AI disinformation refers to the use of generative artificial intelligence to produce and spread, at scale and at low cost, false or misleading content: faked text, images, voices, and videos. AI does not create a new intent to deceive — it amplifies an old practice by lowering the cost of fabricating the false and multiplying its speed of diffusion. Documented examples exist; so do countermeasures, and the most effective are preventive. This article sorts the matter out, without panic and without unproven accusation.
Definition. Disinformation differs from mere error (misinformation, false information spread without intent to harm) by its deliberate character: knowingly propagating false or misleading content to produce an effect. AI is here a tool of amplification, not an autonomous actor: it executes, it does not decide.
A large language model produces what is plausible, not what is true. It generates with the same confidence an accurate piece of information and an invented one — a quotation that does not exist, a fabricated reference, an erroneous fact phrased impeccably. This phenomenon, called "hallucination" or "confabulation," is not an accidental bug: it is a direct consequence of the training objective, which maximises plausibility and not truth. A false text can therefore appear perfectly credible.
To this plausibility is added scale. Where classic propaganda required writers, studios, and time, generative AI makes it possible to produce thousands of variants of a message in a few minutes, in several languages, at near-zero cost. The factor that changes is not the nature of the deception, but its speed, its volume, and the difficulty of attributing it.
Deepfakes — faked synthetic videos, images, and voices — are the most visible example. They allow identity theft, the fabrication of false statements attributed to public figures, and the creation of "proof" that is none. Their danger lies less in technical perfection than in a second-order effect: as the public learns that anything can be faked, doubt extends to authentic content too. A real video can then be swept aside with "it's probably a deepfake." The countermeasure therefore consists not only in detecting the false, but in knowing how to authenticate the true.
Two dated and documented cases illustrate the technique. In January 2024, two days before the New Hampshire Democratic primary, thousands of voters received an automated call reproducing the AI-cloned voice of President Joe Biden, urging them not to vote; in September 2024, the Federal Communications Commission imposed a $6 million fine on political consultant Steve Kramer, the originator of this campaign. In September 2023, two days before the Slovak parliamentary elections, a fake audio recording attributed to the leader of the Progressive Slovakia party, Michal Šimečka, and to journalist Monika Tódová (Denník N) a conversation about electoral fraud; AFP's fact-checking service noted signs of AI manipulation, and both individuals immediately denounced it as fake. In both cases, it is attribution — who is really speaking — that is at stake.
Content farms automatically generate quantities of articles, sites, and posts to occupy space, capture advertising, or push a narrative. Generative AI radically lowers their cost: an entire site of fake "news" can be fed without a human writer. The aim is not always to make people believe a precise thesis — it is often to drown reliable information under the mass, until the very act of distinguishing the serious from the bogus becomes exhausting.
This use is now documented by the platforms themselves. In October 2024, OpenAI reported that since the start of the year it had disrupted more than twenty deceptive operations and networks that had attempted to use its models, notably to generate articles and social-media content destined for influence campaigns. In the first quarter of 2024, Meta for its part dismantled six new coordinated influence operations relying on AI-generated content (fake video news-readers, text, synthetic profile photos). The 2024 report of the European External Action Service (EEAS) on Foreign Information Manipulation and Interference (FIMI) likewise notes a growing use of AI to facilitate the production and automate the distribution of content.
In 2016, researchers Christopher Paul and Miriam Matthews published, for the RAND Corporation, The Russian "Firehose of Falsehood" Propaganda Model. In it they describe a propaganda model characterised by high volume, multichannel diffusion, continuous speed and repetition, and an absence of commitment to truth or even to consistency. The image of the "firehose of falsehood" designates a saturation: drowning the public space under so many contradictory assertions that distinguishing the true becomes exhausting.
This model exploits documented cognitive biases: the illusory truth effect (a repeated assertion seems more true), the weight of the first message received, which "anchors" perception, and the multiplicity of sources — even fake ones — which gives the illusion of independent confirmation. Generative AI is an ideal accelerator of this model: it supplies volume and variety on demand. Its primary aim is not always to convince, but to instil generalised doubt — "we can no longer know who is telling the truth" —, a state of mind that disengages the citizen.
Three questions, which ambient discourse often conflates, must be distinguished: did a disinformation campaign exist? (often, yes, and it is documented); did it aim to influence? (yes, by definition); did it succeed, and to what extent? (most often, indeterminate). Measuring the real effect of an operation on opinions, and even more on behaviours, is extraordinarily difficult, and serious research remains cautious. To assert that a campaign "swung" an election is most often an undemonstrated hypothesis. To acknowledge AI amplification does not, therefore, authorise attributing to it unlimited power over minds.
Good news: the most effective defence does not consist in refuting each falsehood one by one — an endless task, which can even reinforce the familiarity of the false by repeating it. It consists in strengthening, upstream, the capacities for discernment, and in spreading reliable information first. Three levers are documented.
Prebunking (or inoculation). Inoculation theory, formulated by psychologist William McGuire in the 1960s, rests on a medical analogy: exposing a person to a weakened form of a misleading argument, together with its refutation, strengthens their future resistance. The work of Sander van der Linden and Jon Roozenbeek, at Cambridge, tested this principle at scale. A study published in Science Advances in 2022, gathering seven experiments and nearly 30,000 participants, showed that short inoculation videos improved the ability to spot manipulation techniques; conducted with Jigsaw (Google), it notably exposed around 5.4 million internet users on YouTube, nearly a million watching the video for at least 30 seconds. Result: exposing people in advance to manipulation techniques (emotional language, false dichotomies, incoherence, scapegoating, ad hominem attacks) then reduces their vulnerability. One does not vaccinate against a precise lie, but against a technique.
Media literacy. Media and information education completes the apparatus: learning to verify sources, to identify the intent of a message, to distinguish fact, opinion, and propaganda. Several evaluations show its benefits on the ability to recognise misleading content. It is not an infallible countermeasure, but a skill that can be taught and reinforced.
None of these defences offers total immunity, and to claim so would already be a form of disinformation. The effect of prebunking declines over time and calls for "boosters"; it protects better against techniques that are explicitly presented than against novel devices. Media literacy turns no one into an infallible judge of truth. These tools reduce a vulnerability; they do not annul it. To acknowledge this does not weaken the defence — it makes it honest, and therefore durable.
Will AI make disinformation impossible to fight? No. It lowers the cost and increases the volume of the false, but the psychological mechanisms it exploits (illusory truth effect, anchoring, fatigue of doubt) are known, and the countermeasures — prebunking, source verification, literacy — work on these mechanisms, not on each piece of content. The terrain hardens; it does not become hopeless.
How to recognise a deepfake? No technical sign is reliable for certain, and detection with the naked eye is becoming illusory. The most robust reflex is not to scrutinise the pixel, but to verify the source: who is publishing this content, is it confirmed by reliable and independent media, does a traceable original version exist? One authenticates the true rather than chasing the false.
Should every false piece of information be refuted? Not one by one: it is endless, and the repetition of a falsehood can make it more familiar, and therefore more credible. Better to strengthen discernment upstream (prebunking, education) and spread reliable information first. Prevention prevails over the belated denial.
To understand how these mechanisms fit into contemporary military doctrines — from NATO's concept of cognitive warfare to the saturation model, all the way to cognitive-defence countermeasures —, see our investigation into cognitive warfare: the brain as a battlefield.
To go further, the full investigation Cognitive Warfare — Doctrine 2026 brings together the real sources (NATO report, RAND's "firehose of falsehood" model, McGuire's inoculation theory, the prebunking of van der Linden and Roozenbeek) and offers six concrete markers of cognitive defence.
No. It lowers the cost and increases the volume of the false, but the psychological mechanisms it exploits (illusory truth effect, anchoring, fatigue of doubt) are known, and the countermeasures — prebunking, source verification, literacy — work on these mechanisms, not on each piece of content. The terrain hardens; it does not become hopeless.
No technical sign is reliable for certain, and detection with the naked eye is becoming illusory. The most robust reflex is not to scrutinise the pixel, but to verify the source: who is publishing this content, is it confirmed by reliable and independent media, does a traceable original version exist? One authenticates the true rather than chasing the false.
Not one by one: it is endless, and the repetition of a falsehood can make it more familiar, and therefore more credible. Better to strengthen discernment upstream (prebunking, education) and spread reliable information first. Prevention prevails over the belated denial.
Dossier : Guerre cognitive & neuro-armes
MINI10: La Guerre Cognitive — Doctrine 2026
Accueil · Collection · Journal · Dossiers · Sources