A Policy Playbook to Inoculate the Public Against AI Text Falsehoods

In early March, the former French ambassador to Israel, Gérard Araud, shared an extraordinary, but fake, video of missiles raining down on Tel Aviv. Buildings collapsed. Infrastructure burned. Coming days after the first Israeli and US strikes on Iran, the footage seemed to capture Tehran’s devastating reply. Hundreds of accounts shared it. Tens of thousands watched.

The clip was fake. It was one of hundreds of AI-generated videos that flooded X during the conflict, and it fooled an ambassador. Examples like this explain why policymakers are increasingly alarmed by visual misinformation. France’s President, Emmanuel Macron, drove the point home last year by sharing deepfakes of himself, including as a rapper and a beauty influencer, to promote the Paris AI Action Summit.

But the obsession with deepfakes misses the bigger threat. A new report from the International Panel on the Information Environment (IPIE), a Swiss-based scientific consortium of more than 500 researchers across 75 countries, finds that the most immediate and persuasive risk from generative AI is not visual. It is textual. The conclusion draws on 60 effect estimates from 24 peer-reviewed randomized controlled trials involving more than 33,000 participants.

The trend is striking. People are becoming harder to fool with AI-generated images. The cruder fakes give themselves away: the AI slop of seven-fingered hands and the viral “Shrimp Jesus” have become objects of ridicule rather than belief. Recent studies show deepfakes are now judged as less credible than they were just a few years ago. Public awareness has bred a useful skepticism towards what we see online.

The opposite is happening with text. As large language models like ChatGPT have grown more sophisticated, false machine-generated text is now often perceived as more accurate, credible and believable than truthful information. The very tools we have learned to rely on for everyday queries are the ones we struggle to distrust when they mislead us.

This matters because text is cheap, scalable and easy to customize at industrial volumes. It is also being woven into the fabric of search, messaging and conversational interfaces — a deliberate push to embed the technology into everyday institutions before its full effects are understood. Worse, the IPIE’s findings may understate the danger. Most existing studies tested static text passages, but the conversational chatbots people increasingly rely on appear substantially more persuasive than static messages. Current estimates are likely conservative.

The message for policymakers and platforms is clear: they have been right to worry about deepfakes, but they cannot afford to overlook the quieter, more pervasive influence of AI-generated text.

So what works?

The IPIE’s review points to one intervention that holds up consistently across studies: preventive corrective information. The idea is simple. Give people the knowledge and context they need before they encounter misleading content, not after.

In practice, this means brief explanations of how AI systems generate content, reminders that those systems can produce errors, and short prompts that nudge users to think about accuracy and credibility. Across the studies reviewed, these approaches deliver a small-to-moderate but reliable reduction in the perceived credibility of misinformation. The effect sizes (0.17 to 0.43 standard deviations) may sound modest, but applied at the scale of platforms with hundreds of millions of users, the real-world impact is substantial.

Content labelling is trickier. Labels flagging content as AI-generated or potentially misleading can reduce credibility, but their effects swing widely depending on design, wording, timing and source. A poorly built label can have no effect, or even backfire.

Labelling should not be abandoned. It cannot be treated as a plug-and-play fix either. It is a design problem, demanding rigorous testing, refinement and adaptation across platforms and audiences.

That cautious framing reflects a broader limitation worth confronting head-on. For all the strengths of the underlying research, the evidence base has serious gaps. It is overwhelmingly concentrated in the United States and Western Europe, leaving most of the world’s information environments effectively unstudied. Most experiments are run in controlled conditions, which may not reflect how people behave while scrolling a chaotic social feed. And the science is chasing a moving target: research cycles run in years, generative AI cycles in months. By the time a study is published, the threat it describes has often evolved.

Closing these gaps is essential if policy is to keep pace. Independent researchers need better and more timely access to platform data. Cross-regional, properly funded research must become a prerequisite for any policy claiming to work globally. Without that, public policy will remain reactive and incomplete.

The stakes are not abstract. Generative AI is already shaping public discourse from elections to public health. As these systems embed themselves further into our information ecosystems, their influence will only grow.

The question is not whether misinformation can be eliminated. It cannot, and never has been. The question is whether we can reduce its impact and harden the resilience of our information environment.

The evidence says we can. But doing so will require shifting our focus, expanding our knowledge base and investing in interventions that actually work. Above all, it requires accepting that the most significant risks are not always the most visible ones.

If we keep watching only for what we can see, we will miss what is already persuading us.