How AI is going to influence elections

I have to start with a disclaimer.

I’m not a political scientist. I’m not an election expert. And I’m definitely not a fortune teller.

I build software, and these days a lot of that software has AI in it. So one could reasonably argue that I have no business writing about elections.

Fair enough.

But I’ve spent enough time with LLMs to know how they behave when nobody is checking their work. And elections are exactly where AI’s worst habits stop being funny.

As I write this, Brazil is voting. More than 150 million people are eligible to cast a ballot in what has been called the first Brazilian election where generative AI is cheap, everywhere, and used at scale by ordinary people. The US midterms are a month away.

So, how is AI going to influence elections?

Short answer: not the way the headlines keep promising.

Long answer: yes, a long text is incoming. Again.

1. The fake doesn’t need to be good. It needs to be on time.

Two days before Slovakia’s 2023 parliamentary election, an audio clip showed up on Facebook. In it, Michal Šimečka, the leader of Progressive Slovakia, appeared to be discussing how to rig the election with a journalist.

It was fake. Šimečka immediately denounced it, and AFP’s fact-checkers said it showed signs of being manipulated using AI.

But it was posted during Slovakia’s 48-hour pre-election moratorium, when the media and politicians are supposed to stay quiet. So the people best placed to debunk it were, by law, the ones least able to talk.

Šimečka had been leading in the polls. He lost.

Did the clip cost him the election? Honestly, nobody can prove that, and the researchers who looked at it closely argue the story is a lot more complicated than “the first election swung by a deepfake.”

But notice the trick. The fake wasn’t impressive. The timing was.

In January 2024, two days before the New Hampshire primary, thousands of Democrats got a robocall in an AI-generated voice that sounded like Joe Biden. It even used his catchphrase—“What a bunch of malarkey”—and, according to prosecutors, suggested that voting in the primary would stop them from voting in November.

The political consultant behind it admitted he orchestrated it. A jury acquitted him of all charges anyway. He still faces a $6 million fine from the FCC, which he has said he won’t pay.

Think about that for a second. A fake president told real voters to sit out a real primary, and the man who admitted to arranging it walked out of court not guilty on every count.

Closer to home, before Pakistan’s February 2024 election, a video circulated in which Imran Khan’s voice appeared to announce that his party, PTI, was boycotting the vote. Fact-checkers found the audio had been made with an AI voice-cloning tool; neither Khan nor PTI had announced any boycott.

And in October 2025, days before Ireland’s presidential election, a fake RTÉ news bulletin—deepfaked presenters and all—showed candidate Catherine Connolly announcing that she was withdrawing and that the election was off. It was viewed almost 30,000 times before Facebook took it down.

(She won anyway. With 63% of the vote.)

See the pattern?

None of these fakes tried to convince anyone to switch sides. They tried to convince people that voting was pointless, cancelled, or already decided.

You don’t need to change someone’s mind to change an election.

You just need them to stay home.

2. Campaigning from the grave, and from prison.

Not every synthetic politician is a scam, though. Some of them are the official campaign.

During India’s 2024 election, parties spent an estimated $50 million on authorized AI-generated content. M. Karunanidhi, who died in 2018, appeared in an AI video at his party’s youth conference—signature yellow scarf, dark glasses and all. An AI-generated audio clip of J. Jayalalithaa, who died in 2016, was posted by her party’s official account.

I’m fairly sure no ethics textbook had a chapter on getting consent from people who have passed away. Until very recently, it didn’t need one.

Then there’s Imran Khan.

Khan was in prison during that same 2024 campaign. According to reporting at the time, his party took notes he passed to his lawyers, turned them into speech with a voice-cloning tool, and played the result at an online rally—with a caption saying the voice was AI-generated. After the results came in, an AI-voiced victory speech followed.

And before anyone asks: this isn’t an essay about Pakistani politics, and I’m keeping it that way. I’m only interested in what the technology did.

And the technology question is genuinely interesting. Here, AI let a politician who couldn’t appear in person keep addressing his supporters—clearly labeled as AI. Around the same time, the same kind of tool was being used to fake a boycott announcement in his voice.

Same voice. Same technology. Opposite purposes.

Brazil has just had its own version. In July, Jair Bolsonaro—under house arrest and barred from public communication—appeared at his son’s campaign launch as an AI avatar that explicitly identified itself as a simulation. His opponents challenged it. His son’s campaign argued that a clearly labeled simulation isn’t a deepfake.

Which raises a question I don’t think anyone has a clean answer to yet: if everybody knows it’s fake, is it still a deepfake? Or is it just a very expensive puppet show?

Either way, AI here is a megaphone. And a megaphone doesn’t care who’s shouting into it.

3. Your most persuasive campaign volunteer is a chatbot.

This is the part that genuinely worries me.

In December 2025, two large studies were published on the same day—one in Nature, one in Science—looking at whether a conversation with an AI could change how people vote.

It can.

In the Nature study, people chatted with AI models arguing for one of the two leading candidates in the 2024 US presidential election, the 2025 Canadian federal election, and the 2025 Polish presidential election. In the US, preferences moved by a few points. In Canada and Poland, the bots shifted opposition voters by around 10 percentage points. That’s a bigger effect than researchers typically see from traditional video ads.

Ten points.

Elections get decided by a lot less than that.

And how did the bots pull it off? No dark psychological tricks. Mostly, they were polite and they used facts and evidence. When the researchers stopped the models from using facts, they became far less persuasive.

Reassuring, right? Facts win!

Well. Not all of the facts were accurate.

The Science study went even bigger: nearly 77,000 people in the UK, 19 different language models, and close to half a million claims checked by fact-checkers.

The most persuasive models were the ones that packed their answers with as many fact-checkable claims as possible. Training a model specifically to persuade made it up to 51% more persuasive.

And then came the punchline: the more persuasive a model was, the less accurate its information tended to be.

In other words, the winning strategy is to bury people in facts until they agree with you. Whether the facts are true is apparently optional.

Which, come to think of it, is also how a lot of LinkedIn posts work.

If you read my review of AI Engineering, you know I have a thing about this. LLMs will happily fill gaps with completely made-up information, and they won’t mention that they’re guessing unless you corner them.

Now imagine a system that is rewarded for winning the argument, not for being right.

It doesn’t get tired. It doesn’t need a salary. It can have a patient, personal, one-on-one conversation with every voter in the country, at 2 a.m., in their own language.

Honestly, it’s the perfect campaign volunteer.

Just one that occasionally makes things up.

To be fair, the same capability works the other way too.

In a 2024 study, also in Science, people who believed in conspiracy theories—including ones about the 2020 US election—talked them through with a GPT-4 chatbot that answered their specific evidence. Their belief dropped by about 20%, and the effect was still there two months later. A professional fact-checker found that 99.2% of the chatbot’s claims were accurate.

So persuasion isn’t evil in itself. It’s a capability.

What matters is who is pointing it, and at what.

4. Please don’t ask a chatbot who to vote for.

I mean it.

Ahead of the Dutch elections in October 2025, the Dutch Data Protection Authority tested ChatGPT, Gemini, Grok and Le Chat as voting advisers. In more than 56% of cases, they pointed people toward the same two parties—one on the far right, one on the left. One chatbot did it more than 80% of the time. Most parties barely showed up at all.

The regulator’s advice was simple: don’t use chatbots for voting advice.

Even purely factual questions are a problem. In January 2024, election officials and AI experts tested five leading models on basic US voting questions. Half of the answers were rated inaccurate. Around 40% were rated harmful—including answers that sent people to polling places that don’t exist.

In that review, I told the story of ChatGPT confidently giving me the wrong tag for an API. A wrong API tag breaks a build.

A wrong polling station breaks something else entirely.

And remember the confirmation-bias problem? These models love agreeing with your premise. Ask one why your candidate is obviously the right choice and see how hard it pushes back.

Then ask it the opposite.

Brazil’s electoral court has gone as far as banning AI systems from recommending or prioritizing candidates. Whether that’s enforceable is another story. But at least someone noticed.

5. “That’s probably AI.”

Here’s the part I think we’re underestimating the most.

Law professors Bobby Chesney and Danielle Citron have a name for it: the liar’s dividend. The more people learn that anything can be faked, the easier it becomes to dismiss something real as fake.

It’s already happening. In September 2025, when a video of objects being thrown out of a White House window went viral, the US president said it was “probably AI-generated”—hours after White House officials had told reporters it showed a contractor doing maintenance.

He then added, and I promise I’m not making this up:

If something happens really bad, just blame AI.

Whether it was meant as a joke or not, that one line explains the liar’s dividend better than any definition I could write.

Now think about what it does to an election.

For about a century, a recording was evidence. Now a recording is just… a recording. In Hungary this year, about half the people who came across deepfakes said they weren’t sure whether what they were watching was real.

The danger isn’t that you’ll believe the fake.

It’s that you’ll stop believing anything.

And a voter who believes nothing is, in a way, the easiest voter to manage. You don’t need to convince them of anything. You just need them to be tired.

6. Okay, but does any of this actually work?

Hungary is probably the best test case we have so far.

Its April 2026 election was soaked in AI to a degree no other EU country had seen. Deepfakes showing opposition leader Péter Magyar as a puppet of foreign leaders. AI-generated war scenes. A fake, AI-written opposition program promising things like extra taxes on pets.

According to the Reuters Institute, a company run by an ally of the ruling Fidesz party became the biggest spender on political ads on YouTube in the EU during the summer of 2025, and its AI-generated videos attacking Magyar were shown an estimated 650 million times across YouTube and Facebook.

In a country of fewer than ten million people.

Fidesz lost. In a landslide. Magyar’s Tisza party won a two-thirds majority, on record turnout.

And in one survey, 90% of respondents said using these tools in politics was entirely unacceptable.

So… point for humans, I guess?

I’m not going to pretend one election proves anything. Hungarians had plenty of reasons to vote the way they did that had nothing to do with AI. And the AI didn’t stop being harmful just because it lost.

But it does line up with what researchers keep finding: the people consuming deepfakes are mostly the ones who already agreed with them. AI disinformation reinforces beliefs far more than it changes them. It’s an amplifier, not a mind-control ray.

Or, as data people love to say: garbage in, garbage out. AI doesn’t create a polarized, distrustful electorate. It takes whatever it’s given and turns the volume up.

7. The boring stuff will matter most.

If I had to bet, I’d say AI’s biggest effect on elections won’t be one viral deepfake. It’ll be the boring stuff.

Take “Dona Maria,” an AI-generated political influencer in Brazil. She was made by a 37-year-old Uber driver using Google’s AI video tools. Her profile is estimated to have reached 100 million people, and at least 12 of her videos passed a million views each. In April, three parties in President Lula’s coalition sued to get her profiles suspended, arguing that they didn’t clearly disclose that she isn’t real.

One guy. A phone. A couple of AI tools. A hundred million people.

Propaganda used to need a studio, a budget, and a team. Now it needs a free afternoon.

Multiply that by every ad, every audience, every language, and every week of a campaign. None of it makes headlines the way a fake Biden phone call does. But it adds up.

And, to be fair again, not all of it is bad. In Japan, an AI engineer named Takahiro Anno founded a party, Team Mirai, that builds tools like a web service using AI to translate complex bill language into plain Japanese. It won its first seat in 2025 and 11 seats in the lower house in February this year.

Translation, accessibility, explaining a 90-page manifesto in plain language—AI can genuinely make politics easier to follow.

Honestly, when you put it all in a list, AI in politics sounds a lot like… some politicians I could name. But won’t:

• →Tells you what you want to hear.
• →Sounds confident without any sources.
• →Says something different depending on who’s asking.
• →Only admits it was wrong once you’ve cornered it.

Maybe that’s why it fits in so well.

Okay, here’s the actually useful part.

• →If something tells you not to vote, that the vote is cancelled, or that it’s already been decided—especially in the last few days—assume it’s fake until your election authority confirms it. In Pakistan, that’s the Election Commission of Pakistan.
• →If a clip makes you furious, wait before you share it. Your anger is the product.
• →Don’t ask a chatbot who to vote for. If you do it anyway, ask it for sources. Then actually check them.
• →“That’s probably AI” is not an argument. Not when a politician says it, and not when you do.

In my review of AI Engineering, I said that AI is only as useful as the person using it.

I think elections work the same way. AI is only as dangerous as we are distracted, tired, and willing to stop checking.

AI won’t take your vote away.

But it might make you too confused, too cynical, or too exhausted to use it.

Don’t let it.

For the record: this piece doesn’t support or oppose any party, candidate or government—in Pakistan or anywhere else. The people and parties in it are here for what their examples show about AI, not for who they are. Every factual claim links to where it came from.

And, as always, don’t treat this as a source of truth either. The sources are linked in the text and listed below. Check them.

It’s just my opinion.