What happened
On 21 January 2024, two days before the New Hampshire presidential primary, thousands of voters received a phone call in what sounded like the voice of President Joe Biden. The voice opened with one of his stock phrases, “What a bunch of malarkey,” and told Democrats not to vote: “It’s important that you save your vote for the November election.” Voting in the primary, it said, “only enables the Republicans in their quest to elect Donald Trump again.” The call was a fake. The White House confirmed the president had said no such thing, and the caller identification had been spoofed to show the number of a real New Hampshire Democratic organiser.
The voice had been generated with a consumer voice-cloning tool made by the company ElevenLabs. A New Orleans street magician, Paul Carpenter, said he had produced the audio in under twenty minutes at a cost of about a dollar. He had been hired by a political consultant, Steve Kramer, who was working for a rival candidate in the Democratic primary. Kramer admitted commissioning the call and claimed he had done it to draw attention to the dangers of artificial intelligence in elections. A voice-analysis firm, Pindrop, traced the audio to ElevenLabs, which then banned the account responsible.
The regulatory response was fast. On 8 February 2024 the Federal Communications Commission ruled that calls using AI-generated voices are “artificial” under existing telemarketing law, which makes such robocalls illegal. In May 2024 the FCC proposed a six million dollar fine against Kramer, which it finalised as a forfeiture order in September 2024. The carrier that had transmitted the calls, Lingo Telecom, settled with the FCC in August 2024 for one million dollars and a compliance plan. New Hampshire charged Kramer with multiple felony counts of voter suppression and misdemeanour counts of impersonating a candidate. At trial in June 2025 a jury acquitted him on all counts.
What an auditable version would have shown
The audio was created in minutes by a tool that imposed no check on whose voice it was cloning or what the clone would say, and kept no record that tied the output to the person who made it in a form available before harm rather than after. The cloning happened, and nothing stood between the request and a synthetic president telling voters to stay home. An auditable version records each generation as a signed event, the account that requested it, the voice it imitated, and the content produced, and refuses outright the cases that should never run. Tracing the call to ElevenLabs took outside analysis once the calls had already gone out. A system built this way treats the identity of the cloned voice and the nature of the message as something to check and record at the moment of generation, not something to reconstruct after a regulator asks.
Where the gap was
There were two missing controls, and both sit with the tool that made the voice. The first is that nothing refused the request to synthesise a named, identifiable public figure, still less to put words about an election in his mouth. A ConstraintGate is the control on that: cloning a real person’s voice, or generating election content in it, is barred unless an authorised, recorded exception is granted, rather than being a default anyone can invoke. The second is that the only record of who did what was assembled afterwards, by a third party. A ConductRecord is the control on that: a signed log of every generation, the requester and the output, so misuse is attributable from the inside and in real time. The point is not that the call could never be traced. It is that tracing it depended on luck and effort after the votes had been targeted, when the record should have existed by design.
What governance should have looked like
A tool that can put convincing words into the mouth of any public figure is a tool that can be aimed at an election, and this one was, for the price of a coffee. The lesson is that the controls belong at the point of generation, not only in the law that punishes afterwards: a system that synthesises real voices should refuse the highest-risk cases, should require recorded authorisation to impersonate an identifiable person, and should keep a signed record of everything it produces so misuse is visible to the people who run it. The FCC’s ruling and the fines came quickly, and they came after thousands of voters had already been told by a counterfeit president to stay home.
The reference implementation of ConstraintGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- FCC Makes AI-Generated Voices in Robocalls Illegal (Federal Communications Commission)
- FCC proposes $6 million fine for Steve Kramer over New Hampshire robocalls (FCC enforcement, PDF)
- Fake Joe Biden robocall urges New Hampshire voters to skip the primary (CNN)
- Political operative behind fake Biden robocalls found not guilty (NHPR)