AI and Election Disinformation 2024-2025: What Actually Happened | State of Surveillance

TL;DR: 2024 was supposed to be the year AI-powered disinformation broke democracy. In reality, the impact was more complicated. Romania annulled its presidential election after evidence of AI-assisted foreign interference. Up to 25,000 New Hampshire voters received AI-generated Biden robocalls. India, Taiwan, Slovakia, and Germany all saw deepfake incidents. But researchers found that "cheap fakes", low-tech manipulation, were used seven times more often than AI-generated content. The AI apocalypse didn't happen. What happened instead was a testing ground: nation-states and bad actors probing what works, building capabilities, and preparing for what comes next.

What Actually Happened: Case Studies

Romania: The First Annulled Election

In December 2024, Romania's Constitutional Court annulled the results of the first round of its presidential election, a first in modern European history.

The reason: evidence of AI-powered foreign interference. Declassified intelligence documents showed a coordinated campaign using manipulated videos targeting voters on social media. The targeted and sophisticated nature of the interference, very likely foreign-sponsored, led the court to invalidate the results entirely.

This wasn't theoretical. An actual election was overturned because of AI-assisted manipulation.

New Hampshire: The Biden Robocalls

In January 2024, up to 25,000 New Hampshire voters in the Democratic primary received phone calls featuring an AI-generated voice of President Biden. The synthetic Biden urged them not to vote in the primary, a classic voter suppression tactic delivered via next-generation technology.

The calls were traced to a political consultant working for a rival campaign. He now faces criminal charges and significant fines. The incident demonstrated both the capability and the consequences: deepfake audio for mass voter suppression is trivially easy to create, but getting caught carries real penalties.

Slovakia: Election Eve Audio Leak

Two days before Slovakia's September 2023 parliamentary election, AI-generated audio surfaced appearing to show a phone conversation between a journalist and the leader of the liberal Progressive Slovakia party. In the fake audio, they discussed rigging the upcoming election.

Progressive Slovakia suffered an upset loss at the polls days later. The audio was debunked as AI-generated, but not before it had spread widely on social media during the crucial final hours of the campaign, a period when Slovak election law prohibits candidates from responding to attacks.

The timing was surgical.

Taiwan: China's First Confirmed AI Operation

During Taiwan's January 2024 election, deepfake videos promoted hoaxes including a fabricated "secret history" of outgoing leader Tsai Ing-wen. Microsoft later termed the China-based operation its "first confirmed use of AI-generated material by a nation-state to influence a foreign election."

The operation combined AI-generated content with traditional disinformation tactics, a hybrid approach that's becoming standard practice.

India: Limited AI, Amplified Impact

India's 2024 elections saw AI-generated deepfakes showing celebrities criticizing Prime Minister Narendra Modi and endorsing opposition parties. These went viral on WhatsApp and YouTube.

However, the actual scale was smaller than feared. Out of 258 election-related fact-checks conducted by the Indian fact-checking organization Boom Live, only 12 involved AI-generated misinformation. The vast majority of false content used traditional methods, screenshots, out-of-context clips, and simple image manipulation.

AI was a small percentage of the disinformation, but it attracted outsized attention and engagement.

Germany: AfD's AI Campaign

Alternative for Germany (AfD) posted a series of AI-generated political advertisements ahead of the February 2025 federal election. The ads depicted idealized scenarios that didn't exist, manufactured images of prosperity and order designed to promote the far-right party's messaging.

Unlike deepfakes of real people saying things they never said, these were synthetic scenes that never happened at all. The technique raises different questions: not "is this person real?" but "is this reality real?"

Putting It in Perspective

Despite the high-profile incidents, AI-powered disinformation wasn't the dominant threat in 2024-2025. Research from the News Literacy Project found that "cheap fakes", low-tech manipulation like out-of-context clips, misleading captions, and selectively edited video, were used seven times more often than AI-generated content in election-related misinformation.

Why "Cheap Fakes" Still Dominate

Speed: Taking a real clip out of context takes seconds. Creating a convincing deepfake takes longer.
Deniability: "I just shared a video" is easier to defend than "I created synthetic media."
Volume: A single operator can share thousands of misleading clips faster than they can generate thousands of deepfakes.
Effectiveness: Viewers often don't notice subtle manipulations. The viral clip with a misleading caption spreads just as well as an AI-generated one.

The Limited (But Growing) Impact

Researchers at Harvard's Ash Center concluded that "AI was everywhere in 2024's elections, but deepfakes and misinformation were only part of the picture." The apocalyptic predictions didn't materialize, but that doesn't mean the threat isn't real.

The capability exists and is improving rapidly. The incidents that did occur proved the concept works. Nation-states are building infrastructure and testing approaches. What we saw in 2024-2025 may be the foundation for much larger operations in the future.

2025: New Threat Vectors Emerge

Throughout 2025, new AI-enabled election interference tactics have emerged:

Poisoned Chatbots

In Australia's May 2025 federal election, a Russian-linked influence network published fake news stories designed specifically to be indexed by AI chatbots. The goal wasn't to reach human readers directly, it was to contaminate the training data and knowledge bases that AI assistants draw from.

When voters asked AI chatbots about candidates or issues, they received responses influenced by deliberately planted disinformation. This marks a shift from targeting social media feeds to targeting AI systems themselves.

Financial Scams and Voter Suppression

AI tools are being misused for illicit financial scams tied to elections, fake donation sites, synthetic candidate endorsements of investment schemes, and fraudulent campaign communications designed to extract money or personal information.

Voter suppression tactics have also evolved: AI-generated messages claiming polling places have moved, registration deadlines have changed, or voting methods have been altered.

Gendered Disinformation

In elections across India, Indonesia, and Mexico, AI was used to create defamatory images of female candidates, specifically building on and amplifying misogynistic stereotypes. Synthetic content depicting women in sexualized or demeaning situations was deployed as a targeted weapon against specific candidates.

This isn't new, women in politics have always faced gendered attacks, but AI dramatically lowers the cost and increases the realism of producing such content.

How Governments Are Responding

European Union

The EU has been most aggressive in regulation. Under the Digital Services Act, large online platforms like Facebook and TikTok must identify and label manipulated audio and imagery, including deepfakes, by August 2025.

The EU AI Act, which took effect in 2024, requires disclosure when AI-generated content is used in political advertising. Enforcement mechanisms are still developing.

United States

The federal approach has been fragmented. Several states have enacted specific prohibitions on deepfakes in election communications. Those involved in the New Hampshire robocalls face criminal charges, establishing a deterrent precedent.

However, there's no thorough federal law governing AI in elections. The FEC has proposed requiring disclosure of AI-generated content in political ads, but enforcement remains uncertain.

Platform Responses

Major platforms have implemented AI content disclosure policies:

Meta: Requires political advertisers to disclose when AI is used to create or alter footage (effective January 2024)
YouTube: Requires disclosure labels on synthetic content
TikTok: Labels AI-generated content (though enforcement is inconsistent)

These policies are largely self-enforced by advertisers. Platforms have limited ability to detect undisclosed AI content, particularly as generation techniques improve.

The Detection Problem

Detecting AI-generated content is getting harder, not easier. As generation techniques improve, the artifacts that detection tools rely on, unnatural blinking, lighting inconsistencies, audio-video sync issues, become less pronounced.

Current Detection Limitations

Speed: Detection takes longer than generation. By the time a deepfake is confirmed as fake, it may have already gone viral.
Arms race: Generators and detectors are improving simultaneously. Detection advances don't create lasting advantages.
Context collapse: Even when content is debunked, the correction rarely reaches everyone who saw the original.
Liar's dividend: The existence of deepfakes means any inconvenient real footage can be dismissed as fake.

What Detection Tools Can Do

Identify known generation artifacts (but these become obsolete)
Analyze metadata for signs of manipulation
Cross-reference against original source material (when available)
Flag statistical anomalies in audio or video

No detection tool is reliable enough to serve as a definitive arbiter of authenticity.

What Comes Next

The 2024-2025 period was a testing ground. Nation-states and political operatives learned what works, what gets caught, and how platforms respond. Several trends are emerging:

Hybrid Operations

The most effective operations combine AI-generated content with traditional tactics. A deepfake clip gets attention; organic supporters share it; traditional media reports on it; the synthetic origin becomes obscured in the noise.

Targeting AI Systems

Rather than creating content for human consumption, operators are increasingly targeting AI systems themselves, poisoning training data, manipulating chatbot knowledge bases, and exploiting AI-powered recommendation systems.

Personalized Disinformation

AI enables personalization at scale. Rather than creating one deepfake for mass distribution, future operations may generate thousands of variations tailored to individual targets based on their psychological profiles, political leanings, and information consumption patterns.

Audio Over Video

Audio deepfakes are currently more convincing than video and much cheaper to produce. The New Hampshire robocalls suggest a shift toward synthetic audio as the primary vector, phone calls, voice messages, and podcast-style content that doesn't require visual fidelity.

Protecting Yourself

Individual voters can't single-handedly solve the AI disinformation problem. But you can reduce your vulnerability:

Before Sharing

Verify the source appears on the original creator's official channels
Check if fact-checkers have reviewed the content
Wait, breaking news is often wrong news. Verification takes time.
Ask: "Who benefits if I believe and share this?"

Red Flags for AI Content

Extraordinary claims without corroborating sources
Timing designed to prevent response (election eve, late-night drops)
Content that perfectly confirms your existing beliefs
Audio or video quality that seems inconsistent with the claimed source

Media Literacy

Understand that AI content exists and is being used in elections
Don't share content just because it makes you feel something
Recognize that your emotional reaction may be exactly what operators are trying to trigger

The Bottom Line

AI-powered election disinformation in 2024-2025 was real but limited. Romania annulled an election. New Hampshire voters received fake Biden calls. Slovakia's election may have been influenced by synthetic audio released at a calculated moment. These aren't hypotheticals, they happened.

But the AI apocalypse predicted by many experts didn't materialize. Traditional "cheap fakes" remained more common than sophisticated deepfakes. Detection held up better than feared. Legal consequences for perpetrators are beginning to emerge.

The danger isn't that AI disinformation destroyed democracy in 2024. The danger is that 2024 was a dress rehearsal. Nation-states learned what works. Operators tested techniques. Detection gaps were identified. The infrastructure for larger operations is being built.

What we saw was the beginning, not the end.