What this threat is
Deepfakes are the most visible face of this problem: AI-generated video, audio, and images that convincingly depict real people saying or doing things they never said or did. The technology to create realistic deepfakes was, until recently, expensive and technically demanding. It required significant compute, specialized skills, and time. Those barriers are gone. Tools that generate convincing synthetic video or clone a person's voice from a short audio sample are now widely available and require no special technical knowledge. The practical implication is that fabricated audiovisual content of any public figure can be generated by anyone with a consumer device and a few minutes.
Text-based AI misinformation is in some ways the more immediate and pervasive problem. Large language models can generate plausible-sounding articles, social media posts, forum comments, and email communications at massive scale with minimal human involvement. This isn't about individual pieces of false content; it's about the ability to flood an information environment with a coordinated volume of synthetic content that overwhelms human moderators and creates the impression of broad consensus around a false narrative. Documented influence operations have already used AI-generated text to populate networks of fake personas across social platforms, a technique sometimes called "synthetic persona" operations, creating the appearance of organic grassroots support for positions that originate with state actors or organized groups.
Personalization is the dimension that makes AI disinformation qualitatively more dangerous than traditional propaganda. Traditional disinformation operated at the level of mass messages that reached everyone the same way. AI enables messages tailored to individual psychological profiles, using data about a person's beliefs, fears, past behavior, and social connections to construct a version of a claim that's specifically designed to resonate with that individual. A politically engaged person and a health-focused person and a financially anxious person might each receive a different AI-generated framing of the same false claim, each optimized for their particular vulnerabilities. At scale, this is a fundamentally different kind of persuasion than any that existed before.
The "liar's dividend" is a secondary effect that deserves attention. As synthetic media becomes more prevalent and more convincing, the response includes a reflexive skepticism toward all media, including real content. A politician who says or does something genuinely damaging can now claim it's a deepfake, and that claim will be more plausible than it would have been before deepfakes existed. This cuts in two directions: it creates a mechanism for spreading false content, and it creates a mechanism for dismissing true content as false. Both effects degrade the epistemic foundation that democratic discourse requires.
Why it matters
Democratic elections are the most immediately visible stake. Voters making decisions based on fabricated video of candidates, AI-generated false claims about voting procedures, or synthetic content designed to suppress turnout in specific communities face a corrupted information environment at exactly the moment when accurate information matters most. Multiple elections in recent years have featured documented AI-generated disinformation campaigns, and the sophistication and scale of those campaigns is increasing faster than the detection and response capacity of election authorities or platforms.
The damage extends well beyond elections. Health misinformation, including false claims about vaccines, treatments, and disease transmission, causes direct harm when people make medical decisions based on it. The COVID-19 period demonstrated how quickly false health information can spread through online networks, and AI makes that problem significantly worse by enabling the generation of content that mimics the style and format of legitimate medical communication. Financial fraud using AI-generated voice cloning (so-called "vishing" or voice phishing attacks) is already costing individuals and organizations real money at scale. The technology behind these fraud schemes is the same technology that makes AI-generated media useful for legitimate purposes.
The broader damage is to shared social trust and the epistemic commons. A functioning society requires some baseline of shared agreement about what is and isn't true, not perfect consensus, but enough shared reality that people can communicate, make collective decisions, and hold institutions accountable. AI-generated disinformation at scale erodes that baseline by making it genuinely difficult for ordinary people to distinguish authentic from synthetic content, by flooding public conversations with noise, and by making every disputed claim feel like it has equal evidential support regardless of the actual state of the evidence. That erosion is a structural problem that compounds over time and doesn't reverse easily once it reaches a certain depth.
Where things stand today
The capabilities are advancing faster than both detection tools and governance frameworks. Researchers have documented AI-generated synthetic media and text-based disinformation campaigns in electoral contexts across multiple countries. Platform companies have invested in detection systems, but the fundamental asymmetry is that generating synthetic content is computationally cheaper and easier than detecting it, especially as generation models improve. Detection systems trained on known fake content from current models become less effective as the generation techniques evolve.
The EU AI Act's transparency requirements are the most significant regulatory response to date. The Act requires that AI-generated content be labeled as such, that deepfakes (synthetic images, audio, or video) be disclosed unless the context makes it obvious they're synthetic (such as obvious satire), and that operators of AI systems generating synthetic content implement technical measures to enable provenance tracking. Watermarking and content provenance standards are being developed by technical bodies and some major AI companies, including work on cryptographically verifiable content credentials that embed information about the origin and editing history of digital content. These are meaningful steps, but they depend on widespread adoption and on the platforms where content circulates enforcing the standards.
What isn't covered is significant. The EU AI Act transparency provisions don't apply retroactively to content that's already in circulation, don't address the international platforms through which most disinformation spreads, and don't resolve the detection problem for content that's generated without disclosure. The international coordination required to address cross-border disinformation campaigns doesn't exist in any robust form. Research into media literacy tools that help people evaluate content more critically represents a complementary direction, but the evidence on the effectiveness of media literacy interventions at population scale is mixed.
How Better Societies helps
Compliance: The EU AI Act's transparency obligations for synthetic content are specific and technically detailed. Organizations generating or deploying AI systems that produce synthetic media, voice, or text face disclosure requirements, technical implementation obligations, and documentation demands. Our compliance programs help organizations understand exactly what's required, assess which of their AI systems are affected, and implement the disclosure and provenance infrastructure the regulation demands before enforcement catches up.
Summit: Information integrity is a global problem that no single country's regulation can solve. The Better Societies Summit creates space for researchers, platform representatives, election administrators, journalists, and policymakers to coordinate on detection standards, platform governance, and the international frameworks needed to make content provenance systems work at scale.
Accelerator: Some of the most important work on this problem is being done by founders building detection tools, content provenance systems, media literacy platforms, and verification infrastructure. If you're building technology that helps people identify synthetic content, establish trust in authentic content, or navigate a higher-noise information environment, the Better Societies Accelerator is where we support that work and connect you with the organizations that need it most.