Election security used to separate voting-system protection from public communications. Deepfakes make that division less useful. A convincing synthetic recording can impersonate a candidate, election official, journalist, or poll worker without altering a single ballot. Its objective may be to suppress participation, redirect voters, provoke harassment, steal credentials, or cast doubt on authentic evidence. The operational response therefore has to combine cybersecurity, crisis communication, law enforcement, platform coordination, and public verification.

The threat is best understood as an accelerator of familiar tactics. In guidance for the 2024 U.S. cycle, CISA assessed that generative AI capabilities would probably not create wholly new risks, but “they may amplify existing risks.” It listed synthetic video, altered images, voice cloning, and generated text as tools for influence operations, scams, impersonation, and phishing (CISA). Lower production costs let more actors create variants in more languages and adjust them quickly to breaking events.

An incident before New Hampshire’s January 2024 presidential primary showed the combination of media synthesis and traditional distribution. Thousands of potential voters received robocalls using an artificial voice resembling President Joe Biden and telling them not to vote in the primary. The FCC proposed a $6 million fine over spoofing violations and separately ruled that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act (FCC). The case did not require a perfect video or sophisticated platform manipulation; a familiar telephone channel and targeted timing were enough.

Foreign influence campaigns add scale and strategic intent. A September 2024 U.S. intelligence update said Russia and Iran were using generative AI to improve and accelerate parts of their influence efforts, while judging that the technology had not yet revolutionized those operations. It noted that impact depended on distribution, evasion of safeguards, and the actors’ ability to remain undetected (ODNI). That distinction matters. Generating a plausible clip is easier than making the right audience believe and share it.

Speed is the central defensive challenge. A false audio message released hours before polls close may do damage before a forensic verdict arrives. Election offices need pre-established channels where voters can verify polling locations, hours, eligibility, and results. They should publish consistent information across an official website, social accounts, local media, and telephone lines, with backups if one channel is compromised. The U.S. Election Assistance Commission’s 2026 AI resources emphasize communications practices and printable material directing voters to verifiable official sources (EAC).

Authentication also has to move closer to publication. Campaigns, media organizations, and election agencies can secure high-value accounts with phishing-resistant multifactor authentication, restrict who can publish, log changes, and establish out-of-band confirmation for unusual requests. Cryptographic content credentials can record provenance when media is created and edited. Watermarks and detectors may provide signals, but compression, cropping, re-recording, and open generation tools limit any single mechanism. Absence of a marker cannot prove that a file is authentic.

Deepfakes Are Changing the Security Playbook for Elections — supporting image
Deepfakes Are Changing the Security Playbook for Elections — supporting image

Law and platform policy are evolving unevenly. The European Union’s AI Act transparency duties under Article 50 became applicable on August 2, 2026. They require providers to facilitate machine-readable identification of certain synthetic content and require deployers to disclose deepfakes, subject to defined exceptions (European Commission). The EU’s Digital Services Act election toolkit also recommends preparation, stakeholder coordination, incident response, monitoring, and media literacy for very large platforms and regulators (European Commission).

Rules must still distinguish harmful deception from satire, artistic work, accessibility tools, and legitimate campaign speech. A broad ban can chill expression and be difficult to enforce across borders; a label can be ignored or deliberately applied to authentic media to discredit it. Enforcement works best when it targets concrete conduct—fraud, impersonation, voter suppression, undisclosed synthetic political advertising, malicious robocalls—or imposes transparent process duties on actors with control over distribution.

Exercises should test more than detection. Officials can simulate a fabricated recording of a local election director, a cloned voice requesting a password reset, or a fake image alleging ballot destruction. The evaluation questions are practical: Who verifies the claim? Who approves a response? Can officials reach affected communities in their languages? Will platforms and broadcasters preserve evidence? Are law-enforcement contacts current? How does the office correct the falsehood without amplifying it unnecessarily?

Deepfakes do not make evidence meaningless, but they raise the value of provenance, trusted channels, and institutional credibility built before a crisis. The new security playbook is less about identifying every synthetic pixel than ensuring that voters can find authoritative facts quickly and that officials can authenticate themselves under pressure.

Sources