Voice AI meets democracy: ElevenLabs signs first electoral authority pact

2026 is shaping up to be the year voice AI goes to the polls, and the stakes are higher than a standard product launch. ElevenLabs, the voice synthesis company, has signed a memorandum of understanding with Brazil's Superior Electoral Court (TSE), its first formal agreement with a national election authority. The deal adds text-to-speech and speech-to-text to the TSE's 'Electoral Assistant', letting Brazilian citizens ask questions by voice across the court's website, app and WhatsApp channels. It is also a move in the larger fight over who controls AI infrastructure, a contest with four distinct strategies vying for control.

The move is a direct response to a problem that predates AI: election information that is inaccessible to visually impaired citizens, low-literacy voters and non-native speakers. Printed leaflets and dense official emails do not reach everyone. Voice AI, as ElevenLabs frames it, can turn a static document into a conversation. Brazil's TSE is betting that this layer of access can also counter election misinformation by routing people toward verified official answers instead of viral audio clips.

But the same technology that reads ballot instructions aloud can simulate a candidate's voice or fabricate a statement. ElevenLabs acknowledges the tension explicitly in its announcement: 'these same capabilities could be misused by bad actors to deceive or misinform voters.' The company has been building countermeasures since before the 2024 U.S. election cycle, and the barriers to cloning keep dropping: a free open-weight model now clones voices across 11 languages.

Three layers of defense

ElevenLabs' safety stack is a set of overlapping protections rather than a single gate. At the front, the 'No Go Voices' (NGV) system cross-checks uploaded audio samples against a proprietary list of public officials, candidates, celebrities and other high-risk individuals. ElevenLabs updates the list using internal monitoring, current events and user reports, and the process speeds up around election seasons.

The second layer is Voice CAPTCHA, which compares the voice in a submitted sample against a real-time recording of the user reading a dynamically generated script. It exists to stop someone from bulk-cloning a voice they do not own. Behind that sit automated moderation systems that flag content violating its policies on hate speech, scams and child exploitation. Some content is blocked before it can be generated; the rest is queued for human review.

Traceability is the third leg. Since 2023, ElevenLabs has offered a free AI Speech Classifier that lets anyone check whether an audio sample was generated on its platform. More recently, the company began watermarking audio with Google DeepMind's SynthID, so generated audio carries a signal its Audio Detector can read. Detection is not a settled game, and research showing AI detectors can backfire is a reminder that every fix tends to open a new flaw.

From Brazil to the UK and Greece

The Brazil deal is part of a wider push into government services. ElevenLabs has also signed formal partnerships with Greece and the United Kingdom, extending the pattern of small European states handing money to the same voice AI startup. In Greece, the company is working with the government on voice AI for public services, tourism and preservation of the country's linguistic heritage. In the UK, a memorandum of understanding with the Department for Science, Innovation and Technology covers public service accessibility, AI security research and talent development. In Ukraine, ElevenLabs provides the voice layer for the Diia app, a government services portal.

The Brazil partnership, however, is the first that explicitly targets election information. The company has committed to training Comprova, a coalition of 42 Brazilian media organizations, to use its AI Speech Classifier and other detection tools in their verification workflows. In the United States, ElevenLabs collaborates with California's Countering Tech-Enabled Fraud Task Force and deepfake detection firm Reality Defender to strengthen real-time detection tools used by large enterprises and governments.

None of these measures is foolproof, and the company does not claim they are. Benchmarks and live elections are different arenas, exactly the gap a framework that defines trust by the numbers was built to close. The real test will come during Brazil's October elections, when real voters interact with a voice interface for the first time on a national scale, and when the No Go Voices list faces its busiest season yet. If voice AI manages to expand access without amplifying the noise it was meant to counter, it will have passed a trial that few other technologies have faced in public view.