The white paper consists of three parts: an overview of the role of AI in disinformation, a discussion of the effectiveness of countermeasures, and practical insights from experts in the BENEDMO network.

What is the role of AI in disinformation?

The white paper uses a broad definition of AI‑driven disinformation: the use of AI to generate, modify, disseminate or manipulate deliberately false information, including misleading visual content (videos and images), speech and text. This information is deployed to serve political, financial, personal, ideological or other strategic goals.

Examples discussed in the report include:

• synthetic videos imitating well‑known political figures (deepfakes);
• AI generated fake news articles;
• manipulated images for social media posts;
• the use of bots and trolls to spread inauthentic content;
• manipulated audio messages in the context of fraud.

The authors stress that existing academic studies on the share of AI‑generated content in disinformation are still scarce and often based on data from the early 2020s. Since then, generative AI is developing very rapidly and its use is being democratised. Applications such as ChatGPT, Grok and other gen‑AI tools are embedded in everyday communication platforms, making it easy for a wide range of users to create and spread misleading content.

In the Dutch context, the use of generative AI for political purposes becomes visible, among other things, through the CampAlgn Tracker, which analyses posts by parties, candidates and selected influencers around the 2025 general elections. The tracker shows a clear increase in AI‑generated images in the run‑up to the elections, particularly among radical‑right actors. These images tap into existing tensions around migration and the loss of Dutch traditions and are used to project the image of a strong leader.

The white paper points out that AI applications do not necessarily always contain explicitly false factual claims; often they are part of negative campaigning or satire that criticises opponents or reinforces the views and perceptions of one’s own group. This has consequences for possible interventions: simply correcting the factual content is not always meaningful, while explaining the use of AI and the strategy behind such campaigns can be relevant.

Effectiveness of counter‑strategies

The white paper distinguishes between two main pathways to counter AI‑driven disinformation:

• Pre‑bunking and literacy interventions
• De‑bunking, fact‑checking and labelling

Pre‑bunking and (AI) media literacy

Media literacy interventions – in which people are taught in advance how to recognise misleading information – are generally effective in improving detection of mis‑ and disinformation. For AI‑driven disinformation and deepfakes specifically, empirical evidence is still limited, but existing studies show that certain interventions increase both the recognition of deepfakes and self‑reported AI literacy.

At the same time, the authors highlight mixed results and potential side effects. They describe a study in which an AI literacy intervention helped participants to correctly identify fake news articles as false, but in which those same participants subsequently also labelled real news articles as fake more often. In open responses, many participants reported feeling “afraid”, “no longer knowing what to believe”, and that “everything can be fake”.

These findings align with broader literature on inoculation and pre‑bunking: interventions that warn people about the risk of being misled can foster a critical stance, but can also contribute to excessive scepticism, where people start doubting trustworthy information as well.

De‑bunking, fact‑checking and labels

Alongside preventive literacy interventions, the white paper discusses corrective measures deployed after disinformation has spread, such as fact‑checks, warnings and labels. Meta‑analyses generally show that fact‑checking can reduce the credibility of false statements, but that in practice people do not always select or accept corrective information, partly due to confirmation bias. However, this has not yet been systematically applied to the context of AI and visual disinformation. 

In a series of experiments by the BENEDMO lab, focused on the Dutch context, the effectiveness of platform interventions against AI‑generated disinformation was examined, including:

• AI labels or watermarks;
• fact‑check labels;
• community notes;
• direct interventions by factcheckers under misleading posts.

The results show that many standard labels have only limited impact on the perceived credibility of AI‑driven disinformation posts and on the extent to which people believe false claims. Community notes did, in specific cases (for example around disinformation on immigration), lead to a decrease in belief in the false claim.

In another experiment, direct interventions by factcheckers – responding under the post with both factual corrections and context – proved most effective in reducing belief in a false claim. In such cases, it appears useful not only to correct the facts, but also to address the possible intent behind the misleading content, especially among people with strong prior beliefs.

The white paper concludes that there is no simple, generic solution for correcting AI‑driven disinformation. The effectiveness of measures depends on the topic, the form of the content and the audience’s attitudes, and much existing research is based on controlled survey settings that do not fully capture the dynamics of real‑world online environments.

Insights from the BENEDMO consortium

In the third part of the white paper, experts from the BENEDMO network share their experiences with AI‑driven disinformation in practice.

• Media literacy and society
  Media literacy expert Zara Mommerency (Mediawijs) observes both curiosity and resistance among citizens when it comes to AI. She calls for sustained investment in critical media and AI literacy across age groups and sectors, and points to growing distrust towards online information in general.
• Journalistic practice and monitoring
  Editor‑in‑chief Luc van Bakel (VRT NWS) emphasises that newsrooms are simultaneously confronted with more AI‑generated disinformation and with the integration of AI tools into their work. He highlights the lack of scalable detection tools and the need for stronger platform regulation and faster monitoring.
• Detection and fact‑checking with AI
  Guy De Pauw (Textgain) argues that AI is needed to process the volume of problematic content, but that human oversight remains essential. Michaël Opgenhaffen (KU Leuven) sees opportunities to make fact‑checks more accessible through AI (for example via translation and repackaging for different audiences), without compromising verification standards.

Conclusions

Drawing on the state of the science, BENEDMO lab’s own findings and the practical insights from the consortium, the authors conclude that AI has not completely altered the fundamental logic of disinformation, but has significantly increased its scale, speed and accessibility. Deepfakes and AI‑generated content are not consistently more convincing than traditional forms of disinformation, but generative tools make it easier and cheaper for a wide range of actors to produce and disseminate misleading content.

An important consequence is pressure on trust in visual and factual evidence: when people assume that many media can be manipulated, their stance towards the authenticity and legitimacy of information changes.

The white paper underlines that there is no straightforward technical or communicative solution. Effective responses require:

• platform accountability and transparency;
• strengthening verification capacities among media, factcheckers and other institutions;
• sustainable media and AI literacy that fosters critical engagement without sliding into cynicism;
• research that systematically tracks the prevalence, evolution and long‑term effects of AI‑driven disinformation in realistic online environments.

The white paper concludes that AI‑driven disinformation is not a temporary anomaly, but a lasting feature of the digital information landscape. Only a long‑term, collective effort by platforms, media professionals, factcheckers, researchers and policymakers – with continuous monitoring, systematic experimentation and adaptation of practices – can keep democratic information provision sufficiently resilient.

** Read the full whitepaper here **

~~~

Image credits: Yutong Liu & Kingston School of Art / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/