Government Deepfake Response for Election Communications: A Playbook for Election Offices

What a Government Deepfake Response for Election Communications Has to Do

Election officials have one job when a fake clip, call or notice appears: stop it from changing what voters do. That is a communications problem first. If a recording tells voters the wrong date, a wrong polling place or a false rule, the office already knows the correct answer and can publish it immediately. CISA's guidance on public communications describes election officials as the primary sources of official information about elections, and that standing is the strongest tool an office has.

The media itself still needs examination, on a parallel track. Someone has to preserve the earliest copy, establish whether it is synthetic, edited or genuine but out of context, and produce findings that hold up when a journalist, platform or prosecutor asks how the office knows. Mixing the two tracks causes the classic failure, where a correction waits days for a technical verdict while the false claim spreads.

What the New Hampshire Biden Robocall Showed

The clearest documented case in the US is the New Hampshire primary robocall. According to the FCC's forfeiture order, adopted on 26 September 2024 and released on 30 September, thousands of calls went out on 21 January 2024, two days before the primary, carrying a deepfake message in the cloned voice of President Biden with misleading caller ID. The Commission had proposed a $6 million penalty against Steve Kramer on 23 May 2024 and imposed it in full after he did not respond to the notice.

Less than three weeks after the calls, on 8 February 2024, the FCC unanimously adopted a declaratory ruling that calls made with AI-generated voices are artificial under the Telephone Consumer Protection Act. Chairwoman Jessica Rosenworcel said bad actors were using such voices to misinform voters, among other abuses. The ruling gives state attorneys general another basis to pursue voice-cloned robocalls.

For election offices the operational lessons are concrete. The attack arrived by phone, not social media, so platform monitoring alone would have missed it. It displayed the number of a prominent New Hampshire political operative who had nothing to do with it. And the message urged voters not to vote in the primary, something the state could rebut from election law without waiting for any audio analysis.

Preparation Before Voting Begins

CISA's Risk in Focus fact sheet on generative AI and the 2024 election cycle concluded that these tools would likely amplify existing risks rather than introduce new ones, and that many mitigations are long-standing security practices. Most of the work therefore happens before any fake appears:

• Name owners for the procedural correction, technical analysis, platform and carrier reporting, law enforcement contact, translation and accessibility.
• Keep official pages for registration, polling locations, hours and results current on a .gov domain, so every correction can point to one stable source.
• Archive the office's own published audio and video with dates, so investigators can compare a suspect clip with genuine recordings.
• Agree escalation contacts with major platforms, phone carriers and local media ahead of time.
• Run a tabletop exercise with a cloned-voice call, a fake video of an official and a compromised office social account in the same week.

Triage in the First Hours

Not every fake needs the same response. A low-reach parody may only need monitoring, while a false instruction about how or when to vote needs a correction the same day. Decide by the claim and its reach, not by how realistic the media looks, and keep replaying the fake to a minimum: describe it in text where you can.

Preserve before anything is taken down. Capture the earliest available file, URLs, account names, phone numbers and call times, screenshots and engagement counts. Copies that have been re-uploaded several times lose detail, so the first-generation file is the one forensic review needs.

Forensic Review With NeuroDetect

Election offices need findings they can explain. NeuroDetect is Neural Defend's forensic media workspace. An analyst uploads the authorized audio, video, image or document into a case, and the analysis keeps model and version context. Detailed views include the original, explainability heatmaps and attention views, down to regions, video frames and audio segments, so the analyst can point to where a cloned voice or swapped face shows signs of manipulation rather than citing a single score.

An LLM-assisted summary describes the findings in plain language and links each statement to the underlying evidence; the analyst makes the final call. The case exports as a forensic report with observations, uncertainty, provenance, model context and analyst notes, plus structured results. That package is what a platform trust and safety team, a phone carrier or a state attorney general's office will ask for when the office requests removal or refers the case.

Limits Election Offices Should Plan Around

Detection is a signal with error rates. NIST's November 2024 report on synthetic content, AI 100-4, notes that detection methods carry nonzero false positive and false negative rates and that a model built for one language may perform worse on others, which matters for multilingual election communications. Heavily compressed re-uploads can leave too little signal for a firm finding, and an inconclusive result should be reported as inconclusive.

Many harmful election clips are not deepfakes at all. Real footage with a false caption, an old video recirculated as new, or a crude edit can mislead just as well, and a synthetic media detector will correctly find nothing generated. Verify the claim, not only the pixels. Finally, avoid naming a foreign or domestic actor without evidence and the authority to do so; attribution belongs with law enforcement and intelligence partners.