The past few years have brought an extraordinary shift in how digital content is created. Videos and images that once required studios, actors, and expensive equipment can now be produced by generative deep learning models that run on a laptop. These systems can fabricate a person’s face, voice, and gestures with such precision that the results often look indistinguishable from real footage. This technological leap has opened remarkable creative possibilities, yet it has also created a new kind of vulnerability. A recent study from the University of Bristol, published in Communications Psychology as The continued influence of AI-generated deepfake videos despite transparency warnings and summarized in a plain English article on Phys.org, asks a deceptively simple question: What happens when people watch a deepfake video that they already know is fake?[1]
The answer turns out to be more unsettling than many policymakers assume. The researchers found that even when viewers are explicitly told that a video is a deepfake and even when they say they believe that warning, the video's content still shapes their judgments. People continue to treat the confession, accusation, or moral slip-up depicted in the video as meaningful, even though they know it never happened. This finding challenges the dominant policy response to deep fakes, which has focused heavily on transparency labels. If knowing something is fake does not stop it from influencing us, then labeling alone cannot be the solution.
To understand why this matters, it helps to step back and consider how generative models became so convincing in the first place. These systems learn by analyzing millions of real examples, identifying subtle patterns in how faces move, how light reflects on skin, and how voices shift with emotion. Over time, the model becomes so good at predicting what a “real” frame should look like that it can generate new ones that fit seamlessly into a video. The result is like a master forger who has studied every brushstroke of a famous painter. Once the forger has internalized the style, they can produce a painting that fools even trained eyes. The problem is not only that forgery looks real but that it feels real, and our brains are wired to trust what we see.
This is why the psychological impact of deepfakes is not yet fully understood. Humans evolved to treat visual evidence as one of the most reliable forms of information. When we see someone confessing to a crime or violating their own moral code, our instinct is to treat that as meaningful, even if another part of our mind knows it is fabricated. The Bristol researchers suspected that this instinct might override the warning label, and their experiments were designed to test exactly that.
Across three studies involving 175, 275, and 223 participants, the team created both real and AI-generated videos. In one set, a fictional local government official appeared to admit to taking a bribe. In another, a vegan social media influencer confessed to secretly eating meat. Some participants saw the real video, some saw the deepfake, and some saw a control version where the incriminating audio was obscured. Before watching, participants either received no warning, a generic warning about the existence of deepfakes, or a specific warning stating that the video they were about to watch had been identified as a deepfake.
After viewing the video, participants were asked two simple questions. First, did they think the person in the video was guilty of the crime or moral transgression? Second, did they think the video itself was a deepfake? They were also asked to explain their reasoning in their own words, which allowed the researchers to see whether people were relying on the warning, the video content, or both.
The results were both significant and unsettling. Warnings reduced the influence of the deepfake, but did not eliminate it. Many participants acknowledged that the video was fake yet still judged the person as guilty. In other words, they treated the content as meaningful even while rejecting its authenticity. Specific warnings were more effective than generic ones at convincing people that the video was fake, but even specific warnings did not fully neutralize the video’s persuasive power. Generic warnings, meanwhile, did not reliably increase skepticism about the video’s authenticity, though they sometimes altered how people interpreted the content.
This pattern reveals an important aspect of human cognition. The main problem is not deception. The main problem is that people continue to rely on information even after it has been conclusively proven false. Psychologists have long known that even lies and propaganda can leave a residue in memory, shaping impressions and judgments even after they have been exposed and the truth revealed. Deepfakes appear to operate in a similar way. Once a viewer has seen a person confess on screen, the emotional and visual impact of that moment lingers, even if the viewer knows it never happened.
This insight has implications far beyond deepfakes. It touches on a broader cultural shift in which desired beliefs increasingly outweigh objective facts. Public education systems in the United States and the United Kingdom have struggled to maintain a commitment to fact-based reasoning, often replacing it with ideological framing that encourages students to prioritize personal belief over empirical evidence. When people are taught to treat facts as optional, they become more vulnerable to persuasive fabrications, whether those fabrications come from political propaganda, social media rumors, or AI-generated videos. Reversing this trend will require broad-sweeping educational reform that restores critical thinking, scientific literacy, and intellectual humility as core competencies rather than optional virtues, and raises the bar for what is considered the minimum competency for educators.
The Bristol study offers a path forward. The researchers recommend moving beyond transparency alone and investigating how the source and framing of warnings influence their effectiveness. A warning issued by a social media platform may be interpreted differently than one issued by a journalist, a fact-checking organization, or a legal authority. They also emphasize the need to understand the psychological mechanisms behind the continued influence effect. If people rely on a video even when they know it is fake, then the goal should not be simply to label the content but to understand how meaning is constructed in the viewer’s mind.
These findings can help society place generative AI back into the role of augmenting technology rather than a threat. If we understand how people interpret AI-generated content, we can design systems that support healthy skepticism without undermining trust in legitimate information. We can also build educational programs that teach people not only how to spot deepfakes but how to use critical thinking to reason about them. The goal is not to make people paranoid about every video they see but to help them recognize that realism is no longer a guarantee of truth.
The real-world impacts of this research are significant. As deepfakes become more common in political campaigns, legal disputes, and online harassment, the assumption that labeling will solve the problem becomes increasingly untenable. Policymakers will need to consider additional safeguards, such as authentication systems for real videos, stronger penalties for malicious deepfake creation, and public awareness campaigns that focus on reasoning rather than fear.
Deepfakes are not going away, and transparency alone will not protect us. What will protect us is a deeper understanding of how people interpret information, how belief is formed, and how meaning persists even after truth has been clarified, and educational systems beginning at the elementary level that inculcate and teach critical thinking skills. The Bristol study is an important step in that direction and invites researchers, educators, and policymakers to rethink how we prepare society for a world in which seeing is no longer enough.
This article is shared at no charge for educational and informational purposes only.
Red Sky Alliance is a Cyber Threat Analysis and Intelligence Service organization. We provide indicators-of-compromise information via a notification service (RedXray) or an analysis service (CTAC). For questions, comments, or assistance, please contact the office directly at 1-844-492-7225 or feedback@redskyalliance.com
- Reporting: https://www.redskyalliance.org/
- Website: https://www.redskyalliance.com/
- LinkedIn: https://www.linkedin.com/company/64265941
Weekly Cyber Intelligence Briefings:
REDSHORTS - Weekly Cyber Intelligence Briefings
https://register.gotowebinar.com/register/5207428251321676122
[1] https://six3ro.substack.com/p/when-seeing-isnt-believing-why-deepfakes
Comments