If you were scrolling through social media a few years ago, you may have come across a video of a politician or celebrity that looked like a detailed video game animation. It might have appeared real at first glance, but on closer inspection, the mouth moved unnaturally, or the person had six fingers.
These were common signs the content was a deepfake — an AI-generated image, video or audio manipulated to replicate a real person or event, often to spread misinformation.
Those tells are now a lot harder to detect, according to a new study from Utah Valley University's Center for National Security Studies.
Researchers gave an online survey with a mix of deepfakes and real content to over 600 test subjects. It found that only 16% could successfully differentiate between a real video and an imitation. The rest either got it wrong or were unsure.
“This is significantly less than our study that we did only two years ago,” said UVU researcher Kaye Banner in a presentation of the findings.
The improving technology has implications for elections and politics. Part of the survey consisted of real and AI-generated political information, like a ballot initiative.
Banner said deepfake media influenced the opinions of potential voters “just as much” as the real media used in the survey.
“We found no statistically significant difference in the opinion changes of people who saw a real versus a synthetic video,” she said, “this shows just how successful deepfake disinformation operations could be.”
In 2024, New Hampshire voters received AI-generated robocalls impersonating former President Joe Biden, telling them not to vote in the primaries. That same year, tech trillionaire Elon Musk shared an AI-generated video of former vice president and presidential candidate Kamala Harris calling herself a “diversity hire.” President Donald Trump has also frequently posted AI-generated images and videos of himself on social media.
The technology has surfaced in Utah politics, too. During the 2024 governor's race, a video appeared to show Gov. Spencer Cox admitting to forging signatures in a previous election he won. Though quickly debunked, it intensified a campaign already dominated by Phil Lyman's allegations of election fraud. It also helped spur the partnership that led to the UVU study.
Brandon Amacher, the director of the university’s Emerging Tech Policy Lab for the Intermountain Intelligence, Industry and Security Consortium, said the technology used to make such content has only gotten better and more accessible.
“Our students built deepfakes convincingly enough to get these results in a few days,” he said. “If a small university lab can do that, this is no longer a capability that is constrained to highly sophisticated malicious actors. It's available to anyone.”
Amacher explained that there are several misconceptions about who is vulnerable to deepfake technology. It’s no longer just older generations or the less technologically savvy falling victim.
“Democrats, Republicans and independents all detected defects at nearly identical, and I will add, uniformly poor rates. Somewhere between 15% and 19%, and older and younger participants performed at about the same rate,” he said.
Despite the current administration's vocal push to advance AI technology, Utah is making moves to regulate it.
A new Utah law requires websites and social media platforms to remove non-consensual AI-generated explicit images. Once reported, platforms have 48 hours to make reasonable efforts to take the content down. The law’s focus is on “intimate” images, but it adds to the governor’s calls for more state regulation.
Amacher said he believes one of the most crucial takeaways is that anyone can make a convincing deepfake. He emphasised that his study used publicly available tools on an “old student-owned laptop.”
“The first misconception is that this is a fringe problem, something rare and isolated, which has not or could not actually impact voter opinion or the outcome of an election. The data does not support that.”
