Contents6
The debate on AI and disinformation often sets two images against each other: a machine that churns out believable lies without limit, or a tool that can correct false ideas better than a human can. Both images rest on serious studies. But those studies are not equally solid in October 2026, and neither says what AI will do “in general”.
Our position: AI lowers the cost of producing any content, true or false. What protects us is knowing where a piece of content comes from. That is what European law has partly required since 2 August 2026.
What the two key studies show
AI can disinform better than we can. In June 2023, the journal Science Advances published the study “AI model GPT-3 (dis)informs us better than humans”. This pre-registered study involved 697 participants. Its authors conclude:
GPT-3 is a double-edge sword: In comparison with humans, it can produce accurate information that is easier to understand, but it can also produce more compelling disinformation.
The study also shows that participants could not tell the messages generated by GPT-3 from those written by real Twitter users.
Limitation: the study is from the GPT-3 era and covers short messages in the Twitter format. It does not measure an effect on real-world behaviour, only judgements made within an experiment.
AI can also reduce false beliefs. In September 2024, Science published the study “Durably reducing conspiracy beliefs through dialogues with AI”. In it, 2,190 people who believed in a conspiracy theory talked with GPT-4 Turbo, which countered them with facts tailored to their arguments. Belief fell by about 20%, and the effect still held two months later.
Watch out
On 11 June 2026, Science published an expression of concern about the 2024 study
It is not a retraction: with this editorial expression of concern, the journal is warning its readers that questions are under review.
The authors learned after publication that the participant selection criteria had not been applied in the same way in the paper and in their analysis code, and that the published dataset contained extra rows, caused by an error when merging code: some values in the paper are difficult to reproduce. They sent the journal a corrected analysis which, according to them, recovers the original results in direction, statistical significance and magnitude; Science is evaluating it. Until the journal has ruled, we cite this study as a lead, not as proof.
The public is worried. According to the Reuters Institute’s Digital News Report 2025, 58% of those surveyed in 48 markets say they are concerned about their ability to tell what is real from what is fake in online news, a share close to the previous year’s. The survey was run online by YouGov, from mid-January to the end of February 2025 (methodology). The same report notes that, in most countries, the public remains sceptical about the use of AI in news, and says it is more comfortable when humans stay in the loop.
Position 1: AI is primarily a disinformation machine
The first camp starts from an economic observation. Writing a fake article, or making a fake image or a fake voice, used to take time and know-how. That cost is falling to almost zero. Disinformation campaigns can therefore produce more, faster, in more languages, and tailor their messages to each audience. If people already cannot tell a generated text from a human one, as the 2023 study shows, trust in everything that circulates online erodes. Even true content becomes suspect.
Position 2: AI can be a tool for correction
The second camp answers that the same tool can be used for verification. A model that talks patiently, picks up each person’s arguments and counters with tailored facts does what a tired or irritated human no longer does. The 2024 study suggests that a well-designed conversation can shift convictions thought to be unshakeable. The same techniques can help journalists or moderators spot dubious content faster.
This camp has a real argument, but its centrepiece has carried a reservation since June 2026. That has to be said.
What we think
We do not believe AI is “primarily” one or the other. It speeds up whatever it is asked to do. The real change is elsewhere: we can no longer judge a piece of content by its appearance. A well-written text, a sharp photo or a natural voice no longer proves anything. Hence our conviction:
Key takeaway
The useful question is no longer only “is it true?”, but also “where does it come from, and who answers for it?”
For a company that publishes, that means stating the origin of synthetic content, citing primary sources, and keeping a person responsible for what goes out.
A limit of our own view: we do not yet have the hindsight to say how effective content marking really is. We know how to require it; we do not yet know whether it changes how people judge what they see.
What would change our mind: solid studies, in real-world conditions, showing that AI corrects false beliefs at scale. Or, conversely, that it measurably swings elections or behaviour. The evidence available in October 2026 supports neither.
What this changes in practice for a company
The EU AI Act provides, in Article 50(2), that providers of systems that generate synthetic content (text, image, audio, video) must ensure that this content is marked in a machine-readable format. Anyone who uses such a system in a professional context has a part to play too (paragraph 4): they must disclose the deepfakes they produce, and the texts published to inform the public on matters of public interest, unless a human has reviewed them and a person bears editorial responsibility for them. We go through these rules in our analysis of disclosing AI-written text.
Three dates set the timetable:
- July 2026
Regulation (EU) 2026/1744 is adopted.
- 2 August 2026
Article 50 applies.
- 2 December 2026
The period the July regulation gives providers whose system was already on the market comes to an end.
For a small or mid-sized company that runs ads or publishes content, three habits to adopt:
- Disclose generated images and videos in your campaigns. For a deepfake, meaning an image, audio or video generated or manipulated by AI that resembles real people, places or events closely enough to pass as authentic, it is an obligation; beyond that, it is good practice.
- Never publish a figure without its primary source. A figure taken second-hand can be wrong, and AI sometimes makes things up with total confidence.
- Keep human review on everything that goes out to customers. Together with a person who takes responsibility for the publication, it is what exempts a public-interest text from disclosure, and it is what puts the public more at ease, according to the Reuters Institute.
From the field
A client learned this the hard way
On their site, three pieces of content invented by a model had to be unpublished. That is why a figure with no verifiable source does not go into the site you are reading: on this blog, every article starts life as a draft, and it is a human review that publishes it.
The same flood of generated content raises another question, that of what the next models will learn from: we cover it in our analysis of the web feeding on itself.
The Online advertising section gathers our analyses on advertising, starting with our guide to sending CRM sales back to Google Ads. And if you are wondering what these rules change for your own campaigns, write to us to talk about it. For campaigns that work with your site and your CRM, see our digital marketing in one block offer.
Frequently asked questions
Can you tell whether a text was written by AI?
Not reliably by reading it. In the 2023 study, participants could not tell messages written by GPT-3 from those of real Twitter users, and the study covered only that model. That is why European regulation does not rely on the human eye: it asks the providers of tools to mark generated content in a machine-readable format, which makes it possible to identify it as such.
Does an expression of concern mean the study is wrong?
No. It is a warning from the journal, published while it examines questions about the paper; it is neither a retraction nor a correction. In the case of the 2024 study, the authors sent in a corrected analysis which, in their view, confirms their results. Until the journal has concluded, the study should not be cited as established proof.
Where can you check information that looks dubious?
Start with the primary source: the original study, legal text or press release, rather than the article that repeats it. In the Reuters Institute’s 2025 survey, respondents name first the news outlets they trust (38%), official sources (35%) and fact-checking sites (25%), ahead of social media (14%). Younger people rely more on social media and AI assistants.
Sources
- “AI model GPT-3 (dis)informs us better than humans” (new tab)doi.org
- “Durably reducing conspiracy beliefs through dialogues with AI” (new tab)doi.org
- Editorial expression of concern (new tab)doi.org
- Digital News Report 2025 (new tab)reutersinstitute.politics.ox.ac.uk
- Methodology (new tab)reutersinstitute.politics.ox.ac.uk
- EU AI Act (new tab)eur-lex.europa.eu
- Regulation (EU) 2026/1744 (new tab)eur-lex.europa.eu
Jim · founder of Brain Plus AI
Also in online advertising
Stepping back8 min
Should you disclose that a text was written by AI?
Transparency owed to the reader, or a detail of how the text was made, like a spell checker? The EU AI Act settles it only halfway, with a criterion that is useful well beyond the law: does a human answer for the text?
Stepping back6 min
AI-generated content: is the web eating itself?
Text generated everywhere, answer engines that read in place of readers, models that learn from their own output. What is proven, what is not, and what a site must do to stay useful.
Stepping back8 min
Does AI make us dumber? What the studies say
Not all the studies doing the rounds on this subject are worth the same. We rank four of them by strength, from a field experiment to a preprint on 54 people, and say what emerges.