Screenshot from the real version of one of the videos that the team used in their study, showing Amelia Palmer, a fictional vegan influencer. Credit: Clark & Lewandowsky.
Generative deep learning models are artificial intelligence (AI) systems that can create texts, images, audio files, and videos for specific purposes, following instructions provided by human users. Over the past few years, the content generated by these models has become increasingly realistic and is often difficult to distinguish from real content.
Many of the videos and images circulating on social media platforms today are created by generative deep learning models, yet the effects of these videos on the users viewing them have not yet been clearly elucidated. Concurrently, some computer scientists have proposed strategies to mitigate the possible adverse effects of fake content diffusion, such as clearly labeling these videos as AI-generated.
Researchers at University of Bristol recently carried out a new study set out to better understand the influence of deepfake videos on viewers, while also assessing user perceptions when AI-generated videos are labeled as "fake." Their findings, published in Communications Psychology, suggest that knowing that a video was created with AI does not always make it less "persuasive" for viewers.
"This paper grew out of a simple but slightly uncomfortable observation," Simon Clark, co-author of the paper, told Phys.org. "Most of the public and policy discussion about deepfakes assumes that the main problem is deception, i.e., the idea that people will be misled into believing something that isn't true. The dominant response has therefore been transparency, on the assumption that labeling AI-generated content will enable people to disregard its content. From a psychological perspective, that assumption is not well supported."
Screenshot from the deepfake version of one of the videos that the team used in their study, showing Amelia Palmer, a fictional vegan influencer. Credit: Clark & Lewandowsky.
Some past psychological studies suggest that people can be influenced by information they are presented with even if they know it to be false, unreliable, or irrelevant. Clark and his colleague Stephan Lewandowsky wondered whether this pattern also applies to the deep fake videos that people watch online.
"We asked ourselves: what happens when people encounter a deepfake video that has already been identified and clearly flagged as fake?" said Clark.
"The primary aim of our paper was to answer this question. We specifically examined whether transparency warnings reduce the influence of deepfake videos, and whether they eliminate that influence entirely, as policy discussions often implicitly assume.
Assessing people's perceptions of deepfake videos
To assess people's perceptions of "flagged" deepfake videos, Clark and Lewandowsky carried out three online experiments involving 175, 275, and 223 participants, respectively. The participants were asked to watch short videos that showed a person admitting that they had committed a crime or a moral transgression.
"Some of the videos were real, and others were AI-generated deepfakes that we created specifically for the study, using the same scripts and settings," explained Clark.
"Crucially, in some conditions, participants were shown a warning before watching the video, stating that it had been identified and flagged as a deepfake, while other participants saw no warning, or a generic warning about the existence of deepfakes. We also added a control condition, in which participants watched a version of the video with all relevant information obscured by background sounds."
After watching each of the videos, participants were asked a series of simple questions. They were asked whether they thought the video was fake and if they believed the person was guilty of a crime or moral transgression.
"We also asked them to explain their answers in their own words, which allowed us to better understand whether people were relying on the warning, the content of the video itself, or both," said Clark.
Shaping new measures against deepfake content
When they analyzed the participants' responses, the researchers found that most participants were still influenced by the content of a video, even if they were warned that it was fake and generated with AI. Interestingly, many participants acknowledged that a video was fake but still judged the person in the video as guilty of the crime or transgression that they confessed to.
"Warnings do reduce the effect, but they do not eliminate it, even among participants who say they believe the warning and know the video is a deepfake," said Clark.
"We also showed that different types of warnings have different effects. Specific warnings referring to a particular video are more effective at convincing people that a video is fake, whereas generic warnings do not reliably do so. At the same time, generic warnings can still alter how people interpret a video's content, which has implications for how such warnings are used."
In their analyses, the researchers looked at both yes/no responses and more detailed explanations provided by participants. This allowed them to gain insight into how people reasoned about deepfake videos that are "flagged" as AI-generated.
"Many participants continued to engage with the video itself rather than simply deferring to the warning, which helps explain why transparency reduces influence without fully neutralizing it," explained Clark. "This has clear implications for regulation, where disclosure is often treated as a sufficient safeguard."
The results of this recent study could soon inspire other research teams to further investigate the impact of warnings on how users perceive fake videos. These efforts could ultimately inform the development of more effective measures to mitigate the disinformation and confusion that can arise from the diffusion of realistic deepfake content.
"One priority in our future research will be to better understand the psychological mechanisms behind this effect," said Clark. "If people continue to rely on a video even when they know it's fake, then we need to understand what drives that reliance, and how it differs from simple belief in the video's accuracy."
In their next studies, Clark and Lewandowsky also plan to investigate whether the "source" and "framing" of deepfake warnings influence their effectiveness. In their recent experiments, the warnings were issued by the social media platform on which they were posted, yet they would also like to explore what would happen if warnings were attributed to journalists, independent fact-checkers, legal authorities, or other sources.
"More broadly, I am interested in moving beyond the question of belief to the question of meaning," added Clark. "Even when people reject a deepfake as false, it may still shape impressions, moral judgments, or feelings about a person. Understanding those more subtle forms of influence will be essential for assessing real-world harm, and for designing responses that go beyond transparency alone."
Written for you by our author Ingrid Fadelli, edited by Gaby Clark, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you.
