Vermont Senate Committee Hears Expert on Detecting AI in Election Media as S.23 Advances

The Senate Committee on Government Operations heard expert testimony Tuesday on S.23, an act relating to the use of synthetic media in elections, and whether broadcasters or content creators can reliably determine if political material was produced by artificial intelligence.

Julia Zimmerman, a PhD candidate at the University of Vermont, told the committee that diagnostic tools can identify statistical patterns that make AI-generated text or manipulated imagery likely, but no method guarantees absolute certainty. "Anything that promises, like, a 100%, that's, like, probably a little bit shaky," Zimmerman said.

Zimmerman described different detection approaches for text, audio and video. For language, analysts can compare statistical regularities against large human corpora to flag likely AI output. For manipulated images and deepfake video, she said investigators sometimes examine physical inconsistencies — for example, lighting that would violate real-world physics — and that such inconsistencies are strong indicators of fabrication. "There's nothing innate about that that would mean it's something the model can't improve on in the future," Zimmerman cautioned, noting that once a benchmark is public, models can adapt.

Committee members raised whether a "knowingly" standard in S.23 — which would penalize parties that knowingly broadcast AI-generated material without disclosure — is practical for broadcasters that receive ads from third parties. One senator argued that broadcasters should be expected to run checks because they profit from the ads; another responded that small stations with only a few employees and FCC equal-access obligations may lack the resources to perform detailed forensic checks.

Senator Glaxie, who pressed the committee on an exemption for broadcasters, said she supported holding advertisers and creators accountable rather than low-level staff who merely queue content for airplay. "I don't agree that the person who may be volunteering or paid very little that pushes the button on air should be held accountable," Glaxie said, adding that the onus should be on the entity that purchases or produces the ad.

Witnesses and lawmakers discussed the difference in detectability among media types. Zimmerman and committee members agreed that audio-only political messages — including robocalls — are harder to distinguish from human-generated material than paired audio-and-video items because video supplies additional visual cues. She also said some detection tools are open-source and likely free to use, though running or combining multiple checks can require computing resources; she expected the computational cost of checking to be lower than the cost of generating advanced synthetic media.

The committee did not vote on S.23 during the session. Members discussed possible legal standards and practical compliance measures, including a sliding-confidence approach in which organizations rely on multiple detection methods and document due diligence. Zimmerman warned that any benchmark adopted now may require updates as models and detection methods co-evolve.

The committee adjourned after thanking the witness and previewing forthcoming agenda items, including testimony on optometry scope of practice and open meeting law amendments.