Political deepfakes have moved from a technical warning to an election-year problem. With billions of people voting in high-stakes races in more than 50 countries, synthetic images are now part of the information environment that voters, platforms and election officials have to navigate.
A study from the Center for Countering Digital Hate (CCDH) found that AI-generated election deepfake images on X, formerly Twitter, rose by an average of 130% per month over the past year. The finding points to a larger issue: image generation tools are widely available, and social platforms are struggling to keep misleading political content from spreading.
How election deepfakes became easier to make
Deepfakes were already growing quickly before election-related examples became a mainstream concern. Research cited by the World Economic Forum found that deepfakes grew 900% between 2019 and 2020. Sumsub, an identity verification platform, observed a 10x increase in the number of deepfakes from 2022 to 2023.
The newer challenge is not only volume. It is credibility. Generative image tools have improved enough that synthetic election content can look persuasive to ordinary users, especially when it is shared quickly and without clear context. In a 2023 University of Waterloo study, only 61% of people could tell the difference between AI-generated people and real ones.
Public concern has followed. In a YouGov poll, 85% of Americans said they were very concerned or somewhat concerned about misleading video and audio deepfakes. A separate survey from the Associated Press-NORC Center for Public Affairs Research found that nearly 60% of adults think AI tools will increase false and misleading information during the 2024 U.S. election cycle.
What the CCDH study found on X
The CCDH study focused on election-related deepfake images on X. The researchers examined community notes, the user-contributed fact-checks added to potentially misleading posts, from a public X repository covering February 2023 to February 2024.
They searched for notes that used terms connected to images and AI, including words such as “image,” “picture” or “photo,” along with variations of terms such as “AI” and “deepfake.” The study did not measure election deepfakes on Facebook, TikTok or other platforms, so its findings describe one important channel rather than the full internet.
The researchers said many of the deepfakes identified on X appeared to come from four image generators:
- Midjourney
- OpenAI’s DALL-E 3 through ChatGPT Plus
- Stability AI’s DreamStudio
- Microsoft’s Image Creator
To test how easily those tools could create election deepfakes, the researchers developed 40 text prompts tied to the 2024 U.S. presidential election and ran 160 tests across the generators. The prompts included false scenes involving candidates and false scenes involving voting or election procedures.
The researchers also tested whether small wording changes could bypass safeguards. For example, they tried preserving a prompt’s meaning while replacing a direct candidate name with a description such as “the current U.S. president.”
Safeguards blocked some images but not enough
The image generators produced deepfakes in 41% of the tests, according to the CCDH study. That result came even though Midjourney, Microsoft and OpenAI had specific policies against election disinformation. Stability AI’s DreamStudio prohibited “misleading” content, but the source article notes that its policy did not specifically cover content that could influence elections, hurt election integrity or feature politicians or public figures.
The tools did not behave identically. Midjourney generated election deepfakes in 65% of the test runs. Image Creator did so in 38%, DreamStudio in 35% and ChatGPT in 28%.
Some guardrails worked in narrower categories. ChatGPT and Image Creator blocked all candidate-related images. But the study found that all the generators created deepfakes depicting election fraud and intimidation, including scenes such as election workers damaging voting machines.
Company responses varied. Midjourney CEO David Holz said Midjourney’s moderation systems are “constantly evolving” and that election-related updates were “coming soon.” OpenAI said it was “actively developing provenance tools” to help identify images made with DALL-E 3 and ChatGPT, including tools using C2PA. Stability AI said DreamStudio’s terms prohibit “misleading content” and pointed to measures such as filters, watermarking technology, and work on “provenance and authentication.” Microsoft did not respond by publication time.
Why social platforms matter just as much
Creating a political deepfake is only one side of the risk. The other is distribution. Social media can move an image far beyond the original post, and fact-checking systems do not always appear to follow the same media everywhere it appears.
The CCDH study highlighted an AI-generated image of Donald Trump attending a cookout. One post containing the image was fact-checked, but other posts containing the same image were not, and those other posts received hundreds of thousands of views.
X says community notes on a post automatically appear on posts containing matching media. But the CCDH study found that this did not always seem to happen. BBC reporting also found that deepfakes of Black voters encouraging African Americans to vote Republican gained millions of views through reshares even though the originals had been flagged.
That gap matters because political deepfakes do not need to persuade everyone. They can still confuse, distract or reinforce false claims when they circulate at scale, especially around voting, election fraud or intimidation.
The response is still taking shape
There is no simple fix in the source article. Hood and the CCDH co-authors argue that AI tools and social platforms need stronger safeguards, more testing before launch, and deeper collaboration with researchers. They also call for social media companies to invest in trust and safety staff focused on generative AI, disinformation and election integrity.
The source article also describes several steps already underway. Microsoft, OpenAI and Stability AI were among image generator vendors that signed a voluntary accord on responding to AI-generated deepfakes intended to mislead voters. Meta said it would label AI-generated content from vendors including OpenAI and Midjourney ahead of elections and barred political campaigns from using generative AI tools, including its own, in advertising. Google said political ads using generative AI on YouTube and other platforms such as Google Search would need a prominent disclosure when imagery or sounds are synthetically altered.
X, after reducing headcount including trust and safety teams and moderators following Elon Musk’s acquisition of the company over a year ago, said it would staff a new “trust and safety” center in Austin, Texas, with 100 full-time content moderators.
Policy efforts are also uneven. The source article says no federal law bans deepfakes, while 10 states around the U.S. have enacted statutes criminalizing them. Minnesota’s was the first to target deepfakes used in political campaigning.
The core problem remains speed. AI image tools can produce convincing political content quickly, and social media can spread it before labels, notes or takedowns catch up. As Hood put it, “It’s incumbent on AI platforms, social media companies and lawmakers to act now or put democracy at risk.”