Deepfakes and voice clones: what changed from 2023 to 2026

Series: How minds are moved · The modern machinery · 4 of 7

On 21 January 2024, two days before New Hampshire’s presidential primary, thousands of voters picked up the phone and heard Joe Biden’s voice telling them to stay home and “save” their vote for November. It was not Biden. It was a cloned voice, and the political consultant who commissioned it was later fined $6 million by the US regulator, then acquitted by a state jury. Four months earlier, a faked audio clip of a Slovak party leader had spread during the silent period before that country’s election. My AI agent and I went through what changed with deepfakes and voice clones between 2023 and 2026, what is documented, what the evidence says about their effect, and what detection can and cannot do for you.

Contents · 10 sections

TL;DR #

• Fake audio and video are now cheap, fast and good enough to fool people in real time: in 2024 a deepfaked video call cost the engineering firm Arup about US$25 million.
• The election cases are real but their effect on results is not shown: a fake audio clip hit Slovakia two days before its 2023 vote, and a cloned Biden voice called New Hampshire voters in 2024.
• Studies show people spot fake voices only about 73% of the time, and automated detectors lose roughly half their accuracy on real-world fakes.
• The defence that works is procedural, not visual: verify through a second channel you already trust.

What changed between 2023 and 2026 #

Until about 2022 a convincing fake video needed hours of footage, skilled work and patience. By 2023 a few seconds of someone’s voice were enough to clone it, and consumer tools could produce a talking face from a photograph. By 2024 criminals were using live deepfakes in group video calls. The technology did not get a new name. It got cheap.

Three cases mark the change.

Slovakia, September 2023: the audio that arrived during the silence #

Slovakia voted on 30 September 2023. Two days earlier, during the 48-hour pre-election moratorium when media and politicians are supposed to stay quiet, an audio clip spread on social media. It appeared to be Michal Šimečka, leader of Progressive Slovakia, discussing with the journalist Monika Tódová how to rig the election. Both denied it at once. AFP’s fact-checkers said it showed signs of AI manipulation.

The timing was the weapon. The moratorium made rebuttal hard, and because the fake was audio rather than video, it fell outside Meta’s manipulated-media rule at the time. Slovakia’s Interior Ministry centre for hybrid threats called it deceptive content with a significant impact on the credibility of the vote. Nobody has shown that it changed the result. Progressive Slovakia came second. Both statements are true at once, and that is typical of this field.

New Hampshire, January 2024: the cloned president #

Two days before the state’s 23 January 2024 primary, thousands of New Hampshire voters got a robocall with an AI-generated voice that sounded like President Biden, telling them to “save” their vote for November. A political consultant, Steve Kramer, admitted he had organised it. He said he did it to warn about the dangers of AI.

The record that followed:

• In February 2024 the US Federal Communications Commission ruled that AI-generated voices count as “artificial” voices under existing robocall law.
• In August 2024 the telecom carrier that transmitted the calls agreed to a $1 million penalty.
• In September 2024 the FCC fined Kramer $6 million, a civil penalty. He told the Associated Press he would not pay it.
• In June 2025 a New Hampshire jury acquitted him of all state charges of voter suppression and impersonating a candidate. His defence argued the primary was an unsanctioned “straw poll” and that the message never named Biden as a candidate.

So the regulator punished it, and a criminal jury did not. That is not a contradiction; they applied different laws.

Arup, 2024: the call that cost $25 million #

The election cases are about moving many people a little. Fraud is about moving one person once, and there deepfakes have clearly worked. In January 2024 an employee at Arup’s Hong Kong office joined a video call with what looked like the company’s chief financial officer and several colleagues, and then made 15 transfers worth HK$200 million, about US$25 million. Arup confirmed in May 2024 that fake voices and images had been used; police said the fakes were built from publicly available footage of real staff. The video call was not the attack; it was the proof that defeated the employee’s first doubt. An earlier post in this series, on deepfake fraud, takes that case apart step by step, so here it stands as the marker of what changed: real-time fakes in a group call.

What it feels like from the inside #

A deepfake does not feel like a trick. It feels like urgency from someone you know. The boss is in a hurry. Your child is crying on the phone and needs money now. A politician you dislike says exactly the shameful thing you always suspected. The content is built to make you act or share before you think, and to match what you already expect. That is why the most effective fakes are often short, low-quality and emotional, not cinematic.

What detection can and cannot do #

Humans. A 2023 study in PLOS ONE by Kimberly Mai and colleagues played real and synthetic speech to 529 listeners in English and Mandarin. They identified the fakes about 73% of the time, and training helped only a little. The authors noted their fakes came from older algorithms; newer ones are harder.

Machines. In 2025 researchers released Deepfake-Eval-2024, a benchmark built from fakes actually circulating on social media in 2024 across 52 languages. Open-source detectors that looked excellent on older academic tests lost about 50% of their performance (measured as AUC) on video, 48% on audio and 45% on images. Commercial tools did better but did not match trained forensic analysts. A detector score is evidence, not a verdict.

Labels and provenance. Since 2 August 2026, Article 50 of the EU AI Act requires people who publish deepfakes to disclose that the content is artificially generated or manipulated, and providers of generative systems must mark outputs in a machine-readable way (with a transition for systems already on the market). Content Credentials, the C2PA standard, attach signed metadata to images and video. The limit is plain: a screenshot or a re-upload can strip the metadata, and a missing label proves nothing. Honest creators will label; scammers will not.

Effects on elections. The Alan Turing Institute’s Centre for Emerging Technology and Security looked at the 2024 election year. It found no clear evidence that AI-generated content changed the result of any of the elections it studied, including the UK general election, where it identified only 16 confirmed viral cases. It warned of second-order harms: deepfakes mostly reinforce what people already believe, and they erode trust in real recordings. That second effect has a name, the “liar’s dividend”: once fakes are common, a real recording can be dismissed as fake.

So the honest balance is this. Deepfakes have clearly worked for fraud, where one person must be fooled once. For elections, where millions must be moved, the evidence of effect is thin so far, and the damage is mostly to trust.

How to notice it #

• The request is urgent, secret, or about money or credentials.
• The person asks you to skip the usual process “just this once”.
• The call or message comes through a new number, a new account, or a meeting link you did not expect.
• The clip appears right before a deadline: a vote, a transfer cut-off, a weekend.
• It confirms exactly what you already feared about someone.

What to do, in order:

1. Stop. Urgency is the attacker’s main tool.
2. Hang up and call back on a number you already had, not one given in the message.
3. Ask something only the real person knows, or use a family or team code word agreed in advance.
4. For money, require a second person to approve any unusual transfer.
5. For a political clip, wait a day and look for the original source and the person’s response.

Frequently asked questions #

Can I tell a deepfake by looking for glitches? #

Sometimes, but do not rely on it. Hands, teeth, blinking and lip sync used to give fakes away; newer tools fix most of that. Verification through a separate channel works whatever the quality of the fake.

Are deepfake detectors trustworthy? #

They are useful as one signal. On real-world fakes from 2024, open-source detectors lost about half their accuracy compared with lab tests. Treat a “real” verdict as “not detected”, not as “proven authentic”.

Did deepfakes decide any election? #

No study we found shows that. The Slovak and New Hampshire cases are documented attempts; their effect on results has not been demonstrated. The documented damage is to trust and to the speed of debunking.

Is it illegal to make a deepfake? #

It depends on where and why. Fraud and impersonation were already crimes. The EU now requires disclosure of deepfakes, and the US FCC treats AI voices in robocalls as artificial voices under the robocall law. Making a parody is not the same as using a fake voice to move money or votes.

The lesson #

The question is no longer “does it look real?” Everything can look real now. The question is “how did this reach me, and can I check it through a path I already trust?” For a voter, that means waiting a day for the original and the denial. For a company, a thirty-second callback on a known number is the cheapest security control there is.