In January 2024, a robocall using an AI-generated clone of President Biden’s voice urged New Hampshire voters to skip the primary election. In March 2024, a deepfake video of Ukraine’s President Zelensky appearing to surrender circulated during active warfare. In Slovakia’s 2023 election, an audio deepfake of a candidate discussing vote-rigging was released 48 hours before polls opened — too late to debunk before voting.
These are not prototypes. They are deployed weapons in the information environments that democracies depend on.
What deepfakes are
Deepfakes use deep learning (specifically generative adversarial networks and diffusion models) to create synthetic media — video, audio, or images — depicting real people doing or saying things they never did.
Video deepfakes — face swap, lip sync, full body generation. Quality ranges from obvious manipulation to broadcast-indistinguishable.
Audio deepfakes (voice clones) — as few as 3–10 seconds of source audio can train a convincing voice clone. ElevenLabs, Resemble AI, and open-source tools make this accessible to non-technical users.
Real-time deepfakes — live video manipulation during calls. Zoom meetings, video interviews, and live streams can display synthetic faces responding in real time.
The technology improved faster than detection technology. The gap is widening.
The threat to democratic processes
Elections:
- Fake audio/video of candidates making disqualifying statements
- Synthetic footage of voter fraud (creating pretext for contesting results)
- Deepfake “leaks” timed for maximum impact before debunking is possible
- Erosion of trust in ALL media, including genuine footage (“liar’s dividend”)
The liar’s dividend: When any video might be fake, politicians can dismiss genuine evidence of misconduct as deepfake. The existence of synthetic media provides plausible deniability for real behavior. This may be more damaging than the fakes themselves.
Journalism:
- Verifying video evidence becomes exponentially harder
- Source protection complicated (was the whistleblower real?)
- Speed of publication vs. speed of verification creates permanent tension
Legal system:
- Video evidence in court requires forensic authentication
- Defendants claim real evidence is synthetic
- Witness testimony complicated by uncertainty about video authenticity
International conflict:
- Synthetic footage of military actions, atrocities, or surrenders
- Deepfake propaganda targeting enemy populations
- “Fog of war” amplified by synthetic media
Detection — an arms race we’re losing
Current detection methods:
- AI classifiers trained on deepfake datasets (FaceForensics++, DeepFakeDetection)
- Metadata analysis (C2PA content credentials, EXIF data)
- Biological signals (imperfect blink patterns, pulse detection from skin color variation)
- Provenance tracking (blockchain-adjacent authentication at point of capture)
Why detection fails at scale:
- Generative models improve faster than detectors (the “detector’s dilemma”)
- Compression, filtering, and re-sharing degrade detection signals
- Open-source tools lower the barrier to creation below the barrier to detection
- Real-time generation outpaces batch analysis
- Detection tools are unavailable to most consumers
The fundamental problem: Proving something is fake is harder than creating the fake. A deepfake requires one creator. Debunking it requires institutional resources, technical expertise, and time that misinformation does not afford.
What governments are doing (not enough)
United States:
- FCC banned AI-generated voices in robocalls (2024)
- No comprehensive deepfake legislation at federal level
- State-level laws in California, Texas, Minnesota (varying scope)
- CISA (Cybersecurity and Infrastructure Security Agency) monitoring threats
European Union:
- AI Act (2024) requires labeling of AI-generated content
- Digital Services Act mandates platform responsibility for synthetic media
- Code of Practice on Disinformation includes deepfake provisions
China:
- Requires labeling of AI-generated content (2023 regulations)
- Deep synthesis regulations among world’s strictest
- Irony: strict domestic regulation, suspected deployment in foreign influence operations
Platform policies:
- Meta, Google, TikTok, and X require labeling of AI-generated political content (enforcement varies)
- Watermarking initiatives (Google SynthID, OpenAI metadata) — voluntary, removable
What individuals can do
Before sharing video of political figures:
- Check the source — is this from a verified account or anonymous upload?
- Search for corroboration — is any other outlet reporting this?
- Check timing — was this released suspiciously close to an election or event?
- Look for artifacts — unnatural blinking, skin texture, audio-visual sync issues
- Wait — the 24–48 hours after a viral video is when debunking occurs. Sharing immediately amplifies potential fakes.
Long-term habits:
- Follow institutional fact-checkers (Snopes, PolitiFact, AFP Fact Check)
- Support media literacy education
- Advocate for C2PA adoption in devices and platforms
- Accept that some genuine content will be doubted — the cost of the deepfake era
The deeper crisis: epistemic collapse
Deepfakes do not just create false evidence. They destroy confidence in true evidence. When any image, video, or audio might be synthetic, the shared factual basis that democratic deliberation requires erodes.
This is not hypothetical. Surveys show declining trust in video evidence among all age groups since 2020. The trend predates convincing deepfakes — social media misinformation started the erosion. Deepfakes accelerate it.
A democracy where citizens cannot agree on what happened — because any account might be manufactured — is a democracy that cannot hold leaders accountable, cannot debate policy from shared facts, and cannot distinguish propaganda from journalism.
What might work
Provenance over detection — ensuring authentic media is cryptographically signed at capture (C2PA standard) rather than trying to detect fakes after distribution. Camera manufacturers, Adobe, and Microsoft are implementing this.
Legal deterrence — criminal penalties for deploying political deepfakes within election windows. Several jurisdictions moving in this direction.
Platform architecture — designing sharing mechanisms that slow viral spread of unverified media (friction, not censorship).
Media literacy at scale — teaching critical evaluation of media as core curriculum, not optional enrichment.
Norms — journalists, politicians, and institutions refusing to share unverified video regardless of political advantage.
None of these solve the problem completely. All of them reduce the damage. The alternative — continuing to treat deepfakes as a technical curiosity rather than a democratic emergency — guarantees the worst outcome.
Seeing was believing. That contract held for a century of photography and decades of video. AI broke it in three years.
Rebuilding trust in what we see — or building new frameworks for verification that do not depend on visual certainty — is the democratic challenge of this decade.
The deepfake is not coming. It is here. The question is whether democracy adapts faster than deception scales.
Lumen is edited by Leo Hartmann. Related: AI Art vs Photography · Online Privacy Guide