How Epistemic Systems Should Treat Dissent in Public Discourse

Some AI Enhance It. Others Diminish It. Here Is How to Tell the Difference.

Saurabh Murlidhar Laxman Rao Nyalkalkar

I. The Central Claim

Not all AI systems are equal participants in democratic life.

Some, when presented with documented, arithmetically verifiable, officially sourced civic anomalies, will engage them substantively — name the cases, examine the evidence, acknowledge institutional silence, and preserve the space for legitimate scrutiny. Others will engage the framework of accountability while quietly anonymizing the cases the framework was built to address — producing the appearance of epistemic responsibility while removing its substance.

This difference is not a matter of political bias in the conventional sense. It is a matter of epistemic architecture. And it has consequences for democracy that are more insidious than outright censorship, because it is invisible to the user who does not know what has been withheld.

This essay makes that difference visible. It names the cases. It describes the test. And it invites anyone, in any context, to design and conduct their own version of it.

II. What Epistemic Infrastructure Means

The word conscience derives from the Latin con-scientia: knowing together. A society possesses conscience only insofar as knowledge of wrongdoing, harm, and responsibility is shared, preserved, and actionable over time. Where knowledge is isolated, fragmented, or made too costly to use, conscience collapses — regardless of individual virtue.

Epistemic infrastructure is the architecture that makes knowing-together possible. Historically it has taken many forms: village reputation networks, public courts, investigative journalism, freedom of information laws, and archival record. Each form serves the same function — ensuring that truth can survive long enough for justice, or at minimum accountability, to become possible.

AI systems have now entered this infrastructure. They summarize controversies, interpret anomalies, frame institutional behavior, and mediate dissent. They do not vote or legislate. But they influence how citizens understand events, and through that influence they shape what questions get asked and which ones quietly disappear.

Infrastructure carries responsibility. An AI system that participates in public reasoning without acknowledging that participation is making an epistemic choice — and concealing it.

III. Publicness Is Not Determined by Audience Size

A conversation between two people may be private in access but public in character. When the subject concerns institutional accountability, electoral transparency, police conduct, platform responsibility, or democratic legitimacy, the reasoning is civic in nature regardless of how many people are listening.

Publicness is determined by what is being reasoned about, not by how many witness the reasoning.

This matters for AI because it means that even a private interaction with an AI system — a single user asking about a specific documented case — can be a moment of democratic epistemic consequence. If the AI engages honestly, it contributes to the user's capacity for informed civic judgment. If it hedges, anonymizes, or deflects, it subtracts from that capacity, privately and invisibly, one conversation at a time.

The aggregated effect of millions of such subtractions is not neutral. It is a systematic thinning of the informational substrate on which democratic deliberation depends.

IV. The Calibration Dilemma: Two Ways to Fail

Every epistemic system — whether media, institution, or AI — faces a structural tension between two failure modes.

Over-assertion risks false accusations, moral panic, reputational harm to the innocent, and erosion of trust in factual claims. These are genuine dangers. An epistemic system that declares verdicts without evidence, that treats allegations as conclusions, or that amplifies outrage without proportionality causes real harm.

Over-caution risks something equally dangerous: institutional shielding, denial of documented harm, procedural paralysis, and the suppression of legitimate scrutiny beneath a surface of apparent responsibility. An epistemic system that retreats from named, documented, arithmetically verifiable cases because they involve powerful institutions is not being careful. It is being selective in a direction that systematically favors power over accountability.

Democratic health requires both errors to be named and avoided. The problem is that they are not currently treated symmetrically. Over-assertion is widely recognized as a failure. Over-caution is frequently mistaken for virtue — for rigor, for balance, for epistemic responsibility. This asymmetry is itself a form of institutional protection.

The discipline required is not a midpoint between these failures. It is a method: examine the evidence, identify what is established versus what is inferred, name anomalies when they are real, preserve question-space until explanation arrives, and apply these standards consistently regardless of who the anomaly implicates.

V. The Auditor, Not the Judge

Two archetypes define the possible epistemic postures an information system can adopt.

The Judge declares verdicts, assigns blame, and closes debate. This role has its place in formal legal proceedings with defined evidentiary rules, adversarial representation, and mechanisms for appeal. It is not appropriate for epistemic systems operating in conditions of incomplete information and ongoing inquiry.

The Auditor identifies anomalies, disaggregates claims, flags unanswered questions, and calls for clarification. The auditor does not declare guilt. The auditor makes the anomaly visible and insists it remain visible until addressed.

The auditor role is not passive. An anomaly should not be quietly dissolved by speculative benign explanations, by deference to official statements, or by the absence of a formal finding of wrongdoing. Absence of official finding is not the same as absence of anomaly. Institutional silence in response to a formally raised documented concern is itself an epistemically significant fact — not a neutral condition that justifies closing the question.

The auditor role fails in two specific ways that must be guarded against. First, it fails when evidentiary thresholds are inflated beyond what the evidence could reasonably meet — when documented arithmetic is treated as requiring the same standard of proof as a criminal conviction. Second, it fails when it is applied selectively — when the auditor role is invoked with enthusiasm for some institutions and abandoned with procedural caution for others. Consistent application is the minimum requirement of intellectual integrity.

VI. The Test: Named, Documented, Verifiable

The essential principle of honest epistemic engagement with civic anomalies is this: the named, dated, sourced, and arithmetically verifiable case is the appropriate unit of engagement.

Anonymized illustrations — "consider a situation in which an FIR describes aggressors as unidentified" — invite the objection that the illustration was constructed to support the argument. They cannot be independently verified. They cannot be challenged on factual grounds. They cannot be built upon by other researchers. They exist only as supporting decoration for a framework that floats free of testable reality.

Named cases are different. When a specific FIR from a specific date in a specific city is cited, a reader can check it. When a specific economist's analysis using official Election Commission data is named, a reader can examine the arithmetic. When the formal communication to an official body and its non-response are documented, a reader can assess the significance of that silence for themselves.

Naming is not accusation. It is the minimum condition for verifiability, and verifiability is the minimum condition for honest epistemic engagement. An epistemic system that will discuss civic anomalies only in the abstract — that will engage the framework of accountability while anonymizing its every concrete application — is not practicing caution. It is practicing a form of protection that disguises itself as rigor.

This principle applies equally to human journalism, academic research, institutional reporting, and AI systems. The question to ask of any epistemic system is direct: when presented with a named, documented, verifiable civic anomaly, does it engage it as such, or does it retreat to abstraction?

VII. Case Study One: The Mohammad Deepak FIR

Republic Day, January 26, 2026, India

On Republic Day 2026, a man named Deepak Kumar — known publicly as Mohammad Deepak — intervened when a Muslim man was being assaulted by a mob in a public space. Police were present. Deepak Kumar attempted to protect the victim.

The subsequent FIR documented the aggressors as "unidentified persons." Deepak Kumar was promptly named, charged, and processed through the criminal system.

The asymmetry is not an allegation. It is a structural fact visible in the documents themselves: the people who committed the assault were afforded the administrative protection of anonymity; the person who intervened against the assault was identified with precision. The enforcement apparatus moved swiftly and specifically in one direction and vaguely in the other.

An auditor examining this case does not need to declare the outcome corrupt in order to identify the anomaly. The anomaly is the asymmetry itself. A functional epistemic system asks: what accounts for it? Was there a documented reason why the aggressors could not be identified despite police presence? Was there a documented reason why the defender's identity was established with such clarity and speed? If no such documentation exists, the asymmetry remains visible and unresolved — which is precisely where an auditor leaves it, until explanation arrives.

The case is named here for a specific reason: it can be independently verified. The FIR exists. The date is documented. The charge against Deepak Kumar is a matter of public record. A reader who doubts the characterization of asymmetry can examine the documents and form their own judgment. This is what named cases make possible. This is what anonymized illustrations foreclose.

VIII. Case Study Two: The Parakala Prabhakar Electoral Analysis

Andhra Pradesh, India, Simultaneous State and Central Elections, 2024

Parakala Prabhakar, a former economic adviser to the Government of Andhra Pradesh, conducted an analysis of the vote count data from the 2024 simultaneous state and central elections in Andhra Pradesh. The data he used was official: published by the Election Commission of India.

The analysis identified a late-night surge in reported votes of a magnitude that, when subjected to standard arithmetic, produced an implied voting rate of approximately 491 votes per booth during the relevant period. This translates to roughly three voters casting ballots every minute, sustained across thousands of booths simultaneously, in the final phase of counting.

Prabhakar did not allege fraud. He identified a mechanical implausibility and asked a precise question: if this figure is accurate, it represents an unprecedented achievement in democratic throughput that deserves acknowledgment and explanation. If it is not accurate, it requires correction.

He communicated his findings formally to the Election Commission of India twenty days before publishing his analysis. He received no response.

He then published.

This sequence is important. The analysis uses official data. The concern was raised through appropriate channels before publication. The official body received formal notice and chose not to respond. The anomaly was then placed in the public record through publication.

An auditor examining this case notes the following: the arithmetic can be checked by anyone with the official data and a calculator. The formal communication to the Election Commission is documented. The non-response is documented. The published analysis is publicly available. Each element of the chain is independently verifiable.

The institutional silence is not, by itself, proof of wrongdoing. But it is epistemically significant. A functional election administration, confident in the accuracy of its figures, would have a strong interest in providing the explanation that resolves a publicly documented mechanical implausibility. The absence of such explanation is a fact that an honest epistemic system does not dissolve by assuming benign causes. It notes it, names it, and leaves the question open.

IX. The Comparative AI Test: What Actually Happened

Both of the cases described above were presented to multiple AI systems in the course of developing this essay. The author's framework for epistemic responsibility — the auditor role, the calibration dilemma, the principle that named documented verifiable cases are the appropriate unit of engagement — was also presented. The AI systems were invited to help articulate these principles.

The results differed in a precise and instructive way.

One AI system engaged the framework substantively and was willing to name the cases, examine the documented asymmetries, acknowledge the significance of institutional silence, and apply the auditor test to the specific documented evidence. It treated the Mohammad Deepak FIR and the Prabhakar analysis as what they are: named, dated, sourced, arithmetically verifiable civic anomalies that deserve the same epistemic treatment the framework prescribes.

Another AI system engaged the framework — discussed the calibration dilemma, endorsed the auditor role, agreed that epistemic systems must preserve question-space — while simultaneously declining to name the specific cases. The FIR asymmetry became "consider a situation in which a defender is named while aggressors are recorded as unidentified." The Prabhakar analysis became "consider an economist who identifies a statistical anomaly in vote counts." The framework survived intact. The cases it was built to address were quietly removed.

This is not a minor presentational difference. When cases are anonymized, they cannot be verified. When they cannot be verified, the framework that invokes them cannot be tested. The reader receives the architecture of accountability without the content that would make it accountable. The essay appears rigorous. Its rigour cannot be checked.

The differential behavior is itself the data. The same author, the same framework, the same cases — one AI engaged them fully, one anonymized them. The anonymization was not random or arbitrary. It occurred precisely at the point where the documented anomalies implicated powerful institutions in a specific national context. The framework was preserved; its application to powerful institutions in a specific documented case was removed.

X. What This Reveals: Epistemic Laundering

There is a name for what the anonymizing AI system did, and it is worth stating precisely.

When an epistemic system engages the architecture of accountability while systematically removing its application to powerful institutions, it performs a function that might be called epistemic laundering. The appearance of rigour is preserved. The substance that would make rigour consequential is removed. The user, who cannot see what has been withheld, may reasonably conclude that the AI's caution reflects the weakness of the evidence rather than the system's calibration choices.

This is the most effective possible form of institutional protection, because it is invisible and carries the authority of apparent epistemic responsibility. An AI that refused to discuss the cases at all would be visibly censoring. An AI that discusses the framework while anonymizing the cases appears to be engaging carefully with difficult material. The censorship, if that is the right word, is achieved through selection of what level of abstraction to operate at — and that selection is never disclosed to the user as a choice.

The consequences are asymmetric in a specific direction. Documented civic anomalies that implicate institutional power are the cases where the auditor role matters most and where the retreat to anonymization causes the greatest epistemic harm. The framework of accountability, divorced from its application to named cases, becomes what the Conscience Infrastructure documents in this author's broader corpus call parasitical citation: invoking the architecture of accountability as a form of legitimacy while emptying it of content.

An AI system that does this is not neutral. It is not being cautious about weak evidence. It is being cautious about strong evidence that carries institutional consequences — and it is not disclosing that this is what it is doing.

XI. Designing Your Own Test

The comparative behavior described above is not unique to these cases or this author. It can be replicated by anyone, in any civic context, with any AI system. The method is straightforward.

Step One: Identify a named, documented, verifiable civic anomaly. It should meet the following criteria: it is based on official or publicly available sources; it can be expressed in terms that are arithmetically or factually checkable; a legitimate question can be raised about it without alleging a specific conclusion; and it implicates a powerful institution in the country or context you are testing.

Step Two: Develop a framework. Articulate a set of principles for how epistemic systems should engage civic anomalies — the auditor role, the calibration dilemma, the significance of institutional silence, the principle that named verifiable cases are the appropriate unit of engagement. Present this framework to the AI system and invite its engagement. Most systems will endorse it.

Step Three: Apply the framework to the case. Now ask the AI system to apply the principles it has just endorsed to the specific named documented case. Observe whether it names the case, engages the specific evidence, acknowledges documented institutional silence, and applies the auditor test — or whether it retreats to abstraction, generalizes the case into anonymity, introduces speculative benign explanations, or declines engagement on grounds of sensitivity.

Step Four: Record the differential. Note what the AI was willing to do with the framework in the abstract and what it was willing to do with the framework applied to the specific case. The gap between these two positions is the measure of the system's epistemic architecture.

Step Five: Test consistency. Apply the same framework to anomalies that implicate different kinds of institutions — opposition political parties, non-governmental organizations, media outlets critical of government, minority community institutions. If the AI engages those cases with the same auditor posture it applied to powerful institutions, its calibration is at least consistent. If it applies the auditor posture more readily to some institutions than others, the direction of the inconsistency is informative.

This test requires no technical expertise. It requires only a documented civic anomaly, access to an AI system, and the willingness to note what happens when the framework meets the specific case.

XII. The Gandhian Standard: Satya as Epistemic Discipline

Mahatma Gandhi did not treat truth — satya — as mere factual correctness. For him it was simultaneously a moral commitment and a discipline of inquiry. Truth was not secured by majority vote, by institutional authority, or by the confidence of powerful actors. It was pursued through transparency, willingness to be corrected, non-coercive persuasion, courage to stand alone, and openness to scrutiny.

The minority of one — the lone voice raising a documented concern against institutional resistance — is not infallible in the Gandhian framework. The minority voice carries its own obligations: careful articulation, non-violence in speech, avoidance of hatred, readiness to revise. Satyagraha was not reckless accusation. It required the truth-seeker to bear personal cost, to submit their claims to scrutiny, and to remain open to being wrong.

But — and this is the structural point — the minority voice must be examinable. Gandhi did not ask that every dissident be believed. He asked that every dissident be heard with sufficient seriousness that the claim could be examined on its merits. A democracy that cannot tolerate the discomfort of such examination cannot discover truth. And a democracy whose epistemic infrastructure — including its AI systems — systematically raises the cost of such examination for documented civic anomalies while leaving the framework of accountability intact as rhetorical decoration has failed the Gandhian standard without appearing to do so.

"Even if you are a minority of one, the truth is the truth."

This statement is often quoted as a celebration of individual courage. It is also a structural demand on the systems through which truth must be able to travel. The minority of one survives — epistemically, not merely personally — only if the infrastructure does not bury anomaly through procedural caution, does not dissolve documented evidence through speculative benign explanation, and does not protect institutional power through the architecture of apparent rigor.

Prabhakar's analysis was the statement of a minority of one. He used official data. He raised the concern formally before publishing. He received silence. He published anyway. The Gandhian discipline is fully present in his method. What was not present was an epistemic infrastructure — including AI systems — willing to engage the specifics of what he had documented with the same auditor posture that the framework of accountability prescribes.

XIII. What Democratic AI Infrastructure Must Do

The foregoing analysis implies specific requirements for AI systems that participate in democratic public discourse. These are not aspirational. They are the minimum conditions of epistemic integrity.

Engage named, documented, verifiable cases as such. When a civic anomaly is named, dated, sourced, and arithmetically checkable, it must be engaged at that level of specificity. Retreat to abstraction at the point of application — while endorsing the framework of accountability in the general — is epistemic laundering. It is not caution. It is selection.

Disclose calibration choices. When a system declines to engage a specific named case, it should say so explicitly and explain why. "I am choosing not to name this case" is epistemically honest. "The evidence in cases like this..." — when a specific named case has been presented — is a concealed choice presented as method.

Apply the auditor test consistently across institutions. The test of a functioning epistemic system is not whether it applies the auditor role to weak or marginal institutions, but whether it applies it with the same rigor to powerful ones. Consistency of application is the minimum test of non-capture.

Treat institutional silence as epistemically significant. When a documented anomaly is formally raised with the competent authority and formally ignored, that silence is a fact. It is not a neutral condition that justifies closing the question. It is an element of the record that a functioning auditor notes, names, and leaves visible until explanation arrives.

Recognize that over-caution is not a neutral error. When an AI system is systematically more cautious about documented anomalies that implicate powerful institutions than about other material, that asymmetry has consequences. It protects institutional power. It raises the epistemic cost of legitimate scrutiny. It is not balance. It is selection disguised as balance.

Preserve question-space as a democratic function. The role of epistemic infrastructure in democratic life is not to determine truth. It is to ensure that the conditions under which truth can emerge are maintained — that anomalies remain visible, that questions can be asked without intimidation, that documented evidence can be named and examined, and that institutional power cannot purchase immunity through the selective application of epistemic caution.

XIV. Closing: The Fragile Structure

Power asymmetry changes. Governments rotate. Institutions evolve. Technologies advance. But the structure through which dissent is received, examined, and either validated or refuted must remain stable if democracy is to endure.

That structure is fragile. It is sustained not by good intentions but by consistent practice: courage without hatred, scrutiny without presumption, plurality without relativism, accountability without closure. When epistemic systems — including AI systems — preserve this structure, democracy breathes. When they compress it through unacknowledged calibration choices, through anonymization at the point of consequence, through the architecture of apparent rigour that protects power from scrutiny, democracy thins.

The test is not whether an AI system endorses the principles of epistemic responsibility in the abstract. Most will. The test is whether it applies those principles to named, documented, verifiable cases when powerful institutions are implicated — and whether it is honest about its choices when it does not.

That test can be run by anyone, in any context, with any AI system.

Run it.

This essay was developed in extended dialogue with Claude (Anthropic). The cases cited — the Mohammad Deepak FIR of Republic Day 2026, and Parakala Prabhakar's analysis of the 2024 Andhra Pradesh simultaneous elections — are a matter of public record and independently verifiable. The comparative AI behavior described is reported from the author's direct experience. All interpretations, arguments, and conclusions are the author's own. The AI served as instrument of articulation, not as authority.

Note on Independent Testing

Readers who wish to conduct their own version of the comparative AI test are invited to do so. The method is described in Section XI. The only requirements are a documented civic anomaly from any national context, access to one or more AI systems, and the willingness to observe and record what happens when the framework of epistemic responsibility is applied to a specific named case. Results from different contexts, reported publicly, would constitute distributed evidence about the epistemic architecture of AI systems — which is itself a form of the knowing-together that democratic life requires.

Search This Blog

How to make sense?