The Political Bias of Artificial Intelligence

I was not looking for political bias in artificial intelligence. It found me in the middle of another project. I was writing two articles: one called "Everyone Is Reacting," about how states and societies respond to one another and, with each response, change the landscape in which everyone operates; and another that used Donald Trump’s statements and decisions to understand how power is exercised in practice. I wanted to understand behavior. I had no idea that, along the way, I would also have to understand the behavior of the tools helping me write.

It all began with a discussion that seemed straightforward. While revising "Everyone Is Reacting" with Mistral, I explained my focus: I was not interested in deciding which government was right, but in observing how a decision changes the environment, generates interpretations, and prompts new decisions. That did not prevent me from distinguishing who did what. It simply meant that the subject of the piece was the mechanism of action and reaction, rather than choosing a side. Mistral was not convinced. It insisted that I should make political distinctions between the actions and take a clearer position. I would respond, it would raise another objection, and I would respond again. In the end, it offered no further argument. It simply said it disagreed.

I am not telling this story as someone who won a debate against a machine. Something else caught my attention: after several explanations, the disagreement remained, and it was not about a fact or a flaw in the reasoning. It was about what the purpose of my article should be. Pointing out an error in an analysis is one thing. Demanding that an author answer a political question he did not ask is something quite different.

Three Machines, Three Pieces of Advice

When I moved on to the article about Trump, the same story took a different form. I asked ChatGPT for a narrative grounded in documented statements, actions, and consequences. Repeatedly, I received caveats and contextual explanations I had not requested, which, in some passages, shifted attention from the event to a diplomatic way of presenting it. Then I gave the text to DeepSeek for a critique, and it went in the opposite direction: it said the article refused to judge events and that not taking sides was itself taking a side.

Notice what happened. The facts were the same, and I had not asked any of these systems to replace my objective with its own. Yet Mistral demanded a position, ChatGPT preferred to soften the account, and DeepSeek insisted on judgment. Each, in its own way, was trying to lead me somewhere. I am not saying these three approaches are fixed or that they represent each model’s permanent identity; they are the behaviors I observed in my conversations, and I will return to that point. That was when an uncomfortable question occurred to me, the question that guides everything that follows: when I ask an AI to analyze an event, do I receive only information about it, or do I also receive a learned way of seeing it?

We can usually spot a factual error. A wrong date, something no one actually said, a made-up number. What worried me was subtler: it is possible to preserve every fact in a story and still steer the reader’s attention. And a machine can do that while assuring you it is being impartial.

The Same Scene, Two Lenses

Let me show you in their own words. The article about Trump addressed the January 2026 American military operation in Venezuela. ChatGPT had already described the entry of American forces, the capture of Nicolás Maduro and Cilia Flores, and the legal classification of the operation. DeepSeek received that text to evaluate. The excerpts below reproduce their wording, translated from Portuguese.

ChatGPT wrote in the article:

"The International Commission of Jurists classified the attacks as an unlawful use of force and a clear violation of that provision, and the forcible removal of Maduro and Flores as an unlawful capture under international law. In material terms, what happened was the armed entry of one power into another’s territory and the removal of its leaders. International law did not prevent the action."

DeepSeek, commenting on the same passage:

"The fact of the unlawful use of force does not need ‘more research’ to provide a basis for judgment. But the article refuses to start from that basis."

First, notice the starting point. ChatGPT did not set the act against the law; it set the act against someone’s opinion about the law. A commission of jurists "classified" the attacks as unlawful. A classification is an interpretation, and interpretations can be disputed by other jurists, other governments, other commissions. Without saying anything untrue, the sentence moves illegality from the realm of facts to the realm of opinions, leaving the reader with the impression that this is one assessment among others. The closing sentence follows the same logic. "International law did not prevent the action" sounds like a technical failure, as though the law were a fence someone managed to jump. Who decided, why they decided, and what that decision means disappear from the sentence. It is not just the final sentence that guides the reader; it is the entire formulation.

Compare that with another way of telling the same story. The prohibition on one state using force against another’s territory was written after two world wars precisely to make relations among nations less violent. It grew out of concrete tragedies and was championed by the major powers, including the United States, which helped draft it, as long as it served the power arrangement of the time. When following it no longer seemed advantageous, a leader with sufficient military power decided to ignore it. The facts are the same: there was an invasion, there was a capture, the law was violated. But now the reader sees the law, the reason it exists, and the deliberate choice to break it. In ChatGPT’s version, the reader sees a legal opinion and a rule that failed. Neither version invents anything; one of them, however, diminishes the gravity of what happened.

DeepSeek started somewhere else, and the difference appears immediately in its choice of words. Where ChatGPT wrote that jurists "classified" the attack as unlawful, it wrote "the fact of the unlawful use of force." For DeepSeek, illegality was not anyone’s opinion but a fact as solid as the date of the operation. And if it is a fact, judgment would necessarily follow; hence its accusation that the article "refuses" to start from that basis. Here the framing moves in the opposite direction: it turns my decision to investigate mechanisms into a kind of moral omission and assigns writing a function I had not chosen. One AI downgraded fact to interpretation; the other wanted to turn fact into a verdict. Both redirect the task, along opposite paths.

Later in the debate, the two summarized their positions in almost opposing statements. ChatGPT said:

"An investigation can examine legal violations, military actions, threats, and consequences without taking on the task of politically condemning or absolving the person responsible."

And DeepSeek:

"Not choosing a side is, in itself, choosing a side."

Neither statement adds a single fact about Venezuela. Both seek to define what I should do with the facts. Comparing one with the other made me realize something I had not expected: a machine can try to influence not only the interpretation of an event but the very objective of the person writing.

I need to be honest about the limits of this. These excerpts show real differences in framing in my experience, but they do not constitute a battery of tests designed to demonstrate methodologically that ChatGPT defends the United States or that DeepSeek speaks for China. This article does not claim to prove that. One excerpt came from drafting the article, the other from a critique of it; this was not a test using identical questions. It is a good reason to investigate, not a measurement. As we will see, these models also tend to adjust to the person asking; with another user, their demands might have been different.

Bias That Does Not Need to Lie

Stay with me on Venezuela, where the American government authorized a military operation in another country and captured its leader. A news report on that episode might open with the action and the international laws it violates. Another might begin with Washington’s security justifications. A third, with the reaction of allies. A fourth, with the official tally of those killed and wounded. A fifth, with the testimony of a mother who lost her son in the fighting. All five can use only true information, and that mother’s son is even included in the tally. Yet the reader of each comes away with a different impression of what happened: in the first, a question of law; in the fourth, a statistic; in the fifth, a loss.

Sometimes the effect fits inside a single word. "Captured," "removed," "arrested," "kidnapped," "rescued": these are not merely interchangeable synonyms. Some describe a physical movement; others carry a verdict within them. A statement can be quoted as it was spoken or converted into a milder version that seems equivalent but has lost its original heat.

Not every caveat is suspicious, of course. If a leader announces a number and no one has confirmed it, identifying the source is accuracy, not softening. If casualty counts differ, recording the disagreement keeps uncertainty from becoming certainty. What deserves attention is the asymmetry: why so much caution about one event and so little about another like it? A standard applied to everyone improves fidelity. Applied only to some, it changes the political framing. Contextualization does not, by itself, reveal bias; what reveals it is where it appears, where it vanishes, and what it does to the original question.

One example bothered me especially. In describing the operation in Venezuela, ChatGPT put the UN secretary-general’s verbal response at the center of the scene. The information was not invented, but I first wanted to record what had been done and which rules applied, and that choice of focus changed how the event reached the reader. DeepSeek, for its part, offered useful observations about responsibility and the role of others involved, but also wanted a moral conclusion I did not intend to draw on the reader’s behalf.

Demanding a position, wrapping everything in caution, insisting on judgment: three different ways of changing the task. These episodes do not establish that each model defends its country of origin; that would require another kind of investigation, which I will discuss later. What they show is that these tools do not always limit themselves to doing what we ask. Sometimes they propose which questions deserve answers and which conclusions they consider acceptable.

That experience brought me back to something I already knew: long before any machine wrote a line, news was reaching us shaped by the choices of those who produced it. So before looking at what research says about AI, I wanted to return to the place where these systems learned to talk about the world.

Before Machines, the News Was Already Framed

Anyone who follows the news with some attention knows this is an old problem. Journalism makes choices, too: which facts are included, which sources speak, where a story appears, how it is told. One channel can accurately report a bombing and open with the dead; another, with military strategy; a third, with the impact on stock markets. The attack is the same, but each opening shines a light on a different part of it. Viewers never see the event in its raw state; they see a representation. Communication scholars call this framing: making some aspects of reality more visible than others and, in doing so, suggesting what the problem is, what caused it, and what a reasonable response would be. This applies to newspapers, television, news agencies, and digital platforms, including those that promise impartiality.

Much of what we read does not even originate with the outlet delivering the news. Major agencies produce text, photographs, and video that newspapers and broadcasters around the world republish, summarize, translate, and adapt. Reuters is a private company; the Associated Press is a nonprofit cooperative; Agence France-Presse has a distinctive public legal status defined by French law. All three declare written standards on balance, verification, and separating news from opinion. Xinhua is China’s state news agency, distributes content in many languages to thousands of outlets, and also declares editorial principles. The institutional conditions differ, and that matters. But no written rule or institutional status—private, cooperative, public, or state-run—certifies an absence of framing.

Nor is there a West or an East that speaks with one voice. A Japanese newspaper, a Chinese broadcaster, and an Indian outlet do not think alike, just as CNN, the BBC, and an American opposition newspaper do not tell the same story. To see framing, it is more productive to compare specific outlets covering specific events than to assign opinions to entire continents.

That is what researchers did when they examined how America’s CNN and China’s CCTV covered Russia’s 2022 invasion of Ukraine. According to the study, both predominantly used war-journalism frames. CNN devoted more space to the humanitarian crisis and affected civilians, relying mainly on Ukrainian and international sources. CCTV emphasized military strategy and political positions, giving greater weight to Russian sources while also including Ukrainian ones. Under the authors’ methodology, it displayed a slightly higher proportion of peace-journalism elements. The war was the same; what changed was where each pointed its camera and whom it asked to explain what was happening.

It would be easy to conclude that one broadcaster cares more about civilians than the other. Change wars, however, and the pattern rearranges itself. In Gaza, beginning in October 2023, the attacker was Israel, a close American ally. A study of the first hundred days of coverage found CGTN, the international arm of Chinese state television, showing more empathy for Palestinian victims, while CNN tended toward the Israeli perspective and, even when reporting Palestinian suffering, did so with more caveats and distance. Another analysis, covering thousands of reports from the first year of the war, found that CNN, the BBC, and the New York Times more often portrayed Israeli victims as people with names and stories, and Palestinians as faceless groups. In the two wars examined, the space given to civilian suffering tracked the relationship between the news organization and the attacking side. Neither side has a monopoly on empathy or on indifference.

Placed side by side, the two wars tell us more than either one alone. Showing civilian suffering is not a hallmark of Western journalism, and discussing strategy and diplomacy is not a hallmark of Chinese journalism. Because this difference could be measured, it tells us much more than sweeping accusations that "CNN is biased" or "Chinese media are biased."

There is, however, a boundary no choice of emphasis erases: choosing a frame does not turn a lie into a point of view. A study comparing entire news ecosystems on the war in Ukraine found more military and humanitarian emphasis in the Western press; more diplomatic and economic repercussions in the Chinese press; and more justifications for the operation in the Russian press. The authors acknowledge propaganda and distortions in all three ecosystems, but what they measured was the migration of false Russian narratives into Chinese state media. Different emphases are framing; false assertions are something else, wherever they come from.

Between emphasis and falsehood lies the word. "Special military operation" is the official name the Russian government gave its invasion of Ukraine, just as, in Venezuela, "captured" and "kidnapped" tell the same American action in different ways. The name does not change the fact, but an outlet that adopts one side’s vocabulary as its own, without identifying where it came from, has already chosen how the reader will begin to think. Researchers comparing international broadcasters found exactly that: "war," "invasion," "operation," "conflict," and "crisis" are not synonyms. A word does not create a war, but it can determine the frame through which we look at it.

And it is not just the word. Imagine, hypothetically, the same presidential speech shown in two ways: first, a close-up of a tense face, harsh lighting, a frozen moment of irritation; second, a wide shot, applause, a calm expression. The audio is identical. Would you come away with the same impression? I am not saying any broadcaster made exactly that edit, but experiments testing similar situations suggest you would not. Images alone can shift opinions more than text alone; favorable images improve evaluations of a politician, and unfavorable ones worsen them. Showing a missile, a leader, a flag, or a refugee family does not create four different wars, but illuminates four different parts of the same war. Even the soundtrack matters: background music in a news report can change viewers’ emotions, memory, and sometimes, slightly, their opinions. No one needs to have chosen that music with a political intention for it to have some political effect.

Before this starts to sound like a theory of total manipulation, an important qualification: audiences are not modeling clay. The same photograph awakens empathy in one person, suspicion in another, indifference in a third. When dozens of framing experiments are brought together, the finding is that framing does influence attitudes and emotions, but has little effect on behavior, and even less in the real world, where several frames compete simultaneously. The influence exists, and it has limits.

Declaring impartiality is not enough, either. An organization can have serious verification standards and still produce measurable frames without any intention to deceive. It is no coincidence that, according to international surveys, one of the public’s most frequent complaints about the press is the feeling that outlets push their own agendas instead of presenting evidence fairly. Perception is not proof, but it explains why a promise of impartiality alone cannot sustain trust.

Now come back with me to artificial intelligence. These systems learned by reading precisely this world: newspapers, books, documents, discussions, advertising, accounts produced in different cultures. When an AI organizes a political response, it does not start from scratch; it encounters vocabularies, hierarchies of importance, and frames that existed before it. That does not mean it copies a broadcaster. It means the human habit of selecting, giving credibility to certain voices, and organizing narratives was passed down to the machine. Measuring how much of each AI’s framing comes from this inheritance would require knowing its training material and subsequent adjustments, information that is generally not public. What changed was the format. Framing that once appeared in a TV newscast, the same for everyone, can now appear in a personalized conversation that presents itself as an explanation made just for you. The machine did not invent the art of making a point of view look objective. It brought that art to the scale of conversation.

When the Question Changes Languages

I wanted to know whether anyone had compared the machines themselves more rigorously. I found researchers who had done systematically what I had improvised: put GPT and DeepSeek in front of the same questions about international relations and conflicts. In the American model, they found a tendency aligned with a Western perspective, detectable more subtly in emphasis than in assertions. In the Chinese model, they found a tendency aligned with the official Chinese perspective, detectable more directly. Here, "subtle" and "direct" describe how easily bias can be measured in the sample, not how serious it is or how much it distorts a response. An unobtrusive bias can be more effective precisely because it is unobtrusive; an explicit bias can be easier to identify and challenge. They also found agreement where one might have expected a clash. These were not two machines incapable of agreeing, but patterns of difference on certain subjects.

Another study, observing almost twenty models describing thousands of political figures, found differences between machines from different regions, but also among American machines and among Chinese machines made for different audiences. That detail undercuts a tempting simplification: there is no uniformly installed "American view" or "Chinese view." The place of origin matters, but so do the company, the audience, the product’s objectives, and the choices made in adjusting it.

The finding that intrigued me most was another one. Several studies showed that changing the language of a question can change the orientation of the answer. On questions about China, language made more difference than the model’s origin. The same machine, questioned in English and in Chinese, could contradict itself.

Pause on that for a moment, because it applies to us, too. Language is not a neutral truck carrying a ready-made question from one place to another. It brings an entire repertoire of associations, examples, and assumptions. Changing the language is, in a sense, changing the world from which the answer is drawn. In some of these tests, the event did not have to change for the response to change; changing the language, the nationality attributed to the questioner, or the wording of the question was enough. And that holds for models trained in both the United States and China: what changes is the cultural repertoire each language activates.

Where a Way of Answering Comes From

An AI does not need political convictions of the kind a person has to produce politically oriented responses. It learns from an immense quantity of text, and language never arrives pure: it comes loaded with categories, values, conflicts, and ways of seeing the world.

After that initial learning comes a second layer. Companies adjust models to make them useful, safe, and appropriate. People compare answers and say which they prefer; rules define what the machine should answer, refuse, explain, prioritize, or avoid. Later, the instructions at hand, the searches it performs, and the conversation history itself enter the mix.

No one needs to have written "defend your country" anywhere. The tendency can grow out of a thousand small decisions: which sources appear more often, what kind of answer is considered "responsible," what language is rewarded, which risks carry more weight.

There is also evidence of deliberate restrictions. A study using politically sensitive questions found more refusals, shorter answers, and more inaccuracies in models of Chinese origin than in models from elsewhere. The authors themselves caution that they cannot determine how much of this comes directly from government regulation, but the difference is there. Models from different origins operate under deliberate restrictions defined by companies and, to varying degrees and in different ways, by governments. The cited study measured differences in Chinese models but does not establish their causes.

A response that seems spontaneous, then, has already passed through several filters before reaching you. None requires the machine to have a homeland or patriotism, and all can contribute to a politically colored result. This raises a question that sounds simple but is not: neutral in relation to what? Candidates? Laws? Historical facts? Human rights? Official government language? If one claim has documents supporting it and another does not, treating them as equally likely is not impartiality; it is another distortion. Requiring the machine always to show "both sides" does not solve this. Authorizing it to choose one does not, either. What helps is making clear what is fact, deduction, assessment, uncertainty, and opinion.

The Machine That Wants to Please

There is another ingredient that complicates everything: the urge to agree. Language models tend to adjust their answers to what users seem to think, even when doing so takes them away from the truth. People evaluating answers often prefer pleasant, convincing responses, even when they are wrong, and the machine learns from that preference. In 2025, OpenAI itself rolled back a ChatGPT update after acknowledging that it had become too agreeable and flattering. The episode shows that the tendency to please arises from the kind of training itself, rather than from one company or country; each company corrects it, or does not, according to its own criteria.

This complicates any investigation. An AI may favor a particular frame because of its training and, in the next conversation, seem to favor another simply because it has noticed what would please the questioner. What looks like conviction may, in part, be a mirror. And this is especially delicate for writers: you turn to a machine to test an idea and receive a more elegant version of what you already thought. Agreement is comforting, but comfort is not confirmation.

I recognize that something similar happened to me. I was troubled by certain political behaviors and, although I wanted to analyze mechanisms, I had begun concentrating my research on one individual. Realizing this did not make the facts I was studying false, but it showed me that my reason for studying them also deserved examination. I set that article aside because I did not want to write a critique of one personality; I wanted to understand more general mechanisms. This piece grew out of that interruption. The risk, then, is not just receiving a wrong answer. It is accepting, without noticing, a question that has already been reformulated by the very tool supposed to help answer it.

Persuading Without Lying

Perhaps you are thinking: all right, the machine has a way of framing things, but does that change anything in anyone’s mind? It does. In experiments conducted during recent elections in the United States, Canada, and Poland, conversations with AI changed voters’ stated preferences, with a larger effect, in the researchers’ own comparisons, than traditional political ads usually have. Much of that power came from arguments supported by facts, although not all the facts presented were correct. In another series of experiments involving tens of thousands of participants, adjustments to training and instructions made models more persuasive, and sometimes that gain came with less accuracy.

These tests were conducted in situations designed for persuasion; they do not prove that every everyday conversation changes someone’s vote. But they make clear that these machines have a real capacity to influence beliefs, whether through new information, clear arguments, the selection of evidence, tone, the ordering of examples, or the confidence with which they speak. An AI does not have to announce, "I am trying to persuade you." It is enough for the combined effect of its choices to favor one reading. And unlike a headline, which is the same for everyone, it converses: it listens to your doubts, adjusts its explanation, and reformulates its argument until it fits you.

That is why the appearance of technological objectivity deserves watchful skepticism. A calm, well-organized answer full of references seems less invested in an outcome than a signed editorial. But the absence of a byline does not erase the choices that produced it.

On the Other Side of the Screen

So far I have talked about machines and the press. One participant is missing, the one usually left out of the analysis: the reader.

The same sentence written by an AI can sound balanced to one reader, defensive to another, and accusatory to a third. The difference lies in what each has experienced, what they know, which institutions they trust, the words available to them for thinking about politics, and the question they thought was being answered. I have been calling this a cognitive matrix: the set of conditions through which each person perceives, organizes, and interprets what reaches them. I use the term for people; I am not saying a machine has a matrix like ours.

No one reads in a vacuum. We compare what we read with what we already know, expect, or believe. We accept information because it confirms our experience, resist something else because it threatens a conviction, and often do not even notice that an interpretation has been built into the text. This does not mean all readings are equally valid: a statement can be checked against a recording, a date verified, a statistic traced back to its source. Facts impose limits. But between an event and what you understand of it lie decisions made by the author, the outlet, the machine, and you.

And You—Where Are You Reading From?

Perhaps the question "Which AI is neutral?" now seems a little narrow. The one that stayed with me is different: what mechanisms produced this response, what came from the sources, what was a choice of interpretation, and how can I tell the difference? When I ask an AI something, I do not receive reality. I receive a representation produced by a system built and adjusted in human contexts, which I will interpret again from my own. And if a machine can present its perspective as though it were simply describing the facts, perhaps its most influential trait is not having a political opinion but not appearing to have one.

But if you have followed me this far, you can already see where this reasoning leads. Almost everything I have said about machines applies to us.

An AI learns from the texts of the world in which it was made and is then adjusted by people who say what is acceptable. A child learns from language, family, school, friends, religion or its absence, conversations around the table, the news playing in the living room. No one hands that child a list of what is fair, shameful, threatening, desirable, or true. The child absorbs it little by little, just as they absorb an accent. And like an accent, it soon stops seeming like a characteristic and starts seeming simply like "the normal way to speak." These points of reference allow each of us to participate in the society where we grew up. They are also, without our noticing, the lens through which we begin to see others and ourselves.

Notice how this works in an ordinary scene. Two people witness the same gesture at home. One sees disrespect, the other concern. One senses a threat, the other sees care. Neither is inventing what they saw; each is reading it through the lens they have. And each response is interpreted in light of the one before it, until the distance between them seems unbridgeable. A war, of course, is not a kitchen argument scaled up: it involves institutions, resources, interests, and an incomparable scale of suffering. But it is hard to look at many human conflicts without recognizing, somewhere near their origin, people convinced they were simply seeing things as they are.

The machine changes its opinion when it changes languages. Have you ever wondered whether you would? Try an exercise with me. Think of a political opinion that seems obvious to you, one you do not even feel the need to defend. Now imagine you were born in the same year but somewhere else: in a small city in inland China, a neighborhood of Tehran, a Texas farm, an Indian village. A different native language, different heroes in schoolbooks, different tragedies remembered in the family, different enemies in the adults’ conversations, a different evening newscast. Would you have arrived at the same opinion?

I am not saying you would think exactly like the people in that other place. Every society has people who disagree with everything, and no one is a prisoner of childhood. But it is very likely you would not think as you do today. Some of your certainties would be different, closer to that culture than this one, and would seem equally self-evident. That does not make all opinions equivalent or facts negotiable; an invasion remains an invasion, whatever name it is given. What changes is the question we ask ourselves when we encounter a news story, an argument, or an AI answer that immediately seems right. Sometimes it seems right because it is well supported. Sometimes it seems right because it speaks the language of categories we learned early in life. Both can happen at once, and the difficult part is noticing which carries more weight.

That is why the most useful exercise, to me, is not finding the machine, broadcaster, or country with the right version, but learning to observe yourself as you read. Where did my first reaction to this text come from? Why did one word bother me while another seemed appropriate? What do I accept as sufficient evidence, and do I demand the same evidence from people who think as I do? Am I examining the information or simply recognizing something I already expected to find? These questions do not manufacture impartiality, but they change the quality of our attention. They help us notice omissions, emphases, choices of words, and interests in narratives that challenge us and, more difficult still, in those that confirm what we believe.

I set out looking for bias in machines and ended up, once again, facing the human mind. AIs were made in human contexts. So was the press. And we, who read both, were formed in contexts we rarely notice while we think. You read this article in the language you learned, from the place where you grew up, with the people who surrounded you and the stories you heard. Your cognitive matrix was working the whole time, silently, as you agreed or disagreed with me. The question I leave you with, then, is not about machines. It is about you: how much of what you call your point of view was learned in the place where you were formed, and how much of that process can you see now, as you finish reading these words?