In January 2024, robocalls using a synthetic voice clone of President Biden reached tens of thousands of New Hampshire voters the weekend before the state primary, telling them not to vote and to “save their vote for November.” The calls were convincing enough to fool many recipients. The perpetrator — a political consultant — was eventually charged. The technology used was available, legally, for approximately $1 per minute of generated audio.
That incident was, by the standards of what was already technically possible, primitive.
The generation of synthetic media — images, audio, and video fabricated by artificial intelligence — has undergone a qualitative transformation in the past three years. What once required specialized expertise, significant computing resources, and hours of processing time now takes seconds and requires no technical skill. The “lipsync” video — placing convincing words in a real person’s mouth — has become a commodity product. Face-swap technology that was a cinematic special effect in 2015 is now a smartphone app. Audio cloning that once required hours of voice samples can now be done with three seconds of recorded speech.
The quality crosses the human threshold. Studies by MIT, Stanford, and several European research institutes consistently find that people cannot reliably distinguish high-quality deepfakes from real footage with accuracy above chance. We are in an era where seeing is no longer believing — and democracy, which depends on a shared factual reality to function, is facing a challenge it was never architected to handle.
The Attack Surface
The threat is not uniform. Different applications of synthetic media create different types of damage.
The most discussed category is political disinformation. A fabricated video of a candidate making a racist statement, confessing a crime, or announcing a position reversal — released at the right moment in an election cycle — could cause irreparable reputational damage in the hours before fact-checkers can respond. The cycle of damage does not require the deepfake to be believed by everyone; it requires it to reach enough people at a moment when trust is already low and the margin of electoral victory is thin.
The 2024 election cycle saw deepfake incidents in Slovakia, Indonesia, Pakistan, Bangladesh, the UK, and the United States — a partial list. In each case, the specific incident was identified and debunked within hours or days. The aggregate effect — the repeated experience of encountering videos that might be real — is harder to measure but arguably more corrosive. The damage deepfakes do to political discourse may be less about the specific lies they tell than about the generalized uncertainty they create. If every video of a politician could be fake, the default posture becomes one of skepticism toward all video evidence. This is sometimes called the “liar’s dividend” — the ability of bad actors to dismiss genuine evidence by claiming it is fabricated.
The legal system faces a version of the same problem. Video and audio recordings have long been among the most compelling forms of evidence in criminal and civil proceedings. Courts and juries assign them high credibility. As synthetic media quality improves and authentication tools lag behind generation tools, the reliability of audio-visual evidence becomes structurally uncertain. Defense attorneys have already begun raising deepfake challenges as a tactic in cases where the underlying evidence is genuine.
Non-consensual intimate imagery — synthetic pornography using real people’s faces — constitutes a separate and rapidly growing harm. The targets are overwhelmingly women. The damage to reputation, mental health, employment, and relationships is severe and well-documented. Several jurisdictions have enacted specific laws criminalizing non-consensual synthetic intimate imagery, but enforcement across international jurisdictions is inconsistent and technically challenging.
The Authentication Problem
The natural response to unfalsifiable synthesis is better detection. If AI can create convincing fakes, can AI not identify them?
The answer, currently, is: partially and impermanently.
Detection tools exist and work — to a degree. Several academic and commercial systems can identify deepfakes with accuracy rates of 80-95% on standard benchmark datasets. The problem is that this accuracy degrades rapidly on content that has been compressed, re-encoded, or processed — which is what happens to virtually all video shared on social media platforms. A deepfake that passes through WhatsApp’s compression, Twitter’s re-encoding, or Facebook’s processing pipeline becomes substantially harder to detect by automated tools. The very infrastructure of viral content distribution is an inadvertent laundering mechanism for synthetic media.
More fundamentally, the detection arms race has an asymmetric structure. Generation and detection are trained on the same underlying data, and better detection leads to better generation in a feedback loop that has consistently favored synthesis over authentication. Detection tools are always playing catch-up to generation tools, with a lag of months to years.
The more promising technical approach involves authentication at source rather than detection after the fact. Content provenance frameworks — most notably the C2PA (Coalition for Content Provenance and Authenticity) standard, backed by Adobe, Microsoft, Intel, and several major media organizations — cryptographically sign content at the point of creation, creating a verifiable chain from capture device to publication. Cameras, smartphones, and recording devices that implement C2PA embed a tamper-evident signature at the moment of capture. Content without a valid signature is flagged as unverified; content with a broken signature is flagged as modified.
This is technically elegant and practically challenging. It requires hardware and software changes across the entire capture-to-publication pipeline. It requires platforms to build provenance verification into their display systems. It requires users to understand and act on provenance signals. And it does nothing to address the enormous existing archive of unsigned content, for which provenance cannot be retroactively established.
What Democracy Requires
Democracy rests on a set of epistemic conditions — conditions about what citizens can know and believe — that synthetic media is actively eroding.
The minimum conditions are not demanding. Democracy does not require that all citizens have perfect information or that all political communication be honest. It requires that citizens can, with reasonable effort, access enough reliable information to make meaningful political choices, and that there is some shared factual floor below which political debate cannot descend without general recognition that it has done so.
The 20th-century version of that floor was built on institutions: broadcasters with editorial standards, newspapers with fact-checking practices, courts with evidence rules, professional norms in journalism and politics. These were imperfect floors — not incorruptible, not always honest, not accessible equally across class lines. But they were floors.
Synthetic media attacks those floors directly. It attacks the reliability of audio-visual evidence. It attacks the credibility of public statements. It attacks the ability to distinguish authentic from fabricated political communication. It does this cheaply, at scale, with tools available to any actor regardless of resources.
The democratic response to this challenge has been slow and incomplete. Platform moderation of synthetic media has been inconsistent — faster on high-profile cases, slower on the long tail of lower-profile manipulation. Legislative responses have varied enormously: some jurisdictions have enacted targeted deepfake laws; others have nothing. International coordination on synthetic media governance has barely begun.
The Harder Question
The deeper problem is not technical or legal. It is epistemic.
Modern democratic culture has, for several decades, been experiencing a corrosion of shared epistemic foundations. The decline of trust in journalism, science, and expertise. The rise of partisan media ecosystems with minimal factual overlap. The normalization of the “alternative facts” framing in political discourse. Synthetic media did not create these conditions. It arrived into a landscape already primed for them.
The most damaging application of deepfakes is not a single well-crafted video. It is the slow accumulation of a world in which no recording can be trusted, in which the phrase “it could be AI-generated” becomes a universal doubt-injecting response, in which citizens are rationally uncertain whether anything they see or hear about political events is real.
Democracy cannot function in a post-epistemic environment. It requires that citizens can have evidence-based beliefs about what politicians have said and done. If that capacity is destroyed — not by a single lie but by a generalized collapse of trust in audio-visual reality — the civic infrastructure of democratic decision-making becomes hollow. Elections do not stop happening. They just stop mattering in the way they are supposed to.
This is not inevitable. But preventing it requires a deliberate and sustained response — from technologists, platforms, legislators, educators, and citizens — that is currently running well behind the technology it is trying to address.
The deepfake problem is, in the end, a democracy problem. And it is one that democracies are finding very difficult to talk about honestly.
Also explore:
Political Disinformation and the Post-Truth Era
AI Regulation: Who Controls the Controllers? (published June 23)
Sources & Further Reading
- C2PA — the coalition defining provenance standards for digital media.
- Content Authenticity Initiative — the Adobe-led effort behind Content Credentials.
- NIST AI Risk Management Framework — the US framework for managing AI risks, including synthetic content.
Related Articles
- Algorithm Political Polarization: How AI Divides Our Reality
- Understanding AI Art Copyright Issues
- Exploring the Possibility of AI Sentience
Weekly Geopolitical Briefing
Go Beyond the Headlines
Analysis of the forces shaping the world - delivered every week. No noise, no bias, just depth.
Subscribe Free ?No spam. Unsubscribe anytime.
About the Author
António Monteiro
Engineer by profession, geopolitical analyst by conviction. I believe responsibility for the planet's future doesn't belong only to governments and institutions - it belongs to all of us. Knowledge about geopolitics, international conflicts, and the forces shaping the world is the most powerful tool for becoming more conscious, informed citizens. You don't need to be a diplomat to understand what's at stake - you just need to want to go beyond the headlines. At Outside The Case, I analyze conflicts, power dynamics, and global trends with rigor and accessible language, so you can understand what's really happening in the world.
Read more about the author →