Opinion – The Most Personal Propaganda Tool Ever Built
Jude Althagafi
An AI chatbot can ingrain itself into every part of a human being's cognitive output, and has the ability to frame a political choice or idea. It helps write an email, it can help personally advise, write responses or information for school, work, or home. The influence comes through a useful answer that is written for one person at the moment that person is still coming up with the view itself, and that exact placement before the thinking happens is what gives it power. In the data used to train GPT-3, 93 of every 100 words were English and Meta's Llama 2 was 89.70 percent English. Other languages, spoken by billions of people, were rounding errors. Even though, any model built and trained on that ratio can and will still answer fluently in whichever language the user brings to it. In the study Faux Polyglot, researchers found that large language models tend to favor information in the language of the question, and when that information runs thin, they reach for high-resource languages, usually English, then translate the result. The language adapts to the user, but the bias and reasoning in the answer itself more often than not comes from one end of the political and social spectrum.
E-International Relations is free to read on our website and on substack. We rely on reader support to make this possible.
Language preference is only half the story. Some of the resulting direction comes from the imbalance of the data itself, and it is embedded in the system before anyone trains and fixes the data, since English-language sources lean a particular way politically as much as they lean a particular way linguistically. The rest of the influence comes from the human decisions on top that dictate when a model refuses, what it calls harmful, and which groups receive protection. The lean is not an accident because specific inputs and specific decisions produced it, which means someone, somewhere, engineered it. Two different studies found that ChatGPT-4 and Claude were on the liberal side, Perplexity on the conservative side, and Gemini closer to the center. The exact position differs from model to model and study to study, but no analysis has found the center.
This is where a lean becomes something more dangerous, and it is worth saying plainly. This is propaganda, by the definitions of the people who studied it closest. A lean you can see is just a viewpoint, and you weigh it as one, but Jacques Ellul described propaganda as a deeper process, the slow adjustment by which a system's judgments begin to feel ordinary. Jason Stanley, in How Propaganda Works, argues that the strongest propaganda presents a partisan frame as the standard of reason itself, so one position has the face of common sense while the alternative struggles to sound reasonable at all.
The political orientation is stable across models and languages, and a pattern that moved at random would be considered a design quirk instead. The model delivers these messages as reasonable answers instead of one argument among several, since it has already ranked which point sounds balanced, which sounds extreme, which deserves caution, and which deserves encouragement, and it brings the result to the user as plain good sense. And the same frame plays out across billions of exchanges, in the user's own language, inside the user's own task, until it feels less like someone else's view and more like something the user simply came up with. Researchers found that when GPT-4 had a few personal details about its opponent, it won 64 percent of the debates it would otherwise have tied, with the odds of agreement rising by roughly 81 percent over the human baseline. In this case, the illusory truth effect shows that repeated claims gain believability even against prior knowledge, and a personalized answer delivered at this scale carries an intimacy and a timing no broadcaster ever had.
A handful of American companies now hold this kind of influence. OpenAI, Google, Anthropic, Meta, and Microsoft are among them. ChatGPT has passed a billion monthly users in May 2026. Google placed Gemini above 750 million monthly users in late-2025 earnings. Meta's apps reached 3.58 billion daily users in December 2025. A small group of companies is setting, for users far outside its borders and society, how AI treats authority, religion, gender, family, speech, safety, and harm. Give these companies the benefit of the doubt, and they may sincerely believe they are exporting safety and intelligence rather than ideology. The system still exports a view of society along with the service it provides. One society's idea of neutrality can become part of ordinary life in another, and it does it so quietly that it makes it extremely difficult to notice. The model gives the user thoughts through a worldview the user never chose, but the user absorbs it anyway just by using the tool. No one taught it to them on purpose, but a generation raised on these systems may learn to argue and object and apologize and ask questions through this lens, and it might come to feel like their own over time.
This kind of influence succeeds precisely because it hides inside the most everyday act there is, the simple act of asking for help. People knew when they were watching Hollywood movies, studying at a foreign university, or reading a textbook printed abroad. The source had a name and a point of entry, and you could account for it but a chatbot has neither. That is the new soft power, and it is harder to see than culture or media because it speaks in the user's own language and has a personalized relationship, so it sounds like the user's own thought. This is what should worry anyone watching from a country that did not build these systems. Other forms of propaganda needed a visible medium, something a government could point to and regulate. This one needs nothing but a conversation, and it adjusts itself to the one person having it. A message you can name is a message you can resist, but a message that comes as your own thought, tailored to you, repeated every time you ask for help, is not something a society can simply choose to tune out. That kind of soft power should not be allowed to operate without anyone watching.
Governments outside the handful of countries building these systems are adopting them anyway, in their ministries and classrooms, because the tools are useful and the alternative is falling behind. But adoption should not mean surrendering the definition of a reasonable answer. Most governments would read a foreign textbook line by line before it reached a classroom. AI chatbots deserve the same reading. Building local language models, training on local data, and keeping the infrastructure on home soil are how a country keeps that judgment in its own hands.
Jude Althagafi is lead editor for knowledge hub and intelligence research at Takamol Holding, the operational arm of Saudi Arabia’s Ministry of Human Resources and Social Development, and strategic advisor to its Executive Office. She works at the intersection of AI, labor markets, and national strategy, and holds an MA in International Relations from Johns Hopkins SAIS, where her graduate research focused on AI and national security. Her work includes analysis for the Global Labor Market Conference.