AI and Elections: The Officials Nobody Funded
Election administrators were handed a new threat category without being handed any money. Mostly they coped, and the way they coped is the interesting part.
The widespread fear going into recent election cycles was that generated media would overwhelm the information environment: synthetic video of candidates saying things they never said, cloned voices in robocalls, fabricated documents released at the moment they could not be debunked.
The reality was messier and less dramatic, and it put an enormous burden on local officials who had no particular expertise in machine learning.
Recycled footage did more damage than synthesis
- Low-quality synthetic content was abundant, but most of it was aimed at confirming existing beliefs rather than changing them.
- The most effective false claims were usually text and recycled real footage, not novel generated media.
- Detection and debunking were slower than distribution, but the gap narrowed where officials had prepared.
- Trusted local messengers mattered more than platform labelling.
Pre-register the channel, then staff the phone line
Administrators and civil society groups converged on a set of practical measures: pre-registering the official channels so impersonation was visible, rapid-response verification teams, prebunking campaigns that explained techniques before they were used, and relationships with platforms that allowed speedy escalation.
None of this is technically sophisticated. All of it requires people and coordination, and it is the first thing cut when budgets tighten.
Trust did more work than detection technology
The resilience of an information environment turns out to depend less on detection technology than on institutional trust. Where the public had a reason to believe its election administrators, synthetic content did limited damage. Where that trust was already thin, generated media was fuel rather than cause.
One staff member and three pre-written explanations
The measures that reduced harm were mostly unglamorous: a staff member with a direct line to platform trust teams, a documented process for verifying a viral claim within hours, and pre-written explanations of common manipulation techniques.
Funding for those measures is typically temporary and tied to a specific election cycle, which means the capability is rebuilt from scratch each time. Institutional knowledge that is not funded between cycles does not persist.
When a photograph stops being evidence
The deeper concern among officials is not any particular technology but the erosion of the assumption that a photograph or a recording is evidence. Once that assumption is gone, the burden of proof shifts to everyone, and the advantage moves to whoever is willing to assert things confidently.
No technical measure addresses that. It is a problem of institutional trust, and it predates generative AI by decades.
A practical point for anyone in this field: the measures that work are mostly about speed of correction rather than prevention. A false claim that is debunked within an hour does far less damage than one that circulates unchallenged for a day, and the difference is entirely a function of how quickly a trusted voice can respond.
Image credit and licence details for every photograph on this site are listed on the credits page. This article is editorial content; it carries no sponsored material.