top of page
HENSEI-LOGO-V2.png

AI and Disinformation: How Algorithms Really Shape Public Opinion, Beyond the Conventional Wisdom

Writer: Stéphane Guy
Stéphane Guy
Nov 9
12 min read

How does AI shape public opinion? Through three channels: it sorts what you see (the recommendation systems behind social feeds), it rephrases what you learn (chatbots and answer engines) and, more recently, it argues back (conversational persuasion). The three don’t carry equal weight. The strongest studies suggest the first, the news feed, matters less than we’ve been repeating for a decade. The second is underrated. The third works mainly because it stacks facts, true or false, faster than you can check them.


This article separates what is demonstrated, what is contested and what is narrative. It draws on work published between 2018 and 2026 on AI misinformation and disinformation, including when that work contradicts a popular idea.


Des magazines sur un présentoir de kiosque
Photo by Markus Spiske on Unsplash

In short


  • The news feed is not the prime culprit. In the experiments run with Meta around the 2020 US presidential election, switching to a chronological feed did not reduce polarization.

  • Power is shifting to the answer. 10% of people surveyed use a chatbot for news every week (Reuters Institute, 2026), while 45% of the assistant answers tested by the EBU and the BBC contained at least one significant issue.

  • An AI persuades by piling up facts. The more a model is optimized to persuade, the less accurate it gets (Science, December 2025).

  • “LLM grooming” remains contested. NewsGuard describes AI contaminated by Russian propaganda; a peer-reviewed study in the Harvard Kennedy School Misinformation Review sees mostly information gaps.

  • The same tool can correct. A dialogue with an AI cut conspiracy beliefs by about 20%, durably. It all depends on what you ask it to optimize.


What hasn’t changed: the engine of rumor is us


In March 2018, three MIT researchers published a study in Science on how true and false stories circulated on Twitter (now X) between 2006 and 2017. Two numbers came out of it: a false story was 70% more likely to be retweeted than a true one, and an accurate story took about six times longer to reach 1,500 people. Strip the bots out of the dataset, and the gap persists. Co-author Sinan Aral put it down to psychology: “False news is more novel, and people are more likely to share novel information”.


That study predates ChatGPT by nearly five years, which is what makes it valuable. It establishes that the appetite for novelty and outrage exists without AI. What generative AI brings is a near-free supply: a plausible fake, in any language, now costs nothing to produce. For fake videos and fake voices, we devoted a full article to deepfakes. We don’t revisit that here. This dossier is about the mechanics of opinion, not the manufacture of fakes.


The news feed: a less guilty suspect than we thought


If you’ve read an article about “filter bubbles” in the past ten years, you know the story. The algorithm shows you what you like, you only see your side, you radicalize. It’s seductive. It has also been put to the test.


In July 2023, four studies in Science and Nature delivered the first results of a collaboration between Meta and 17 outside researchers, centered on the 2020 US presidential election. In one experiment, about 40,000 volunteers took part, and a randomly assigned group was switched to a chronological feed, with no algorithmic ranking.


The result is a trade-off, not a victory. Those users spent far less time on the platforms and saw more sources with ideologically mixed audiences. But they also saw more political content and more sources rated untrustworthy. On polarization, co-author Jennifer Pan summed it up: the chronological feed is “no silver bullet.” Another strand of the program confirmed strong ideological segregation in news exposure on Facebook (NYU CSMaP). The algorithm doesn’t invent the silos; it finds them already there.


Let’s be precise about the limits. The experiment lasted three months, in the middle of a campaign, on people whose habits had already been shaped by years of algorithmic feeds (that’s my hypothesis, not the authors’). And Meta immediately read the results as vindication, which calls for the usual pinch of salt.


What happens when the same algorithm meets a coordinated operation? Romania offered a textbook case. On December 6, 2024, its Constitutional Court annulled the first round of the presidential election, won by Călin Georgescu, after the declassification of intelligence files describing a campaign of thousands of accounts on TikTok and Telegram (Romania Insider). Eleven days later, the European Commission opened formal proceedings against TikTok under the Digital Services Act, the EU’s rulebook for large online platforms, to check whether it had properly limited risks to the integrity of the vote. TikTok disputes the scale of the account activity described in the files (King’s College London), and Georgescu contests the court’s decision. The alleged facts were therefore unproven when those sources were published, and that’s the lesson too: with algorithmic influence, the evidence arrives after the vote.


The two stories don’t contradict each other. An algorithm alone, as tested in 2020, doesn’t explain polarization. An algorithm combined with thousands of coordinated accounts is a different problem, and that’s where regulators are looking. My view: the filter bubble handed us a convenient, abstract culprit. The less glamorous issue is data access. As long as researchers depend on the platforms’ goodwill, every study is a snapshot.


The real shift: from a list of links to a single answer


That’s the old question. The new one plays out in a text box, and answer engines are now reaching even community platforms, as with Reddit Answers. According to the Reuters Institute’s Digital News Report 2026, weekly use of chatbots for news rose from 7% to 10% worldwide in a year. It reaches 17% among 18–24s, against 5% among those 55 and over. Only 20% of respondents trust news delivered by these tools, versus 37% for news in general. Among people who use them, trust climbs to 44%, versus 17% among non-users, notes Amy Ross Arguedas, who wrote the report’s chapter on AI chatbots.


Look closely at that 44%. It doesn’t say chatbots are reliable. It says use and trust go together, and the data can’t tell us which drives which. 


Reliability, though, is precisely the question. Coordinated by the European Broadcasting Union (EBU) and led by the BBC, an October 2025 study had journalists from 22 public-service media organizations in 18 countries assess more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity: 45% had at least one significant issue, and Gemini reached 76%, mostly because of its sourcing. NewsGuard’s August 2025 audit measured a 35% rate of repeating false claims, against 18% a year earlier (NewsGuard). Part of that jump is a trade-off: the previous year’s chatbots often declined to answer, while this year’s answered every prompt. The January 2026 audit stayed above 28% (NewsGuard).


A word on method, because these numbers often circulate badly. NewsGuard’s audits test ten provably false claims. That is not ChatGPT’s error rate on your everyday questions. It measures how vulnerable assistants are when you hand them a news trap, and that trap is exactly the situation we care about: a rumor in circulation. NewsGuard also notes that chatbots treat how often a false claim is repeated as “a signal of consensus and reliability” (July 2025 audit). What gets repeated looks like what’s true.


A large language model (LLM) verifies nothing: it produces the most plausible continuation of a text. Plugged into the web, it adds a second flaw: it repeats what it finds. Part of these failures are hallucinations, claims invented with aplomb. But attributing a sentence to the wrong outlet, or citing a dubious site as a reference, is a sourcing failure, and that is what dragged Gemini down in the EBU study.


Why is that worse than a bad news feed? On a feed, ten posts contradict each other in front of you. With an assistant, you read one coherent paragraph. The disagreement has been erased before it reaches the screen, and so has the reflex to cross-check. An editor-in-chief would say the chatbot has done a newsroom’s job without bearing a newsroom’s accountability.


Poisoning the source: “LLM grooming” or data voids?


In March 2025, NewsGuard published an investigation that made the rounds worldwide. A Moscow-based network dubbed “Pravda,” credited with 3.6 million articles in 2024 (a figure from the American Sunlight Project), had allegedly contaminated Western AI. Tested on 15 false narratives pushed by the network, ten chatbots repeated them 33.55% of the time, and seven cited Pravda articles as sources (NewsGuard). France’s Viginum agency ties the network to TigerWeb, an IT company based in Russian-occupied Crimea. The tactic has a name, “LLM grooming”: flood the web with content so it ends up in AI answers.


In October 2025, a team publishing in the Harvard Kennedy School Misinformation Review reran the exercise (four chatbots, 416 responses) and offered a different reading. Chatbots pick up these sources on topics that reliable media have barely covered: these are “data voids.” In most cases, the chatbots flagged the sources as unverified or disputed, and the analysis, which its authors describe themselves as a “small-scale analysis finds little evidence for this” case, rejects the grooming thesis (HKS Misinformation Review). The authors also fault NewsGuard’s report for not publishing its full prompt set. An opinion piece published by Al Jazeera adds that NewsGuard tested the AIs only on questions tied to the Pravda network: if grooming really worked, its traces would show up elsewhere (Al Jazeera).


I don’t take sides, and you should be wary of anyone who does with confidence. The Misinformation Review study is small. NewsGuard is a private company that also sells services and datasets to AI companies (NewsGuard). That doesn’t disqualify its work; it’s simply public knowledge.


Take a fictional but plausible case. On a Tuesday at 6 p.m., a rumor of a fuel shortage starts on a social network. By 7 p.m., three sites with no newsroom pick it up. At 8 p.m., you ask an assistant: it searches, finds those three pages, no serious source has reacted yet, and it answers with caution but without debunking. Nobody “groomed” the AI. The void simply filled with whoever arrived first. The Misinformation Review authors, by the way, remind us that such voids can be created artificially. Whether AIs are targeted or merely badly served, the door stays open: publish fast, in volume, on topics nobody has covered yet.


And we have to say something about ourselves. We optimize our articles to be cited by AIs, a practice known as GEO (generative engine optimization). The techniques (direct answers, clear structure, explicit sources) are those of plain editorial common sense, and also those of a well-organized propaganda network. What differs is the truth of what you publish and the intent. The line is thin, and it may be the weak point of the whole edifice.


AI : The patient persuader


That leaves the third channel, the most recent: an AI that no longer just informs you but tries to convince you. In early December 2025, Science and Nature published two studies on the subject. The first, led by the Oxford Internet Institute and the UK AI Security Institute, had nearly 77,000 British participants converse with language models on 707 political issues, across 91,000 conversations (OII). What changes minds? Not model size, whose role is modest. The dominant factor is information density: answers packed with fact-checkable claims, which alone account for roughly half of the explainable variation in persuasiveness between models (Oxford). Tailoring arguments to the individual added comparatively little (Science abstract), and the Nature team likewise found that chatbots persuade with facts and evidence rather than sophisticated psychological techniques (Nature). And the bill is steep, as the OII writes: “the more persuasive a model is, the less accurate its information tends to be.”


The second study, in Nature, put chatbots favoring Kamala Harris or Donald Trump in front of 2,306 American participants ahead of the 2024 election, then repeated the experiment during the 2025 Canadian and Polish elections (Cornell Chronicle). Let’s keep a critical distance: participants were paid to take part, and we don’t know whether a hostile voter would agree to debate an AI outside the lab (PsyPost).


What these studies overturn is our picture of manipulation. We feared psychological profiling, the message tailored to your fears. The most effective weapon is more mundane: a patient interlocutor who answers every objection with figures. A pub argument with a library’s memory.

Same mechanics, opposite use: in a study published in Science in September 2024, Thomas Costello, Gordon Pennycook and David Rand had 2,190 people who believed in a conspiracy theory talk with GPT-4 Turbo. Belief fell by about 20%, the effect still held two months later, and the AI did not talk anyone out of conspiracies that actually happened (Science). The effect is weaker among people who distrust AI, David Rand says, but “it still works even for people who strongly distrust AI” (Nieman Lab).


Put the two studies side by side. The mechanism is the same: a model that answers point by point with specifics. The outcome depends on what it’s asked to optimize. Pushed toward persuasion, it sacrifices accuracy. Tasked with correcting, it pushes false beliefs back. A neutral tool doesn’t exist. There is an objective, and it belongs to someone.


What the law changes, and what it leaves out


Since August 2, 2026, Article 50 of the EU AI Act has applied: a chatbot must signal that it is an AI, and anyone publishing deepfakes, or AI-generated text on matters of public interest, must say so. Text carries a notable exemption: no disclosure is required where the content has undergone human review or editorial control and a person or organization holds editorial responsibility for the publication (European Commission, AI Act Service Desk). Only machine-readable marking received a delay, until December 2, 2026 for systems already on the market, under the Digital Omnibus on AI, adopted as Regulation (EU) 2026/1744 and in force since July 27, 2026 (EUR-Lex). On the platform side, the Digital Services Act is the framework behind proceedings like the one opened against TikTok. For a wider look at what the law can and can’t do when AI causes harm, see our piece on AI and legal liability.


The limit is plain to see: knowing a text comes from an AI says nothing about its accuracy. Labeling protects against deception about origin, not against an error on the substance or an effective argument. On my reading, the European text deals with the visible symptom, the fake passing for real, and barely touches the quality of the answers.


Keeping control: five habits that hold up against the data


  1. Click the source. Clicking through from chatbots is rare in the general population: across 27 markets, 4% of all respondents say they always or often do, versus 19% for search (Reuters Institute). Since only 9% of that sample uses chatbots for news, we can’t say how many users actually check. Either way, it’s the highest-return gesture. A figure without a link you can open should be treated as a rumor.

  2. Distrust the first two days of a news story. Data voids fill in as the press catches up with the event. On a story less than forty-eight hours old, prefer direct sources (wire agencies, live coverage) to assistant summaries. It’s a heuristic, not a law.

  3. Ask for the objection. AI persuasion relies on a one-sided accumulation of arguments. Demand “the best objection to this position, with its sources,” then compare.

  4. Check two numbers, not all of them. You don’t have time to verify everything. Pick the two most surprising figures in an answer and trace them to their origin. If one falls, throw out the rest.

  5. Keep two or three outlets you read directly, with no intermediary. A summary is only useful if you know what a good one looks like.


My take: the wrong culprit, the right worry


We spent ten years watching the news feed, and the power is moving into the answer. My position is deliberate: the filter bubble was the wrong trial, and we should stop treating it as a universal explanation. The legitimate worry is delegation, the same one behind the cognitive cost of outsourcing your thinking. When 10% of the people surveyed, and 17% of 18–24s, ask a machine every week what’s happening in the world, we’re no longer measuring an algorithm’s influence on a feed. We’re watching one intermediary replace another, whose editorial desk nobody has audited.


I’m asking for three simple things: clickable sources by default, an error rate published and audited by third parties, and researcher access to the data. Nothing revolutionary. None of it is a given.


I may be wrong. The 2023 and 2025 experiments are snapshots, and this field ages in eighteen months. One question stays open, and I don’t have a clean answer: if an AI can cut false beliefs by 20%, should we let it do so at scale, when nobody asked it to? And who decides what counts as “false”? I’d rather leave the question standing than manufacture a comfortable conclusion.


FAQ


  1. How does AI influence public opinion?

    Through three channels: content sorting (recommendation systems on social media), the rephrasing of news (chatbots and answer engines) and conversational persuasion. According to recent studies, the second and third matter more than people think, because they deliver a single answer, dense with figures and hard to contradict in the moment.


  2. Do algorithms really trap us in filter bubbles?

    Less than is commonly believed. In the experiments run with Meta around the 2020 US presidential election, replacing the algorithm with a chronological feed did not reduce polarization. The studies do confirm strong ideological segregation in news exposure, which the algorithm did not create but does not erase.


  3. Do AI chatbots give false information about the news?

    Sometimes, and often through poor sourcing. An EBU and BBC study found at least one significant issue in 45% of the answers tested. NewsGuard’s audits, which deliberately bait assistants with false claims, measure repeat rates between 28% and 35% depending on the edition. Those figures don’t describe ordinary use, but they show where the risk sits.


  4. Can AI also be used to fight disinformation?

    Yes, under specific conditions. A study published in Science in 2024 showed that a dialogue with GPT-4 Turbo reduced belief in a conspiracy theory by about 20%, with the effect still visible two months later. The result depends on what the AI is asked to optimize: convincing at any cost, or correcting with evidence.


  5. How can you spot AI-generated disinformation?

    No visual or stylistic clue is reliable on its own. Trace it back to the source instead: who is publishing, since when, who else confirms it? Be wary of very recent claims relayed by sites with no identifiable newsroom. And keep in mind that an AI’s writing can be perfectly polished.


  6. What does the EU AI Act say about AI-generated content?

    Article 50 of the AI Act, applicable since August 2, 2026, requires telling users they are interacting with an AI and labeling deepfakes and AI-generated text on matters of public interest, unless the text has been human-reviewed under editorial responsibility. The Digital Omnibus delays only machine-readable marking for systems already on the market, to December 2, 2026. The Digital Services Act separately governs platforms, as in the proceedings opened against TikTok after Romania’s 2024 election. Labeling tells you about a content’s origin, not its accuracy

Comments


Commenting on this post isn't available anymore. Contact the site owner for more info.
bottom of page