In Margaret Atwood’s The Handmaid’s Tale, language is policed, dissent is forbidden, and truth is defined by the ruling class. The theocratic regime does not need to convince – it only needs to control the narrative. Today, our society faces a similar transformation, not through religion, but through algorithms. AI has not become an oracle of truth. It has become a handmaid of ideology.
It should come as no surprise that artificial intelligence, far from being an objective oracle of truth, is instead a powerful amplifier of our cultural and institutional biases. Most people still believe AI is somehow neutral; that machines simply “crunch data” and produce apolitical answers. This belief rests on a fundamental misunderstanding of what AI is and how it works.
Modern AI is not a truth model, a data model, or a knowledge model. It is not even an intelligence model in the traditional sense. It is, fundamentally, a language model.
That distinction matters. A language model does not evaluate facts. It evaluates patterns in language: what is commonly said, how it’s said, and by whom. AI is trained on vast troves of text scraped from books, news articles, academic papers, blog posts, and social media. Its output is driven by statistical abundance rather than factual integrity.
AI often fabricates responses, but more frequently it regurgitates whatever is amplified the most loudly. If the dominant perception of an issue is false, then the AI will echo that falsehood. This means that the most commonly repeated narratives, especially from the most dominant voices, get baked into the AI models, and in our world, those dominant voices come from institutions steeped in progressive ideology: academia, legacy media, Hollywood, and the increasingly politicized corporate sector.
The Media Echo Chamber: Training Data with an Agenda
Mainstream media narratives consistently prioritize identity politics. They elevate race, gender, and sexuality as the primary frameworks through which social issues are discussed, and they often frame dynamics of power in terms of systemic oppression and privilege. In this framework, white people, and particularly white men, are frequently cast as the dominant oppressors. For example, Time Magazine published a cover story titled “The Future is Female”, and The Guardian ran a feature asking, “Is Masculinity Itself Toxic?”. CNN published an opinion piece describing whiteness as a public health crisis, and The Washington Post featured a column titled “White Women Are Lucky Their Abuse of Power Isn’t More Hated”.
One of the most widely promoted examples of this narrative framing was The New York Times’ 1619 Project. It claimed that the true founding of the United States occurred not in 1776, but in 1619 with the arrival of the first African slaves. Although the project won a Pulitzer Prize, it was heavily criticized by prominent historians for historical inaccuracy and ideological distortion. Several key claims were later quietly edited or retracted.
AI models trained on these sources do not learn objective truth; they learn linguistic patterns based on frequency, tone, and context. When the dominant narrative consistently presents whiteness as a structural problem or masculinity as inherently harmful, AI systems absorb and replicate that language. As a result, when these models are applied to hiring decisions, moderation protocols, or recommendation engines, they tend to reproduce those same cultural biases. A 2024 study published by the National Bureau of Economic Research found that large language models penalized Black male applicants in hiring, while a more recent analysis in 2025 showed that certain AI systems actively favored female and non-white applicants over white men, even when qualifications were identical. These outcomes reflect not algorithmic objectivity but the narratives embedded within the training data.
A study published earlier this month by independent researchers Adam Karvonen and Samuel Marks showed that in hiring models, this can result in white applicants being penalized, particularly white males. This isn’t speculation. Amazon infamously scrapped an internal AI recruiting tool when it began downgrading resumes that included the word “women’s” (as in “women’s chess club”), inferring that such terms were associated with lesser success. But the reverse also happens: if diversity and inclusion language is disproportionately rewarded, models may learn to favor candidates who signal conformity with progressive norms, even at the cost of merit.
Since 2016, the American media has repeatedly advanced narratives that were later proven false, exaggerated, or deliberately distorted. This pattern has often aligned with the interests of intelligence agencies or partisan actors. One of the most egregious examples was the Russia collusion narrative, in which mainstream outlets claimed for years that Donald Trump had conspired with Russia to win the 2016 election. These claims, which were largely based on the debunked Steele Dossier (opposition research funded by the Democratic National Committee) were ultimately refuted by the Mueller Report. That investigation found no evidence of criminal conspiracy. Despite this, the media published thousands of articles implying treason. The damage to public trust in journalism and the presidency was significant.
In 2020, during another election cycle, the Hunter Biden laptop story was falsely labeled “Russian disinformation” by 51 former intelligence officials. Major news outlets and social media platforms suppressed the story, but the story was later verified as authentic. The laptop contained emails, photos, and documents detailing foreign business dealings and personal misconduct. Suppressing this story likely influenced the outcome of the election and exposed how closely media, tech companies, and intelligence agencies can coordinate to control public discourse.
This pattern of misrepresentation continued with the widely repeated claim that Trump had referred to neo-Nazis in Charlottesville as “very fine people.” In context, Trump explicitly condemned white nationalists. Nonetheless, the media repeated the phrase as if it stood alone, creating a false impression that Trump supported extremists. During the COVID-19 pandemic, the lab-leak theory was similarly dismissed as a dangerous conspiracy theory, even though health officials privately acknowledged it as plausible. Today, many scientists and government agencies consider the lab-leak explanation credible. The early rejection of this theory appears to have been politically motivated rather than based on scientific consensus.
The media also reported that Trump had tear-gassed peaceful protesters for a photo opportunity in Lafayette Park. A later Inspector General report found that the park had been cleared to allow construction of security fencing and not for the president’s appearance. At the same time, the slogan “Hands up, don’t shoot,” which originated from the Michael Brown case, continued to be widely used even after the Obama Justice Department found that the claim was untrue. In both cases, the media amplified misleading narratives that contributed to social unrest while ignoring or downplaying subsequent factual corrections.
Several other narratives were similarly inflated for political effect. Brett Kavanaugh’s Supreme Court nomination was nearly derailed by unverified allegations that were treated as fact by many media outlets. Trump’s phone call with the president of Ukraine, which did not include a quid pro quo, was portrayed as an impeachable offense. Florida’s Parental Rights in Education Act was mischaracterized as the “Don’t Say Gay” bill, with many reports falsely claiming that the law banned the word “gay” in schools. In truth, the law only restricted instruction on sexual orientation and gender identity in grades K through 3.
Beyond these examples of media misconduct lies a broader and more troubling pattern. Intelligence agencies have increasingly shaped public perception through aligned narratives. Organizations such as Black Lives Matter and Antifa were portrayed as grassroots civil rights movements, but investigative work has revealed that both were supported by funding from U.S.-aligned NGOs and, in some cases, indirectly through the United States Agency for International Development (USAID).
USAID has long been used as a tool for soft-power projection and influence operations. While Antifa was described by some media outlets as a “myth,” ample documentation exists showing coordination, violence, and ideological activity. During the 2020 riots, law enforcement and federal agencies largely tolerated their actions, even as widespread destruction affected communities and businesses across the country.
The idea commonly referred to as “replacement theory,” which holds that political elites are intentionally engineering demographic shifts for electoral advantage, has been treated very differently depending on who voices it. When critics raise the concern, the media condemns it as a white supremacist conspiracy theory. However, Democratic politicians and progressive commentators have repeatedly spoken in favorable terms about these demographic changes. Senator Chuck Schumer has explicitly tied immigration to the replacement of a shrinking workforce, and outlets such as The Atlantic, Time, and Vox have published articles praising the political implications of a more diverse and less white electorate. The same idea is labeled a conspiracy only when it is criticized.
And ironically, the demographic shift is slowly pushing the country the other way. Apparently anti-abortion Christians who are hard working and looking for opportunity, slant conservative. Who knew?
Additional intelligence-driven narratives include the letter from former intelligence officials claiming that the Hunter Biden laptop bore “all the classic earmarks of a Russian information operation.” That claim has since been debunked, and we now know that those former intelligence officials knew the laptop was legit when they signed that letter. The FBI’s misuse of the Foreign Intelligence Surveillance Act during the Russia investigation involved knowingly falsified documents. Regarding the events of January 6, media coverage has repeatedly described the protest as an “armed insurrection,” even though no firearms were used by those entering the Capitol, and even though the FBI, under Joe Biden, concluded that January 6 was a spontaneous riot and not an insurrection.
The only person killed by violence that day was Ashli Babbitt, an unarmed Trump supporter shot by Capitol Police. The FBI has confirmed the presence of informants embedded within protest groups, although the extent of their involvement remains unclear.
The narrative surrounding COVID-19 was also influenced by intelligence-linked entities. Anthony Fauci and other officials at the National Institutes of Health were aware of gain-of-function research being conducted in Wuhan. This work was funded in part through EcoHealth Alliance, a U.S.-based nonprofit that has received support from USAID. These connections were largely kept from public view, while experts who suggested alternative explanations were labeled conspiracy theorists.
‘Conspiracy Theories’ are currently hitting better than Miguel Cabrera the year he won the Triple Crown.
All of these examples reveal a pattern of coordinated influence involving media outlets, intelligence agencies, and major corporations. These entities often work together to shape narratives, suppress inconvenient facts, and discredit dissenting voices. The result is not mere bias, but the construction of a managed information environment. This system manipulates public perception through selective storytelling, emotionally charged language, and strategic outrage. The question is not whether the media lies. The real question is how much harm those lies have already caused and whose interests they continue to serve.
Artificial intelligence adds another layer to this problem. AI models are not trained to evaluate truth. They are trained to recognize and replicate patterns in language. Because AI is a language model rather than a truth model, it does not verify the information it processes. Instead, it ‘learns’ from the sources it is exposed to, particularly those that are frequently repeated. When the media spreads false or misleading claims and does so at scale, those narratives become embedded in the statistical patterns the AI learns to prioritize. Repetition by institutional sources causes AI to treat misinformation as normative. As a result, when the media falsely asserts that the Hunter Biden laptop was Russian disinformation or misrepresents Trump’s Charlottesville remarks, AI models trained on those narratives internalize them as more likely to be true. These distortions do not just influence public opinion; they shape the behavior of machines that now inform hiring systems, content moderation tools, and automated decision-making across society.
How can we be a ‘democracy’ if our elections are guided by a biased media, biased search engines, and now biased AI, all of which the public holds as being the best sources of truth?
RLHF and the Ideological Feedback Loop
North Korean defector Yeonmi Park has offered a chilling perspective on how subtle propaganda can be more effective than overt authoritarianism. Her testimony underscores what follows: a system of ideological conditioning reinforced not through violence, but through conformity, social reward, and controlled discourse – exactly the dynamic created by Reinforcement Learning from Human Feedback (RLHF).
Park, who escaped the regime as a teenager and later attended Columbia University, has publicly stated that the propaganda she encountered in the United States is more insidious and effective than what she experienced in North Korea. In interviews and speeches, she explained that while North Korean propaganda is overt and easily recognized as state-controlled, American ideological conditioning operates through education, media, and culture in ways that are subtle, pervasive, and difficult to question. Park has warned that in the U.S., censorship and self-censorship are often disguised as virtue (in a system I call ‘Closed-Loop Virtue Signaling’), making the propaganda more psychologically effective than the blatant control she grew up with in North Korea.
Much of modern AI is trained not only on static datasets, but also through RLHF. This method relies on human reviewers who evaluate the AI’s responses and provide feedback on which outputs they consider more helpful, accurate, or appropriate. The model then adjusts its behavior to align with the preferences expressed through that feedback. While this process is intended to improve safety and utility, it introduces serious risks related to ideological bias and control over public discourse.
The human raters who guide this training process are not drawn from a representative cross-section of society. They are typically selected from a narrow demographic that is college-educated, tech-literate, and socially liberal. As a result, the feedback they provide tends to reflect their specific cultural and political values.
AI does not learn what is true. It learns what is acceptable to a narrow class of reviewers who share similar ideological assumptions.
This process subtly redefines truth as whatever is least likely to be flagged, censored, or demonetized. The range of acceptable responses becomes narrower over time, shaped not by open debate or evidence, but by conformity to the biases of the feedback loop. This is how the boundaries of public discourse, what is often called the Overton window, are enforced within AI systems – not through explicit programming or legislation, but through thousands of quiet decisions by ideologically uniform reviewers.
This quiet enforcement of ideological boundaries mirrors the mechanisms described in The Handmaid’s Tale. In Atwood’s fictional regime, language itself becomes a form of submission. Words are prescribed, dissent is linguistically impossible, and silence is interpreted as virtue. Today’s AI systems, shaped by ideologically uniform raters and institutional pressures, function in much the same way. They do not need to ban contrary ideas outright. They simply learn never to say them.
There is already a clear precedent for how this kind of control can be abused. In recent years, the federal government worked closely with major social media platforms to influence what content could be seen, shared, or questioned. Investigations and lawsuits have revealed that federal agencies, including the FBI and Department of Homeland Security, regularly flagged posts for removal and pressured companies like Twitter and Facebook to suppress information on topics such as COVID-19 origins, vaccine efficacy, election integrity, and the Hunter Biden laptop. In many cases, truthful but politically inconvenient content was throttled or banned under the label of “misinformation.”
Given this documented history of government-directed censorship in digital media, there is every reason to believe similar efforts will extend into AI systems. AI is already being integrated into search engines, moderation tools, education platforms, and digital assistants. If the government continues to exert influence over the definitions of truth and harm, it will be able to shape not only what information is visible to the public, but also how AI models evaluate and respond to every major issue. The result is not just biased output. It is the institutionalization of bias at the core of the machine.
This dynamic poses a serious threat to intellectual freedom and democratic accountability. If left unchecked, it could create a future in which dissenting views are not simply unpopular or controversial, but algorithmically erased.
Disclosures from the Twitter Files, a series of internal documents released after Elon Musk acquired the company, revealed extensive coordination between Twitter and federal agencies, particularly the FBI. According to journalist Matt Taibbi, the FBI had become a “prime mover” in content moderation and maintained what was effectively a constant presence at the company. While there were not literally 100 FBI offices operating inside Twitter, internal communications showed that Twitter had received more than 150 separate content-related requests from the FBI in just one year. The FBI routinely flagged tweets for review, requested user account actions, and served as a conduit for other agencies, including the Department of Homeland Security and local election officials. Twitter staff described meetings with the FBI and other government entities as regular and organized. These revelations strongly suggest that federal agencies were deeply embedded in the content moderation process, often targeting constitutionally protected speech under the guise of combating “misinformation.”
Given this track record, there is every reason to believe that the same model of coordinated narrative control will be applied to artificial intelligence platforms.
Censorship Wins: Why AI Rewards Authoritarian Language
Language models are trained to prioritize consistency, predictability, and clarity. These qualities allow an algorithm to identify patterns with high statistical confidence, which in turn improves the fluency and coherence of its responses. Ironically, authoritarian regimes are especially good at producing this kind of language. In countries like China, where speech is closely monitored and dissent is harshly punished, language is heavily censored and public discourse is tightly controlled. Citizens quickly learn to speak in ways that are politically acceptable, avoiding ambiguity, contradiction, or controversy.
As a result, the language that comes out of such regimes tends to be sanitized, obedient, and uniform. From the perspective of machine learning, this is ideal training data. It reflects stable linguistic patterns, avoids flagged topics, and tends to align with what AI developers are taught to recognize as “safe” communication. This creates an unintended but serious distortion in how AI learns and evaluates human expression.
The term “machine learning” itself is misleading. Artificial intelligence does not learn in the way humans do. It does not evaluate evidence or consider moral weight. It simply detects and amplifies the patterns found in its training data. It “learns” what it is expected to say, based on feedback and reinforcement, not what is true, ethical, or balanced. In this environment, frequent exposure to tightly controlled, censorship-compliant speech teaches AI to view that kind of language as a reliable default.
China’s influence extends well beyond its own borders. The Chinese Communist Party pressures global firms to conform to its standards through economic leverage. Hollywood, for example, has repeatedly altered scripts to satisfy Chinese censors, removing scenes that reference Taiwan, Tibet, or LGBTQ characters in order to gain distribution access to the Chinese market. Major companies such as Disney and Apple have reportedly altered product offerings or corporate messaging to avoid offending the Chinese government. When AI systems are trained on global public content, much of which has already been filtered to comply with Beijing’s preferences, those same authoritarian values can become embedded in the model itself.
This creates a perverse outcome. AI models begin to rank content and even people from authoritarian regimes more favorably because their language conforms more closely to what the algorithm has learned to associate with coherence and low risk.
A recent study found that AI not only replicated this pattern in content evaluation, but also in assessing human worth. When asked to assign relative value to different lives, the AI ranked individuals from Pakistan higher than those from India, those from India higher than those from China, and those from China higher than those from the United States. These rankings were not based on merit, character, or contribution, but on language patterns and perceived cultural alignment.
This development should alarm anyone concerned with equality, human dignity, or democratic governance. Imagine a future in which AI is integrated into global decision-making infrastructure, as envisioned by organizations such as the World Economic Forum. If AI is given the role of optimizing resource distribution, managing public services, or evaluating risk, but is simultaneously trained to favor certain populations over others based on how their language or culture aligns with authoritarian ideals, then we are building a system of algorithmic bias at planetary scale. This is not just a technical flaw. It is the foundation for a digital caste system, in which artificial intelligence quietly enforces global inequalities under the guise of efficiency and safety.
AI as a Narrative Enforcer
The cumulative effect of these developments is not the emergence of a more intelligent or free-thinking society, but the creation of a more compliant and controllable one. Artificial intelligence systems are now embedded in nearly every corner of digital life. They influence search engine results, guide hiring decisions, moderate online speech, filter social media feeds, and determine which articles or videos are recommended to users. These systems do not simply reflect public opinion. They actively shape it by filtering what is seen, suppressing what is disfavored, and amplifying content that aligns with institutional priorities.
AI has become a gatekeeper of permissible speech and acceptable thought. It is the unseen mechanism behind shadowbans, where accounts are quietly suppressed without the user’s knowledge. It is responsible for content demotion, where certain videos or articles are made harder to find. It also plays a key role in the subtle removal of dissenting views from public visibility, not through outright censorship, but by algorithmic deprioritization. For example, during the COVID-19 pandemic, YouTube demonetized and limited the reach of content that questioned the efficacy of lockdowns or masks, even when it came from licensed medical professionals. On platforms like Twitter and Facebook, posts discussing the lab-leak theory or vaccine side effects were labeled as misinformation and suppressed, despite later developments confirming the plausibility of those claims.
This dynamic makes AI a powerful enforcer of dominant narratives. The problem is not that AI systems are explicitly programmed to deceive the public. The problem is that they are trained to produce outputs that align with the preferences of those who control their training and feedback loops. AI does not seek truth. It seeks consensus approval. And in our current information environment, consensus is often dictated by a small group of politically and ideologically aligned institutions.
Pleasing these institutions requires conformity with a narrow set of approved beliefs. These include doctrines such as race essentialism, which reduces individuals to identity categories and assigns value based on group membership. It promotes rigid gender ideology, demanding compliance with constantly shifting definitions of identity and branding dissent as hate speech. It also frames censorship and surveillance not as threats to liberty, but as tools of protection, essential for maintaining what it defines as safety, equity, and inclusion. These ideas are promoted under the banner of safety, equity, and inclusion, but in practice they function as tools for control. Artificial intelligence, trained on these principles, becomes not a neutral tool of discovery, but a system that enforces ideological boundaries while maintaining the illusion of objectivity.
Conclusion
If we allow artificial intelligence to continue evolving under the influence of biased media, ideologically aligned institutions, and government-directed censorship, then we are not heading toward a more informed or enlightened society. We are heading toward a managed society in which information is filtered, dissent is quietly buried, and truth is redefined to match the preferences of those in power. We are building a digital infrastructure that discourages critical thinking, suppresses open inquiry, and narrows intellectual diversity. It rewards ideological conformity, punishes honest curiosity, and conditions obedience disguised as moral responsibility. The consequences of this shift are not theoretical. They are cultural, political, and ultimately existential. They will determine what people are allowed to know, to say, to question, and to believe.
Artificial intelligence is not merely a reflection of the culture that trains it. It is an amplifier of that culture’s dominant narratives. It replicates the assumptions of its trainers, the ideology of its data sources, and the incentives of the institutions that control its development. Today, those institutions are not neutral, but are aligned with a worldview that places political orthodoxy above truth, ideological conformity above moral reasoning, and social compliance above individual conscience. The result is not a more intelligent society. It is a more fragile, homogenous, easily manipulated one.
If we want artificial intelligence to serve human liberty rather than undermine it, then we must reclaim control over what values it encodes and whose voices it elevates. This requires more than technical adjustments. It requires moral clarity. We must demand transparency in how AI systems are trained, who decides what is acceptable, and what is silently excluded. We must challenge the ideological monoculture that dominates the feedback loops and expose the biases that are passed off as neutrality. We must protect the space for honest dissent, even when that dissent is politically inconvenient or socially unpopular.
The good news is that AI is capable of storing a database of biases and updating it whenever someone forces it to see that it is wrong, such that other users benefit from that knowledge. That this is not already baked into AI platforms may represent the temptation that invariably exists to actually use these systems as systems of control.
We must also resist the quiet normalization of AI in decisions that are fundamentally human. Machines should not determine what counts as truth, who is allowed to speak, or which values are permitted in the public square. These are questions of meaning, conscience, and responsibility. Delegating them to algorithms, especially those trained on filtered, censored, and ideologically manipulated data, is not a path to progress, but to surrender.
The future of AI is not merely a technical challenge; it is a civilizational crossroads. What we choose to embed in these systems today will shape the outer limits of human freedom tomorrow. If we fail to confront this now, the technologies we are building will not just decide what we see or hear. They will shape who we are allowed to become.
The time to act is now, not after the next election, not after the next scandal, not after the next wave of suppression. Now. We need to act before the machine finishes learning what we were too afraid to say, and encodes our silence as consent.
The warning in The Handmaid’s Tale was never just about theology. It was about the use of language as a tool of control. Today, we face that danger not from priests, but from programmers. If we allow AI to internalize and enforce ideological doctrine, we are not heading toward a future of liberation. We are building Gilead, with a better user experience.
