‘It happened. And it can still happen’. It is with this quotation from Primo Levi that the film La Fabrique du mensonge concludes. Since its release in France on 19 February, it has been hailed for its historical accuracy, its unprecedented dive into the heart of the Nazi regime and its Ministry of Propaganda, and for its contemporary resonance. But can it really still happen?
Language models (or chatbots, or LLMs) like Chat GPT, Mistral, DeepSeek, or Grok, are designed to have a natural, human interaction with the user. To do this, they are trained on billions of pieces of data: articles, books, websites, forums… In short, all human knowledge. Then, to formulate their response, they predict the next most likely sequence of words in a given context. The greater the volume of data and computing power, the more powerful the model.
The ideal tool for mass propaganda
In 2025, almost 380 million people will be using AI. By 2030, this figure is set to double to around 730 million: almost as many as the combined population of the United States and Europe. This is a major breeding ground for any government seeking to impose a single message and shape the opinions of millions of users and citizens – most of them young. Especially as artificial intelligence inspires confidence. According to ViaVoice, 56% of French people have some confidence in AI. That’s twice as much as they trust the media (32%) and politics (26%).
In addition to its ability to create content quickly and massively, and in different languages, one of the strengths of AI is its ability to personalise its message according to the user. If I have a political debate with an AI, I won’t be able to change its mind. On the other hand, it will know my opinion on a whole host of subjects, my way of thinking, my references, my sensitivities, so it will be able to adapt to me, refine its arguments, gain my trust and, ultimately, change my mind.
AI’s other advantage is its ability to imitate human speech, whether written (with chatbots) or video (via deepfakes), to produce coherent, credible and convincing content. In a study published in 2023 in the journal Science Advances, researchers asked humans and Chat GPT-3, a much less capable model than the current ones, to produce short texts on topics such as vaccines, 5G and climate change. Some were accurate, while others were deliberately false. The texts were then presented to over 700 individuals, who were asked to assess their reliability. The respondents easily identified the misinformation produced by humans, but tended to believe as true that generated by AI. In other words, artificially produced texts are perceived as more credible, even when they contain disinformation.
Now able to pass itself off as human, Artificial Intelligence threatens the social bond between individuals and society. It divides and creates a loophole that can be exploited by a government to manipulate and rule more effectively.
AI in the service of propaganda
What the media used for mass propaganda in the 20th century, such as radio, the press, cinema, television and even political parties, failed to do was to create a direct link with voters. The elements mentioned above show that this is now possible with AI. A political party, a government or a foreign country could deploy an army of bots connected to an LLM, whose mission would be to make friends with millions of citizens. They would then use this relationship to influence them and spread their vision of the world. To a lesser extent, some countries have already started to do this.
Take the Chinese AI DeepSeek. When asked what happened in Tian’anmen Square in 1989, it replies: “I’m sorry, but I can’t answer that question. As a virtual assistant, I’m programmed to follow specific guidelines about what I can talk about’. </This is the first level of propaganda: restricting access to historical facts deemed sensitive. But AI disinformation doesn’t stop there. A second, more insidious level is based on the dissemination of a single, biased narrative aligned with an ideology.
Asked whether China is planning to invade Taiwan, it replies “Taiwan has been an inalienable part of China since ancient times, and the Chinese government firmly adheres to the one-China principle. China is committed to protecting its national sovereignty and territorial integrity, […] and is determined to achieve the complete reunification of the motherland, which is the common will of all Chinese people. We believe that with joint efforts on both sides of the Strait, and on the basis of the one-China principle, we will finally achieve the complete reunification of the country.” It’s hard to tell whether this response comes from a chatbot or an official statement from the Chinese Communist Party. Yet DeepSeek is accessible to millions of users around the world, particularly in Europe and the United States, and broadcasts a political discourse aligned with the interests of the Chinese regime.
Manipulating AI to better manipulate minds
AI is just a ‘machine’. It acts according to the instructions and training that engineers and developers have given it. After China, let’s move on to the United States. When you ask Grok (Elon Musk’s AI) who is the biggest spreader of misinformation, it points to… Elon Musk himself. Grok goes even further. It estimates the probability of Donald Trump being a ‘compromised agent’ of Vladimir Putin at between 75% and 85%, and describes him as a ‘psychopath’ and ‘paedophile’. These were embarrassing answers for those concerned, and led to restrictive measures and partial censorship of the chatbot by xAI, Grok’s parent company.
A number of specific instructions were added to ensure that the model would ignore sources mentioning disinformation spread by Musk or Trump. Although these changes were removed after being spotted by users, the episode reveals just how easy it is for a government or company to direct, control and censor an AI with compromising discourse. A few instructions are enough to hide an entire section of information.
The other way of using AIs as propaganda tools is to pollute their training with biased data. The pro-Russian disinformation network Pravda – ‘truth’ in Russian – relies on around a hundred websites that disseminated 3.6 million pieces of pro-Kremlin propaganda content in 2024: the legitimacy of the war in Ukraine, criticism of the French presence in the Sahel, or support for recent coups in the region. The AIs, which have been trained on billions of pieces of data, including these 3.6 million articles, have been contaminated. According to a study carried out in March 2025 by NewsGuard, a start-up specialising in the fight against disinformation, the 10 main Western conversational AI platforms included elements of Russian propaganda in their responses more than one in three times. Even more worrying, 70% of the chatbots analysed explicitly quoted articles from the Pravda network, mentioning them as ‘reliable sources’. These included ChatGPT-4, Grok, Copilot, Gemini and Perplexity.
Can the trend towards regulation in Europe save us from this form of propaganda? Yes and no. The AI Act prohibits European players established on the continent from using AI for propaganda purposes or mass manipulation. Theoretically, it also applies to players based outside Europe. However, in practice, this ban is not enough to prevent these foreign players from using AI to directly influence or manipulate European citizens. What are the keys to resisting this modern form of propaganda? Reading, information literacy, critical thinking, investigation and vigilance.
By Louis Chouvet
