In 2023, there were about 500,000 deepfakes on the internet. By 2025, that number reached 8 million — a sixteenfold increase in two years. Creating a convincing fake video, a cloned voice, or a fabricated news article is no longer a special-studio matter but a matter of minutes and near-zero cost.
This is not a technological novelty — it is a systemic test for the information environment. The question is no longer "do deepfakes exist" but "how do we live with them."
Scale: what the numbers say
The growth rate is astonishing. Estimates of online deepfake volume (which the UK government also relies on): ~500,000 in 2023, ~8 million in 2025. Video and audio deepfakes pose a particular danger because viewers trust footage more than text, and research on synthetic voice fraud shows people cannot distinguish a cloned voice from a real one in a short clip.
According to NewsGuard's monitoring, as of March 2026 at least 3,006 AI-generated "content farm" sites were identified in 16 languages — a 44% increase in just five months from 2,089 in October 2025. Creating a false news article that looks credible, with an author byline and archive photo, now takes minutes, while a real newsroom needs staff and time.
Per Sumsub's Identity Fraud Report 2025–2026 (analysis of over 4 million fraud attempts in 2024–2025), the share of multi-stage attacks in identity fraud rose from 10% to 28%. 40% of surveyed companies said they had encountered fraud, 52% of end users said they had been victims; 75% of respondents expect fraud to move under AI control.
On the political front, cases are being documented too: per research by the International Panel on the Information Environment (IPIE), 215 cases of AI-generated deepfake election disinformation were recorded at competitive elections in 50 countries in 2024. In Canada's 2025 federal election, 5.86% of election-related images circulated were found to be AI-generated. In an Adobe survey in Australia, 69% of respondents said they were worried about malicious deepfakes influencing elections, and only 12% were fully confident they could detect them.
Three fronts: politics, money, and the information environment
The deepfake threat manifests in three directions — each with real examples.
Politics. In Ireland, a malicious deepfake video spread showing presidential candidate Catherine Connolly supposedly "withdrawing" from the race — an official complaint was filed with the Electoral Commission. In Indonesia, a wave of deepfake fraud in the name of President Prabowo Subianto promising financial aid covered 20 provinces (WEF Global Cybersecurity Outlook 2026). In Honduras, at the 2025 elections, the EU observer mission documented at least 11 sophisticated deepfakes designed to mislead voters — including materials imitating a national newspaper's design with a cloned voice of an electoral council member.
Money. In financial fraud, voice cloning is the fastest-growing weapon: a call from a "CEO" or a "relative" now sounds convincing. In the World Economic Forum's Global Cybersecurity Outlook 2026 report, 94% of respondents named AI the biggest change factor in cybersecurity next year. The FBI's Internet Crime Complaint Center (IC3) in 2025 circulated warnings about fraudsters using fake social profiles, cloned voices, and convincing videos featuring celebrities.
The information environment. An important balance is needed here: Brookings researchers, studying fact-checked disinformation in the 2024 US election, found that less than 1% of it was confirmed as AI-generated. So today's picture is closer not to "AI is inventing new lies to swing elections" but to "AI cheaply multiplies low-quality, false or misleading content, crowding out trustworthy information and making verification harder for everyone."
The World Economic Forum assesses this wave as follows:
"AI is accelerating the scale and complexity of cyber-enabled harm — demanding stronger verification standards, cross-platform coordination, and protections for vulnerable groups."
Why people fall for it: the psychology
Technology is one thing, human psychology another. Research shows several stable vulnerabilities.
First, we trust the frame. Video and audio feel more "documentary" than text — it's hard to doubt what you see with your own eyes. Second, modern cloning is so good that in a short clip a synthetic voice is indistinguishable, while independent detection tools still can't achieve stable accuracy against the newest models.
Third, distribution mechanics work in favor of fakes: research shows that the most realistic fabrications, despite low spread frequency, get disproportionately high engagement. That is, a small number of highly convincing fakes reach a wide audience. Fourth, fraudsters exploit urgency and strong emotions — panic, anger, or greed rob people of the chance to "think before sharing."
Fifth, the "a familiar person sent it" effect: we evaluate not the information but the person who sent it. A message forwarded by a trusted friend automatically gains trust — even though they didn't verify it either. Through this chain, lies spread geometrically, while refutations always arrive late and reach fewer people.
Protection: seven rules of media literacy
The good news is that protection is not expensive technology — it's habits.
1. Check the source. Trust not the first place you saw the sensational news, but the official channel: the government agency's official page, the company's verified account, a reputable news agency. If a candidate "withdrew," it should be on their official page.
2. Use reverse search. Check a frame of a suspicious image or video in search engines — was it used before in another context, how long has it been circulating? This one-minute habit exposes most fakes. Listen to the video's audio separately: cloned speech may have unnatural pauses, intonation jumps.
3. Confirm in at least two or three independent sources. A message in one Telegram channel is not yet news. Have reputable publications written about it independently of each other?
4. Don't give in to emotions. "Share urgently!", "Everyone must know!" — such calls are a classic sign of fraud. Before sharing a message that triggered strong emotions, wait at least a few minutes and verify.
5. Agree on a family "password." The simplest protection against voice-cloning fraud: agree in advance with family members on a secret word or question for emergencies. To a call saying "Mom, I'm in trouble, send money," ask for the password.
6. Pay attention to content labels. The industry is introducing origin-verification standards like C2PA — in the future, trustworthy content will carry metadata of "who, when, how created it." For now, the absence of such a label is a reason to be cautious.
7. Don't spread — report. Don't spread suspicious content even "to warn" — for the algorithm, any spread is a signal of "interesting." Use the platform's complaint function.
