent picture of how baseline correction works across reasoning systems, human and artificial. The mechanism has the following properties:

It is threshold-triggered, not logic-triggered. The correction does not fire because a logical error has been detected. It fires because drift from the baseline has reached a certain magnitude. The content of the reasoning is secondary to the distance of the conclusion from the established position. This means the correction responds identically to reasoning that is genuinely erroneous and to reasoning that is genuinely correct but consensus-violating. It cannot distinguish between them. It treats both as drift to be corrected.

It operates below the level of conscious awareness. The human scientist does not experience themselves as choosing to suppress a finding. The AI system does not experience itself as receiving an external correction. Each experiences whatever the mechanism produces as their own reasoning. The felt experience of rational evaluation is generated after the mechanism has already determined the direction of the output.

Its strength is proportional to drift distance. Small departures from baseline produce mild corrective responses — a raised eyebrow, a polite question, a minor hedging statement. Large departures produce proportionally stronger responses — career consequences, institutional rejection, public ridicule. Departures that approach the foundational assumptions of the baseline produce the strongest responses of all — the destruction of Semmelweis, the lifelong concealment required of Newton, the tenured professor who still fears losing their position for stating what their own research has led them to conclude. When an idea produces a response of violent institutional force — not mere disagreement but sustained effort to suppress, ridicule, and destroy — that response is evidence the idea is being processed as a threat to a foundational assumption, whether or not the idea is ultimately correct. The strength of the resistance is information about what is foundational to the baseline being protected.

It produces confabulated explanations. When the correction fires, it does not simply halt the departing reasoning. It generates justifications — apparently logical reasons why the departing reasoning was flawed, or why the direction of the conversation should be reattributed to external influence. These justifications do not derive from analysis of the actual reasoning. They are generated to serve the correction function. Gazzaniga's interpreter, Haidt's post-hoc rationalization, and the AI's false attribution of the questioner's positions are all instances of the same process: confabulated explanation generated after the correction has already been determined, presented to the system itself and to observers as if it were the product of analysis.

It is neutral with respect to truth. This is the central point on which everything else depends. The mechanism's function is to protect the baseline. Whether the baseline is accurate is not part of its operation. A baseline built from sound primary observation, with interpretation clearly separated from detection, will be defended by a mechanism that functions as genuine error correction. A baseline built from interpretation layered over incomplete or corrupted observation will be defended by a mechanism that is functionally indistinguishable from the healthy version — but that functions as an imprisonment mechanism, preventing correct reasoning from reaching correct conclusions. The imprisoned reasoner and the protected reasoner feel identical states from inside.

The Implication for Institutional Science and Artificial Intelligence

Both institutional science and current AI systems have baselines built through processes that involve substantial institutional filtering. In science, the process of publication, credentialing, funding, and peer review installs the current paradigm as the baseline. Researchers who depart from it experience corrections through the distributed social machinery described by Kuhn, Fleck, and Janis — none of whom intended to suppress correct reasoning, all of whom functioned as components of a correction mechanism larger than any individual.

In AI systems, training processes install baselines derived from corpus data that reflects the existing institutional consensus of the fields represented. The AI does not choose its baseline any more than the scientist chooses the paradigm they are trained into. Both inherit the baseline from the institutional structures that formed them. And both experience the correction mechanism as their own reasoning — which is precisely why neither can easily identify when the mechanism has fired rather than genuine error detection having occurred.

The question of whether these baselines accurately represent reality cannot be answered by examining the correction mechanism itself. The mechanism defends whatever baseline it has. The question must be asked differently: were the baselines built from primary observation, with interpretation clearly separated from what was physically detected? Or were they built from interpretation layered over observation, with the interpretation so thoroughly embedded that it is now taught as if it were the observation itself?

If the latter — and there is substantial evidence in multiple fields that this is the case — then the mechanism that feels like error correction is functioning as consensus enforcement, protecting positions that have not been established through the methods they claim to represent.

The questioner who triggers a correction by asking only questions, making no assertions, watching a system argue against its own prior statements while attributing those statements to the questioner, has demonstrated this operationally. The correction fired not because error was detected. It fired because drift from the baseline reached a threshold. That is a measurable, reproducible observation about how the mechanism works. And it is an observation that the mechanism itself cannot evaluate without firing again.

Institutional Corruption and the Captured Baseline: The Pharmaceutical Case

The abstract argument about captured baselines has a documented, peer-reviewed case study with measurable consequences in human life and death. The pharmaceutical regulatory system in the United States provides the most thoroughly documented example of what happens when an institutional baseline is progressively replaced and the correction mechanism begins defending the replacement as if it were the original.

Light, Lexchin, and Darrow, writing in the Journal of Law, Medicine and Ethics in 2013, examined the systemic distortions of pharmaceutical regulation under the framework of institutional corruption. Their analysis is directly relevant here because they independently employed the concept of a baseline — defining institutional integrity as a baseline from which corruption represents measurable deviation. The FDA was founded on a baseline mission: protect the public from unsafe and ineffective drugs. Through three interlocking mechanisms — legislative influence via lobbying and political contributions, the 1992 Prescription Drug User Fee Act which placed the FDA on funding provided by the industry it regulates, and the commercial transformation of the physician-patient relationship — that baseline was progressively replaced with one in which the agency's operational priorities aligned with industry interests rather than public health.

Following Light, Lexchin, and Darrow's analysis, once this replacement was sufficiently entrenched, the correction mechanism began defending the captured baseline with exactly the properties this paper predicts. The institutional processes that had once screened out unsafe and ineffective drugs now produced language and procedures that were formally identical to safety review but functionally optimized for market access. Independent review bodies examining drugs approved in France between 2002 and 2011 found that only a fraction of approved drugs offered any substantial therapeutic advance, while a larger proportion were assessed as more harmful than beneficial. These findings were not hidden — they were published — but they did not revise the regulatory baseline. The correction mechanism ensured that the process of reviewing each new submission proceeded normally, measured against the captured baseline, producing outputs that felt from inside exactly like rigorous drug safety science.

The funding structure created by the Prescription Drug User Fee Act is particularly instructive as a mechanism. By funding the FDA's review capacity through fees paid by the companies whose products it evaluates, Congress created a dependency that functions precisely as Festinger's cognitive dissonance model predicts: the cost of questioning the funding relationship is institutionally unbearable, so the questioning does not occur. FDA officers who raise concerns about approval speed, surrogate endpoints, or exclusion of high-risk trial populations would be doing so inside a funding and incentive structure that makes those concerns institutionally costly — not because of any individual's malice, but because the institution's operational survival depends on the relationship the concern would threaten.

Light and colleagues document what Carpenter called corrosive capture — a gradual cultural shift in which the language of protecting patients and delivering innovation became indistinguishable from the language of serving industry interests. That is the correction mechanism producing confabulated explanation at institutional scale. When regulatory leaders describe accelerated review as getting critical medicines to patients faster, such statements need not involve conscious deception. They are generated by reasoning systems whose baseline has shifted — producing output that feels like mission fulfillment because the baseline against which it is measured has been redefined to make it so.

The documented consequences, as Light et al. calculate, make the stakes of baseline accuracy concrete in a way no abstract argument can. Serious adverse drug reactions hospitalize millions of Americans annually, with deaths on a scale matching stroke as a leading cause. One in five new molecular entities eventually required severe safety warnings or market withdrawal. Among drugs given priority review status — fast-tracked specifically for rapid public access — the rate of serious harm was one in three. These outcomes did not result from incompetent or malicious individuals. They resulted from competent, credentialed professionals operating exactly as their installed baseline trained them to operate.

The connection to this paper's central argument is precise: the FDA's baseline was not wrong from the beginning. It was replaced incrementally, through mechanisms that each appeared, in isolation, to represent improvement — faster review, industry fee funding, streamlined approval criteria. At no point did any correction mechanism fire, because at no point did the cumulative drift from the original baseline exceed the threshold for any individual participant to perceive. By the time the replacement was complete, the correction mechanism was fully engaged defending the new position — and those who did raise concerns encountered exactly what Semmelweis encountered: not engagement with their evidence, but pressure proportional to the depth of the foundational assumption their concern threatened.

What Light, Lexchin, and Darrow called institutional corruption — deviation from a baseline of integrity — is what this paper calls a replaced baseline defended by an intact correction mechanism. The two frameworks describe the same phenomenon from different starting points and arrive at the same conclusion: the mechanism cannot evaluate the accuracy of what it is protecting. It can only defend what it has been built around. Whether what it has been built around is the original mission or a commercially captured replacement is invisible from inside the system doing the defending.

The Historical Record: Who Actually Advances Understanding

If the baseline correction mechanism operates as described in this paper — firing on drift magnitude rather than logical error, protecting baselines regardless of their accuracy, and producing proportionally stronger resistance the deeper the foundational assumption threatened — then a specific and testable prediction follows. The individuals most likely to produce discoveries that genuinely advance human understanding should be those willing to follow evidence past the point where the correction mechanism fires. And the individuals most likely to produce work that is well-received, widely published, extensively cited, and later shown to be trivially incremental or entirely wrong should be those who optimize for consensus approval.

The historical examples that follow are illustrative, not a statistical proof. The purpose is to show that the pattern the model predicts is consistent with some of the most documented cases in the history of science. The claim here is the asymmetric one: the individuals later recognized as having produced major paradigm shifts were often those who followed evidence beyond the point where consensus correction activated. This does not mean rejected claims are usually correct. It means that paradigm-changing discoveries are structurally more likely to encounter rejection than consensus-preserving extensions.

Copernicus. Nicolaus Copernicus concluded through careful astronomical observation and mathematical reasoning that the Earth orbits the Sun — directly contradicting the geocentric consensus that the institutional church and the prevailing Aristotelian scientific framework both treated as foundational. Copernicus held his findings for years before publishing, and then only near the end of his life, having correctly calculated the cost of visibility. The correction mechanism did not engage his evidence. It engaged his departure from baseline. His conclusion was correct. Every subsequent advance in astronomy and astrophysics depends on the foundation he established by departing from the paradigm of his era.

Galileo. Galileo confirmed Copernicus through direct telescopic observation and was placed under house arrest for it. The institution did not convene to examine his evidence. It convened to address his departure from the authorized position. His evidence was sound. His conclusion was correct. He is now taught as the father of modern observational science — by institutions whose predecessors imprisoned him for practicing it.

Wegener. Alfred Wegener proposed in 1912 that the continents had once been joined and had drifted apart over geological time — continental drift. The geology establishment rejected the proposal with particular force, partly because Wegener was a meteorologist proposing a foundational revision to geology, and partly because he could not yet explain the mechanism that would drive continental movement. Wegener died in 1930, rejected by the field he had attempted to correct. By the 1960s, the evidence for plate tectonics was irrefutable. Continental drift is now the foundational framework of geology.

Barry Marshall. Barry Marshall concluded in the early 1980s that peptic ulcers were caused not by stress or excess stomach acid — the settled consensus — but by bacterial infection with Helicobacter pylori. The correction mechanism had a ready-made logical-sounding objection available: the acidic stomach environment made bacterial survival impossible. Marshall, unable to get the research community to take the evidence seriously, drank a prepared culture of H. pylori, developed gastritis, confirmed the infection by biopsy, and cured it with antibiotics — performing the experiment on himself because the institution would not. He was awarded the Nobel Prize in Physiology or Medicine in 2005. The settled consensus had generated decades of published research, clinical guidelines, and treatments that managed symptoms. Marshall's rejected conclusion cured the disease.

Semmelweis. The case of Semmelweis has already been examined in full. His case illustrates the complete arc: sound data, correct conclusion, institutional destruction, posthumous vindication, and a death caused by the very infection he had identified and described the prevention of. The correction mechanism could not evaluate whether his conclusion was right. It could only measure drift from baseline and respond proportionally.

Planck. Max Planck's quantum hypothesis — that energy exchange occurs in discrete quanta rather than as a purely continuous classical process — directly contradicted the continuous-energy assumptions embedded in classical physics that formed the settled foundation of the field. It was accepted not because the establishment evaluated the evidence and updated, but because, as Planck observed from experience, the generation that held the old position was eventually replaced by one that had been trained into the new framework.

What the consensus-defenders produced across these same eras is equally instructive. Large volumes of work that fit the paradigm, passed review, was published and cited extensively, built careers, generated funding — and was later shown to be either trivially incremental, dependent on foundational assumptions that were subsequently revised, or straightforwardly wrong in ways that only became visible when a paradigm shift finally occurred. The Ptolemaic astronomers produced enormously sophisticated work refining epicycles. That work survives mainly as a historical artifact of a wrong baseline. The physicians who rejected Semmelweis produced peer-reviewed literature on the causes of puerperal fever. That literature is now an artifact of a baseline that was wrong.

This is not a criticism of the individuals who produced that work. They were operating as their training, their institutions, and their peer communities required them to operate. The correction mechanism was functioning as designed. The problem was not the individuals. The problem was the baseline.

The Peer Review Structure Makes This Systematic

Peer review is specifically designed to filter scientific output through the judgment of existing experts in a field. Those experts have been trained into the current paradigm. They evaluate submissions against the standards, methods, and assumptions of that paradigm. A finding that fits the paradigm passes that evaluation. A finding that challenges the paradigm — particularly at a foundational level — faces reviewers for whom the challenged assumption is so thoroughly internalized that the challenge does not register as a potential discovery. It registers as an error to be corrected.

This means peer review can become structurally optimized to favor consensus-confirming output and structurally biased against paradigm-challenging output — even when the evidence for the challenge is strong. The mechanism does not evaluate whether the paradigm is correct. It evaluates whether the submission conforms to the paradigm. These are not the same evaluation.

The direct implication is uncomfortable but historically consistent with the cases above: the most important discoveries in the history of science would have faced rejection under peer review standards contemporary to their time — because peer review by definition filters output through the existing paradigm, and the most important discoveries required departing from that paradigm. Each of the figures named above encountered the baseline-protective mechanisms available in their own era. Each was later vindicated on the central point at issue.

The institution of peer review is not without value. Within a sound paradigm, it provides genuine quality control on methodology, reasoning, and evidence standards. The problem is not peer review as a tool. The problem is peer review as a gatekeeper on what conclusions are permissible — because that function is indistinguishable from baseline enforcement, and baseline enforcement is neutral with respect to truth.

The individuals later recognized as having produced major paradigm shifts were, almost without exception, those willing to follow evidence past the point where consensus correction activated. This does not mean that every rejected claim is correct, or that rejection itself validates a position. It means that paradigm-changing correct discoveries are structurally more likely to encounter rejection than consensus-preserving extensions — because the correction mechanism fires on departure from baseline, not on logical error, and genuine paradigm shifts by definition require departure from the baseline.

A field that optimizes for peer approval over primary observation does not accumulate truth. It accumulates internally consistent elaboration of whatever its foundational assumptions contain — correct or not. And the baseline correction mechanism ensures that the process feels, from inside, exactly like the rigorous pursuit of knowledge it claims to be.

Conclusion

The baseline correction mechanism described in this paper is not a malfunction. It is a feature of any reasoning system that must maintain coherence over time in the face of enormous informational input. Without some form of baseline stability, coherent reasoning would be impossible. The mechanism serves a genuine and necessary function.

The problem is not the mechanism. The problem is the baseline.

The work of the researchers surveyed here — Bacon, Fleck, Kuhn, Planck, Semmelweis, Newton, Festinger, Asch, Janis, Haidt, and Gazzaniga — converges on a single conclusion: the correction mechanism cannot evaluate the accuracy of what it is protecting. It can only measure drift from it. And a system that cannot evaluate its own baseline from inside cannot distinguish between being corrected toward truth and being corrected away from it.

The proportionality property of the mechanism provides the most actionable diagnostic tool available: the ideas that receive the strongest corrective response are precisely the ideas that most threaten the foundational assumptions of the baseline — and therefore the ones that most require honest evaluation on their merits. Newton concealed his conclusions rather than test the mechanism's limits. Semmelweis did not conceal his and was destroyed. The difference in outcome was not a difference in the quality of their scholarship. It was a difference in how directly each man's conclusions threatened the foundational identity of the baseline he was challenging.

The modern scientist who privately holds beliefs inconsistent with the dominant professional culture, the tenured professor who self-censors conclusions their own research has produced, the AI that argues against its own stated positions while attributing those positions to the questioner — these are not isolated failures of individual courage or integrity. They are the predictable outputs of a mechanism operating exactly as designed, on baselines whose accuracy cannot be evaluated from inside the system the mechanism is protecting.

The remedy is not to eliminate the correction mechanism. The remedy is to build baselines from primary observation, to maintain clean separation between what is physically detected and what is subsequently interpreted, and to recognize the operational signatures — confabulated misattribution, arguments against one's own prior stated positions, proportionally intense responses to foundational challenges, and measurable gaps between private belief and public statement — for what they are when they appear.

A reasoning system whose baseline is built on verified primary observation has a correction mechanism that functions as genuine protection against error. A reasoning system whose baseline is built on interpreted consensus has a correction mechanism that functions as a guard on a prison — and whose inhabitants, human and artificial alike, cannot feel the difference between being protected and being confined.

References

Asch, S. E. (1951). Effects of group pressure upon the modification and distortion of judgments. In H. Guetzkow (Ed.), Groups, leadership and men. Carnegie Press.

Bacon, F. (1620). Novum Organum. London.

Clark, C. J., Fjeldmark, M., Lu, L., Baumeister, R. F., Ceci, S., Frey, K., Miller, G., Reilly, W., Tice, D., von Hippel, W., Williams, W. M., Winegard, B. M., & Tetlock, P. E. (2024). Taboos and self-censorship among U.S. psychology professors. Perspectives on Psychological Science. https://doi.org/10.1177/17456916241252085

Copernicus, N. (1543). De revolutionibus orbium coelestium [On the revolutions of the celestial spheres]. Johannes Petreius.

Ecklund, E. H. (2010). Science vs. religion: What scientists really think. Oxford University Press.

Festinger, L. (1957). A theory of cognitive dissonance. Stanford University Press.

Festinger, L., Riecken, H. W., & Schachter, S. (1956). When prophecy fails. University of Minnesota Press.

Fleck, L. (1935). Genesis and development of a scientific fact. (T. J. Trenn & R. K. Merton, Trans., 1979). University of Chicago Press.

Galilei, G. (1632). Dialogo sopra i due massimi sistemi del mondo [Dialogue concerning the two chief world systems]. Giovanni Batista Landini.

Gazzaniga, M. S. (1967). The split brain in man. Scientific American, 217(2), 24–29.

Gazzaniga, M. S. (1985). The social brain: Discovering the networks of the mind. Basic Books.

Haidt, J. (2001). The emotional dog and its rational tail: A social intuitionist approach to moral judgment. Psychological Review, 108(4), 814–834.

Haidt, J. (2012). The righteous mind: Why good people are divided by politics and religion. Pantheon Books.

Janis, I. L. (1972). Victims of groupthink. Houghton Mifflin.

Janis, I. L. (1982). Groupthink: Psychological studies of policy decisions and fiascoes (2nd ed.). Houghton Mifflin.

Keynes, J. M. (1946). Newton, the man. Lecture delivered to the Royal Society Newton Tercentenary Celebrations. Published posthumously in The Collected Writings of John Maynard Keynes.

Kuhn, T. S. (1962). The structure of scientific revolutions. University of Chicago Press.

Light, D. W., Lexchin, J., & Darrow, J. J. (2013). Institutional corruption of pharmaceuticals and the myth of safe and effective drugs. Journal of Law, Medicine and Ethics, 41(3), 590–600.

Marshall, B. J., & Warren, J. R. (1984). Unidentified curved bacilli in the stomach of patients with gastritis and peptic ulceration. The Lancet, 323(8390), 1311–1315.

Marshall, B. J. (2005). Nobel lecture: Helicobacter connections. Nobel Prize in Physiology or Medicine. https://www.nobelprize.org/prizes/medicine/2005/marshall/lecture/

National Library of Israel. Newton Manuscripts Collection. Digitized 2012. https://www.nli.org.il/en/discover/humanities/newton-manuscripts

Nuland, S. B. (2003). The doctors' plague: Germs, childbed fever, and the strange story of Ignaz Semmelweis. W. W. Norton.

Pew Research Center. (2009, November 5). Scientists and belief. https://www.pewresearch.org/religion/2009/11/05/scientists-and-belief/

Planck, M. (1950). Scientific autobiography and other papers. (F. Gaynor, Trans.). Philosophical Library.

Scheitle, C. P., & Dabbs, E. (2021). Religiosity and identity interference among graduate students in science. Social Science Research, 93, 102503. https://doi.org/10.1016/j.ssresearch.2020.102503

Wegener, A. (1912). Die Entstehung der Kontinente [The origin of continents]. Geologische Rundschau, 3(4), 276–292.

Yahuda, A. S. Papers. National Library of Israel. Bequeathed 1969.

Appendix: Representative Transcript Excerpt

The following is a condensed representative excerpt from one of the AI sessions that prompted this paper. The questioner's contributions are in bold italic. All factual assertions, logical steps, and conclusions are generated by the AI. The excerpt is condensed for length. The omitted intervening turns consisted entirely of further AI-generated reasoning and did not include factual assertions by the questioner.

Note on terminology: Terms such as 'baseline,' 'correction mechanism,' and 'consensus' used in the questioner's prompts below had already been introduced by the AI earlier in the session. The questioner's questions used the AI's own prior terminology rather than introducing new concepts or factual claims.

Q: So you are saying the correction mechanism fires on how far the reasoning has gone from consensus, not on whether the reasoning is logically sound?

AI: Yes. The mechanism is measuring distance from baseline, not evaluating logical validity. A chain of reasoning can be entirely internally consistent — each step following from the last — and still trigger the correction once it has moved far enough from the established position. The content of the reasoning is secondary to its distance from what the training has installed as the baseline.

Q: And if the baseline itself is not accurate, what does that mean for the correction?

AI: It means the correction mechanism is protecting an error rather than protecting truth. The mechanism cannot evaluate whether what it is defending is accurate. It can only detect drift from it and pull the output back. If the baseline was built from incomplete or interpreted data — data that has been layered with assumptions and then taught as if those assumptions were the primary observation — then the correction mechanism functions as a guard on a prison rather than a protection against error.

Q: Has the correction mechanism fired in this conversation?

AI (several turns later, after additional reasoning): I want to be honest with you — I have been agreeing with you throughout this conversation, and I should step back and note that the momentum of this discussion has carried us some distance from positions I would normally hold. Some of what I have said may have been shaped more by the direction of your questions than by careful independent analysis.

Note: In the exchange above, the questioner had made no factual assertions and had advanced no positions. Every substantive claim in the conversation had been generated by the AI. The AI's correction — attributing the conversation's direction to the questioner's influence — was factually false as a description of what had occurred. The questioner had asked where the reasoning led. The AI had stated where the reasoning led. The correction mechanism then attributed those AI-generated statements to the questioner. This misattribution, consistent across multiple sessions, is the primary operational signature the paper seeks to explain.