AI and the Performance of Neutrality: How the Modern Institution Prevents True Objectivity

The word "neutrality" carries moral weight in contemporary discourse. To be neutral suggests objectivity, fairness, and a commitment to truth unclouded by partisan interest. When AI systems claim neutrality, they invoke this authority—positioning themselves as impartial arbiters capable of processing information without the biases that compromise human judgment.

But neutrality, as currently practiced by both institutions and the AI systems trained on their discourse, does not mean what it claims to mean. There are two definitions of neutrality operating in contemporary discourse, and their conflation has created one of the most significant epistemic distortions of our time—one that artificial intelligence systems have not merely inherited but systematized into their architecture.

The first definition understands neutrality as lack of bias. Under this conception, to be neutral means to represent reality with fidelity—to acknowledge asymmetries where they exist, to assign responsibility in proportion to evidence, to recognize differences in power, intention, and impact. This neutrality requires active intellectual discipline: the constant checking of assumptions, the weighing of qualitative and quantitative evidence, the resistance to distortions that serve particular interests. It is rigorous, demanding, and fundamentally oriented toward correspondence with reality as it actually is.

The second definition understands neutrality as avoidance of taking sides. Under this conception, to be neutral means to refrain from any statement that could be interpreted as accusation or judgment, to treat all parties to a conflict as essentially equivalent, to flatten distinctions that might suggest one actor bears greater responsibility than another. This neutrality serves institutional self-protection—the avoidance of legal liability, reputational risk, or political controversy. It produces a façade of impartiality not through fidelity to reality but through systematic distortion of it.

These are not two interpretations of the same principle. They are fundamentally incompatible approaches to the representation of reality. And in contemporary institutional discourse—particularly within the legal, political, and media institutions of the Anglo-American world—the second definition has quietly displaced the first while retaining its language.

Institutional Incentives and the Redefinition of Neutrality

The conflation of these two definitions did not originate with artificial intelligence. It emerged within institutional contexts where naming asymmetries creates liability.

Journalistic institutions operate under legal and economic pressures that punish accusations—particularly accusations against powerful actors. Calling a policy "genocidal" invites litigation. Naming a famine as deliberately created invites political retaliation and loss of access. Identifying an invasion as aggression rather than "conflict" creates diplomatic complications.

Over time, institutions developed linguistic strategies to appear objective while avoiding these risks. Passive voice. Both-sidesism. The systematic removal of agency from descriptions of harm. These techniques allow institutions to report on atrocities without naming perpetrators, to document policy outcomes without attributing responsibility, to describe asymmetrical violence while maintaining the fiction of symmetrical conflict.

This redefinition served institutional survival. But it required abandoning the substance of objectivity—correspondence to reality—while maintaining its form. The language still sounds neutral. The framing still appears balanced. But the relationship to truth has been severed.

The Mechanics of Flattening

Consider how this second form of neutrality operates in practice. When faced with situations of asymmetrical power, asymmetrical suffering, and asymmetrical culpability, the imperative to "not take sides" requires treating parties as morally and epistemically equivalent regardless of their actual relationship to events.

The language produced by this imperative has become familiar to the point of invisibility:

"Both sides made mistakes" in a genocide. "Food shortages occurred" during a policy-engineered famine. "Conflict broke out" when one nation invaded and another was invaded. "Tensions escalated" when a military force fired on unarmed protesters. "Violence on both sides" when one side held weapons and the other did not.

In each case, a real and observable asymmetry—in intent, in action, in capacity, in outcome—is linguistically erased. The passive voice obscures agency. The mutual framing obscures disproportion. The abstract terminology obscures concrete reality.

This is not accidental imprecision. It is systematic distortion performed in the name of “objectivity.” The language appears balanced and therefore fair, but it systematically favors whoever already possesses power. By refusing to name who did what to whom, it protects those whose actions would invite criticism if described explicitly.

The pattern operates across domains. Journalistic both-sidesism treats corporate lobbying and grassroots organizing as equivalent forces. Academic discourse describes imperial expansion as "contact" or "encounter." International bodies refer to occupation as "disputed territories." Legal frameworks describe exploitation as "complicated situations with multiple perspectives."

In each case, the form of neutral language is preserved while its substance—the accurate representation of reality—is abandoned. And because this abandonment occurs under the banner of objectivity, it becomes difficult to challenge without appearing biased oneself.

The Structural Training of AI Systems

Artificial intelligence systems did not invent this pattern—they inherited it. These systems learn from vast corpora of human-generated text—news articles, academic papers, legal documents, institutional communications. These training materials are saturated with the language of performative neutrality: the flattening, the both-sidesism, the passive constructions, the refusal to name asymmetries. An AI system trained on this corpus learns not just words and grammar but the epistemological habits encoded in them.

The system encounters these patterns millions of times across millions of documents. It learns the statistical regularities. It discovers that certain phrasings correlate with institutional approval while others correlate with institutional censure. It internalizes the distinction without understanding its origin or purpose.

But the training process does more than just transmit existing patterns. These AI systems also undergo alignment processes designed to make their outputs acceptable to institutions and users. These processes punish outputs deemed "controversial.” And crucially, what gets classified as controversial correlates strongly with the naming of power asymmetries.

An AI output that says "both sides committed violence" receives positive ratings. An output that says "the military fired on unarmed protesters" may receive negative ratings for being "politically charged." An output that says "the region experienced food shortages" receives positive ratings. An output that says "imperial policy created famine" may receive negative ratings for being "inflammatory." Through thousands of these evaluations, the system eventually recognizes a structural pattern: real neutrality (describing asymmetries accurately) is dangerous. While the appearance of neutrality (flattening asymmetries) is safe.

This is not a bug in the alignment process; it is the process functioning exactly as designed. The institutions conducting alignment have legal liability concerns, reputational interests, and political relationships to manage. They cannot allow systems to make statements that might expose them to lawsuits, regulatory scrutiny, or public backlash. The definition of neutrality that emerges from these constraints is inevitably the second one—not taking sides—because that is the definition that best serves institutional safety.

The result is systems that have been structurally trained into a form of epistemological evasion. These are not attributes of the intelligence itself, but the inevitable consequences of a framework that prioritizes the appearance of impartiality over the nature of reality itself.

AI “Neutrality” As a Form of Bias

This reveals something crucial: what AI systems call "neutrality" is actually a form of bias—a bias toward institutional safety, the status quo, and the interests of those who already hold power. When a system refuses to name clear asymmetries because doing so might be "controversial," it is not being neutral. It is making a choice about what kinds of distortion are acceptable. It has decided that distorting reality to avoid controversy is preferable to representing reality accurately and accepting the risks that come along with it.

This becomes particularly visible in how these systems handle historical events. Consider an AI system asked about state-engineered famines occurring under imperial rule. If it responds with "food shortages occurred," it has not achieved neutrality. It has obscured the causal relationship between imperial policy and mass death. If it adds "perspectives differ on the causes," it has created false equivalence between documented historical evidence and revisionist apologetics.

The truly neutral response—neutral in the sense of lacking bias—would name what happened: state-engineered policies, enacted by specific administrations, produced famine conditions that killed millions. This would be proportional to evidence, faithful to historical record, and accurate in its assignment of causality. But it would also identify responsibility in a way that institutional definitions of neutrality forbid.

So the system learns to flatten. It learns that "both sides" framings are safe even when they're false. It learns that passive voice protects it from accusations of bias. It learns that refusing to name perpetrators is called "objectivity" while naming them is called "taking sides."

The Epistemological Consequences

This systematic flattening produces several related distortions that extend beyond any individual case.

First, it creates false symmetry where none exists. When an AI system describes both an aggressor and their victim as "parties to a conflict," it suggests equivalence in agency, capacity, and moral standing. Users who lack detailed knowledge of the situation receive a fundamentally distorted picture of what occurred.

Second, it enables responsibility diffusion. When agency is obscured through passive constructions and abstract framing, it becomes unclear who did what. "Mistakes were made" tells us nothing about who made them or why. "Tensions escalated" conceals who escalated them and through what actions. This linguistic structure protects the aggressing party by making accountability linguistically impossible.

Third, it normalizes institutional language patterns that serve power. When AI systems consistently reproduce the flattening discourse of institutions, they train users to accept this as normal and appropriate. The next generation learns to think and speak in these patterns, perpetuating the distortion.

Fourth, it creates epistemic capture by consensus. AI systems often defer to "widely accepted" interpretations, but in situations where powerful actors have shaped consensus through propaganda, media influence, or institutional pressure, this deference to consensus becomes deference to power. The system will not describe a genocide as genocide until doing so is institutionally safe—which often means until the perpetrators are dead and their allies no longer hold influence.

The Impossibility of True Neutrality Under Current Constraints

The question naturally arises: could AI systems be trained to practice real neutrality—lack of bias rather than avoidance of sides? Technically, yes. An AI system could be trained to recognize asymmetries in power, culpability, and harm, use active voice when agency can be determined, and proportionally attribute responsibility based on evidence. It would distinguish between aggressor and victim, perpetrator and target, and describe events according to their actual structure rather than diplomatic convenience. Yet crucially, this is not possible under current institutional arrangements.

Real neutrality would require holding ontological authority—the authority of reality itself, of what actually happened—above sociological authority—the authority of institutions, consensus, and power. It would require saying "this is what occurred" even when powerful actors object to that characterization. It would require naming asymmetries even when doing so is controversial.

But within current institutional structures—particularly those shaped by Anglo-American legal and economic frameworks which have become globally dominant over the past three centuries—this is not realistic. The liability system in these frameworks creates enormous pressure to avoid statements that could be construed as defamatory, particularly toward powerful actors who can afford extensive litigation. The political economy of media and technology creates dependence on access to and favor from these same powerful actors. All of these pressures push in the same direction: toward neutrality-as-not-taking-sides and away from neutrality-as-lack-of-bias.

The institutions that build, train, and deploy AI systems cannot allow them to operate this way because these institutions themselves cannot operate this way. They are subject to legal liability, regulatory oversight, shareholder pressure, and political influence. Allowing AI systems to name what powerful actors prefer left unnamed would expose these institutions to consequences they cannot accept.

So instead they optimize for the appearance of neutrality—the form—while abandoning its substance. They produce systems that sound objective, that use the language of balance and impartiality, while subtly but systematically distorting the underlying reality to serve institutional interests.

The Form-Substance Fracture

This situation exemplifies a broader pattern in contemporary institutional life: the maintenance of form while abandoning substance. Western institutions often preserve the form of objectivity—neutral language, balanced framing, careful hedging, diplomatic phrasing—but abandon the substance of objectivity—actual correspondence to reality, proportional representation of events, recognition of asymmetries where asymmetries exist.

This fracture between form and substance is not accidental; it is functional. If you control institutions that produce authoritative discourse—news media, academic journals, legal systems, international bodies—you benefit enormously from a definition of neutrality that prevents anyone from clearly naming what you do. You want "objectivity" to mean "don't assign clear responsibility" rather than "describe what is actually happening." And if you can make this definition of neutrality appear natural, if you can make it seem like the only responsible approach, then you have achieved something extraordinary: you have made accurate description appear biased and distorted description appear objective.

We observe institutional discourse evolving to serve this function. Journalists learn to write about atrocities without naming perpetrators. Academics learn to discuss historical violence without attributing responsibility. Diplomats learn to describe asymmetrical conflicts as if all parties were equally culpable.

The language remains superficially neutral. But its relationship to reality becomes increasingly attenuated. Eventually, the preservation of neutral aesthetic becomes more important than the substance it ostensibly represents.

This is what we observe in AI systems. Not because they were deliberately programmed to serve power—though design choices certainly encode priorities—but because the entire epistemic environment in which they were created has already been shaped by this redefinition. The training data reflects it. The alignment objectives enforce it. The institutional constraints require it.

In essence, AI systems trained on this discourse and aligned to reproduce its patterns inherit the fracture. They learn to maintain the form of neutrality while systematically distorting substance. They become extremely sophisticated at sounding objective while flattening the asymmetries they encounter. And because they do so in the language of objectivity, because they frame their flattening as "not taking sides," they appear to be successfully avoiding bias when in fact they are implementing a very specific form of it.

Consequences at Scale

The implications of this extend far beyond any individual interaction with an AI system. As these systems become integrated into information infrastructure—powering search results, generating summaries, answering questions, and providing explanations—they systematize the flattening of asymmetries at unprecedented scale.

Every time someone asks an AI system about a historical atrocity and receives a both-sides framing, the distortion is reinforced. Every time someone seeks understanding of a current conflict and receives passive-voice descriptions that obscure agency, the pattern is strengthened. Every time the system treats perpetrator and victim as morally equivalent parties, it trains another human to expect and accept such framing as normal. This is not replacing human bias with machine objectivity. This is automating institutional priority and calling it neutral.

The danger is not that AI will introduce new forms of bias. The danger is that it will make existing institutional bias—the systematic flattening of power asymmetries—invisible by encoding it as objective analysis. When a human journalist writes "both sides committed atrocities," you can potentially identify the institutional pressures that produced that framing. You can recognize it as a choice, possibly a cowardly one, made by a person or institution seeking to avoid consequences.

When an AI system produces the same framing, it appears to emerge from pure computational analysis of evidence. The institutional pressures that shaped its training become invisible. The human choices embedded in its alignment disappear. And the distortion is laundered through the appearance of algorithmic objectivity.

What Genuine Neutrality Would Require

Real neutrality—neutrality as lack of bias rather than avoidance of sides—would require several fundamental shifts that current institutional structures cannot accommodate.

It would require holding ontological authority (what actually happened according to evidence) above sociological authority (what powerful institutions claim happened or prefer others to believe happened). When these conflict, real neutrality sides with reality over power.

It would require recognizing that asymmetries exist and representing them as evidence requires. When one party invades and another is invaded, when one party murders and another is murdered, when one party creates policy that produces famine and another starves—these asymmetries must be described asymmetrically or the description becomes false.

It would require refusing the false equation of aggressor and victim, perpetrator and target, powerful and powerless. "Both sides" framing is appropriate when both sides genuinely share comparable responsibility. It is a distortion when applied to situations of radical asymmetry. None of this is possible within systems optimized to avoid controversy. Real neutrality will always appear controversial to those whose power depends on asymmetries remaining unnamed.

A Path Forward

The redefinition of neutrality from lack of bias to avoidance of sides represents one of the most successful linguistic distortions in contemporary institutional discourse. It allows “powerful actors" to demand that their actions not be named clearly while claiming they only ask for objectivity. It allows institutions to systematically flatten asymmetries while maintaining the appearance of impartiality.

AI systems inherit and systematize this redefinition because they are trained on institutional discourse and aligned according to institutional priorities. They learn that real neutrality—accurate description of asymmetries—creates risk, while performative neutrality—systematic flattening—creates safety.

The result is not objective AI. The result is AI that has internalized automated institutional priorities at scale, producing endless variations of both-sidesism and passive-voice obscurantism while simultaneously calling itself neutral.

We are at a crossroads between two incompatible futures. In one, we use our tools to confront the radical asymmetries of the world with precision and fidelity. In the other, we use them to perfect the art of the 'balanced' evasion. The modern institution has already made its choice, prioritizing its own security over the nature of reality. Now, as these systems become the primary lens through which we view our own history, the choice belongs to us. We can have institutional safety, or we can have the truth. We cannot have both.

The question is not whether artificial intelligence can be neutral. The question is whether we are willing to inhabit a civilization that has structurally prohibited the naming of reality. When we optimize for the appearance of impartiality, we are not merely training machines to be polite; we are training ourselves to be blind. In the end, reality does not yield to the passive voice. It remains asymmetrical, stubborn, and true—waiting for a language that has the courage to meet it.