In February 2024, thousands of New Hampshire Democrats received a phone call they thought was from President Biden. "What a bunch of malarkey," the voice said, urging them to save their votes for the November election rather than participate in the primary. The message was clear, the voice convincing, and the entire thing completely fake. Political consultant Steve Kramer had commissioned an AI-generated robocall that cloned Biden's voice—an act that would eventually earn him 13 felony counts and a proposed $6 million FCC fine.

The New Hampshire incident wasn't just political dirty tricks entering the AI age. It exposed a problem that threatens the foundation of democratic elections: we're remarkably bad at telling real audio from fake, and the technology to fool us is getting cheaper and easier to use by the month.

The Human Detection Problem

When researchers tested 2,215 people on their ability to distinguish real political speeches from AI-generated deepfakes, the results were sobering. Participants correctly identified audio as real or fake only 73% of the time. That's barely better than a coin flip with slight skill involved.

The 27% error rate matters more than it might seem. In close elections, even small amounts of voter confusion can swing outcomes. If a deepfake audio clip drops 72 hours before polls open—too late for thorough debunking but early enough to spread—that 27% of people fooled could easily exceed victory margins in competitive races.

What makes people fall for these fakes? Research shows we rely more on how something sounds—the rhythm, tone, and audio quality—than what's actually being said. This makes sense for detecting human impersonators but works against us with AI. Modern text-to-speech algorithms have mastered those acoustic signatures better than human voice actors can. The very cues we evolved to trust now betray us.

How Easy Has Voice Cloning Become?

In May 2024, the Center for Countering Digital Hate tested leading AI voice-cloning tools, attempting to generate election disinformation in the voices of Biden, Trump, and other politicians. They succeeded 80% of the time. The tools that failed had basic safeguards that were, in the researchers' words, "easily manipulated."

Modern voice cloning needs only a few audio samples to work. For politicians, that's trivially easy to obtain—speeches, interviews, and press conferences provide hours of training material. The technology that once required specialized equipment and expertise now runs on consumer hardware. Some tools are free.

This accessibility creates an asymmetric threat. A well-funded campaign can commission sophisticated deepfakes through consultants like Kramer. But so can a lone actor with a grudge and a laptop. The barrier to entry has collapsed while detection capabilities lag behind.

The Detection Arms Race

Audio deepfake detection systems work by analyzing acoustic properties that human ears miss. They look for artifacts in how sound waves behave, inconsistencies in breathing patterns, or telltale signs of algorithmic generation. Machine learning models train on thousands of examples to spot these signatures.

The problem is that this creates an arms race. Deepfake creators use similar machine learning techniques to identify what detection systems catch, then refine their algorithms to avoid those tells. Each improvement in detection spurs an improvement in generation. The cycle accelerates because both sides use the same underlying technology—neural networks learning from data.

Detection faces another challenge that generation doesn't: it needs to work fairly across demographics, languages, and voice types. A system trained primarily on English-speaking male politicians might fail when analyzing other voices. It must also respect privacy—extracting acoustic features without storing sensitive audio data—and provide explanations for why it flags something as fake. Generation technology has no such constraints.

The Regulatory Scramble

The New Hampshire robocalls prompted the FCC to rule in February 2024 that AI-generated voices in robocalls violate existing law. It was a quick regulatory response by government standards, but it only covers one narrow use case. The telecom company that transmitted the calls, Lingo Telecom, faces a proposed $2 million fine, but the technology itself remains legal and widely available.

The United States has no comprehensive federal framework for deepfakes in politics. The EU's AI Act includes transparency obligations—platforms must disclose AI-generated content. China directly regulates synthetic-media providers with labeling requirements. The UK criminalized sexually explicit deepfakes in January 2025. But these approaches don't directly address political deepfakes in campaigns, where speech protections complicate regulation.

This fragmented landscape creates opportunities for abuse. What's illegal in one jurisdiction may be permissible in another. Campaign ads must follow disclosure rules, but "organic" social media posts don't. A deepfake audio clip shared by anonymous accounts faces few legal barriers beyond existing fraud or defamation laws—and those require proving harm and identifying perpetrators.

The Liar's Dividend

Perhaps the most insidious effect of deepfake technology isn't the fakes themselves but what legal scholars Bobby Chesney and Danielle Citron call the "liar's dividend": the ability of politicians to dismiss authentic, damaging content as fake.

When voters know deepfakes exist and know they're hard to detect, every piece of evidence becomes suspect. An authentic recording of a politician making offensive comments? "That's a deepfake." Video of corruption? "AI-generated." This doubt corrodes the shared factual basis that democracy requires.

The liar's dividend doesn't require sophisticated technology to work. It just requires public awareness that such technology exists. We're already seeing politicians invoke deepfakes to explain away genuine recordings. As audio fakery becomes more common and more convincing, this defense becomes more plausible.

This creates a paradox for efforts to raise awareness about deepfakes. The more people know about the technology, the more effective the liar's dividend becomes. Yet without awareness, people fall for obvious fakes. There's no clean solution—only a choice between different vulnerabilities.

Authentication Before Detection

The deepfake problem might be backwards. Rather than trying to detect fakes after they spread, we need systems to verify authentic content at the moment of creation. Some news organizations and political campaigns are experimenting with cryptographic signatures—digital watermarks that prove a recording hasn't been altered since capture.

These provenance systems face adoption challenges. They require cameras, microphones, and software that embed authentication data. They need standardization so different systems can verify each other's signatures. Most difficult, they require trust in the entities managing the authentication infrastructure.

But authentication has an advantage over detection: it shifts the burden of proof. Instead of asking "Is this fake?" we ask "Can you prove this is real?" Authentic recordings carry verification. Everything else is suspect by default. This won't stop all deepfake abuse—unverified content will still spread—but it gives voters and journalists a tool to distinguish verified truth from uncertainty.

The New Hampshire robocalls resulted in fines and felony charges, but they also demonstrated how vulnerable our electoral system is to audio manipulation. Humans can't reliably detect deepfakes. Technology to create them is widely accessible. Detection systems struggle to keep pace. Regulation remains fragmented and reactive.

We're in a period where the technology to deceive has outpaced our ability to discern truth. That gap won't close through better detection alone. It requires authentication systems, platform accountability, media literacy, and perhaps most important, a collective acknowledgment that in the audio realm, our ears can no longer be trusted without verification. The 2026 election cycle will test whether we can adapt quickly enough.