When AI Impersonates Family: Protecting Seniors with Deepfake Awareness Education

One day, a video arrives from a family member. The face on the screen is unmistakably familiar, and the voice sounds exactly the same. They say they have been in an accident and urge you to send money immediately. But what if both the face and the voice were generated by AI? How calmly would we be able to verify what is real?

Oh, H., Lee, E., Shin, S., Lee, K., Pyun, M., & Lim, H. (2026). DeepAware: Using Experiential Deepfake Simulations to Enhance Cybersecurity Awareness in Older Adults. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, 1–23. https://doi.org/10.1145/3772318.3791492

When Familiar Faces and Voices Become Tools of Crime

AI-generated fake videos are no longer a problem limited to politicians and celebrities. With only a short voice recording and a few photographs, AI can imitate an ordinary person and be combined with crimes such as family impersonation scams, fraudulent investment solicitations, and fake news.

Contrary to the common image of older adults as a single group that is uniformly vulnerable to fraud, they are far from homogeneous. Some use digital devices with ease, while others feel anxious when faced with an unfamiliar screen or unfamiliar terminology. Simply warning people to “watch out for deepfakes” is therefore not enough. Education must help learners recognize the risk and carry out verification steps even when they are under emotional pressure.

Instead of Showing a Fake Video, Let Learners See Themselves as the Deepfake

DeepAware, developed by researchers at Seoul National University, is an experiential deepfake education platform designed for older adults. The researchers first interviewed five digital educators in South Korea to identify the difficulties older learners encounter and establish principles for the educational design. They concluded that older learners needed short, step-by-step instructions, scenarios closely connected to everyday life, repeated verification, and practical response strategies that would reduce fear.

Participants provided a photograph of themselves and approximately one minute of recorded speech. The system used these materials to generate personalized videos featuring a person resembling the participant and a cloned version of the participant’s voice. The program proceeded through an introduction to deepfakes, real-world fraud cases, personalized simulations, response strategies, and review quizzes.

The personalized scenarios included a situation in which the participant’s own voice recommended an investment, another in which someone impersonated a family member injured in a traffic accident and demanded an urgent money transfer, and fake news portraying the participant as either a lottery winner or a suspect in a fraud case.

Importantly, this was not an overseas study conducted with non-Korean participants. Although the paper was written in English and presented at the international CHI conference, the study was conducted in South Korea with 21 older adults whose native language was Korean. The participants’ average age was 65.8, with ages ranging from 61 to 80. All used smartphones and lived in urban areas of South Korea.

What Changed After a Single Session?

This was an exploratory pre–post study without a control group. Each participant took part in a session lasting approximately 90 minutes and used DeepAware independently for about 30 minutes. Before and after the intervention, the researchers measured five factors based on Protection Motivation Theory. They also analyzed participants’ think-aloud comments during use and conducted interviews after the session.

Source: Reconstructed by the author based on the pre- and post-test results reported in Oh et al. (2026).

The numbers indicate clear changes. Participants showed gains not only in knowledge and risk perception, but also in their confidence that they could respond and their belief that the recommended response strategies would work. The mean system usability score was 76. Satisfaction averaged 4.55 out of 5, usefulness 4.48, and immersion 4.26. By contrast, the perceived realism of the videos remained moderate at 3.21.

This suggests that the learning effect may have come not from technically flawless videos, but from the personal relevance of seeing one’s own face and voice placed in a criminal scenario. Only five of the 21 participants had previously heard the term “deepfake.” For many, this was not an opportunity to review familiar information, but their first structured encounter with a new type of risk.

Did Detection Ability Improve, or Did Coping Confidence Increase?

One distinction is essential when interpreting the findings. The study identified immediate psychological changes in knowledge, risk perception, and self-efficacy after the educational session. It did not test whether participants became more accurate at detecting deepfake videos or whether they would actually stop a money transfer and verify the situation when confronted with a real scam.

It would therefore go beyond the evidence to claim that DeepAware has been proven to prevent deepfake crime. A more accurate conclusion is that it demonstrated the potential to strengthen the awareness and coping confidence needed for fraud prevention.

DeepAware was also more than a tool that simply provided personalized videos. It combined explanations of key concepts, real-world examples, simulations, response strategies, and quizzes. The study cannot separate whether the increase in post-intervention scores resulted from the shock of seeing one’s own face, the repeated explanations and quizzes, or the combination of these components. Because there was no control group, it is also impossible to determine whether personalized education was more effective than a conventional educational video of the same duration.

Shock Can Educate, but It Can Also Create Fear

Seeing one’s own face and voice manipulated can transform an abstract risk into a personally relevant reality. This is the study’s most compelling strength. Yet the same strength may also become a weakness.

If the shock leads learners to think, “I should verify the situation next time,” it can serve an educational purpose. If it instead leads them to conclude, “I cannot trust anything anymore,” or “It would be safer not to use a smartphone,” the intervention may have the opposite effect. Increasing fear among older learners is not the same as increasing their ability to respond.

A threatening scenario must therefore be followed by repeated practice of short, actionable steps. One example is a three-step routine: stop, call back, and verify. When someone urgently requests money, the learner should end the call, contact the family member again using a previously saved number, and verify the situation through a prearranged family password or another family member. These procedures are likely to remain useful longer than visual techniques for spotting unnatural lip movements, even as deepfake technology becomes more sophisticated.

What Matters More in AI Classes for Older Adults

In my own work teaching AI and video production to older adults, I have observed wide differences in how quickly learners understand the same explanation and how much repetition they require. Even when learners follow the steps successfully during class, they may forget the sequence when they try again several days later. After a long break, such as a vacation period, it may be necessary to review previous material from the beginning.

When an unfamiliar tool and an unexpected change on the screen appear at the same time, anxiety about “having pressed the wrong button” can become more immediate than the lesson itself. Fraud-prevention education should therefore rely not on a single intense experience, but on a structure that repeatedly combines short explanations with hands-on behavioral practice.

Educators should also avoid placing too much emphasis on asking learners to identify a fake by sight or sound. Indicators such as subtle tremors in a voice, unnatural lip movements, or errors in the background may become increasingly unreliable as the technology advances. Instead, education should focus on recurring patterns in criminal situations, such as urgent money transfers, investment solicitations, and requests for verification codes.

Learners should practice ending a call, calling a family member back, and locating the appropriate reporting number. The central goal of the educational design should not be to distinguish real content from fake content with perfect accuracy. It should be to stop the transfer of money or personal information and carry out a verification procedure when a situation appears suspicious.

The Paradox of Safety Education That Requires Personal Data

DeepAware uses the learner’s own facial photograph and voice recording. The researchers obtained consent before participation, allowed participants to preview the generated results, provided a function for stopping the experience, and conducted a debriefing afterward. They also reported that the generated materials were deleted after the educational session.

Nevertheless, the intervention requires sensitive biometric information, including facial and voice data, to be transmitted to external generation services and storage systems. This creates a paradox: an educational program designed to teach the risks of deepfakes may introduce another form of personal-data risk.

If this approach is adopted by educational institutions, data minimization must take priority over entertaining or shocking personalization. Learners should receive plain-language explanations of how long the original data and generated content will be stored, which service providers will process them, how they will be deleted, and whether reuse is prohibited. Alternatives should also be considered, such as using only partial facial features, processing the data entirely on the learner’s device, or providing an equally complete educational experience for those who choose not to use personalization.

A Small but Important Starting Point, with Questions Still Unanswered

The sample consisted of only 21 participants, all of whom lived in urban areas and actively used smartphones. It remains unclear whether the same results would apply to older adults with limited digital access, those with cognitive or sensory difficulties, or those living in rural and fishing communities.

Because the outcomes were measured immediately after a single 90-minute session, the study also does not show whether the effects would remain several weeks or months later. Participants knew the purpose of the research and had agreed to provide sensitive personal data, which may also have encouraged responses consistent with the researchers’ expectations.

Future research should randomly compare conventional example-based instruction with personalized education and examine whether participants perform verification behaviors during realistic scam-message or phone-call simulations. Follow-up assessments should be conducted after a longer period, and participants with different ages and levels of digital proficiency should be included. Most importantly, researchers need to determine whether increasing the intensity of personalization strengthens learning or merely increases anxiety and distrust.

After “I Could Be Fooled Too,” What Should Learners Be Taught to Do?

DeepAware is significant because it shifts deepfake fraud education from explanation to experience. The finding that older Korean adults perceived the threat as more serious and felt more confident in their ability to respond after encountering criminal scenarios created with their own faces and voices is also noteworthy. The design is particularly valuable because it treats older adults not as passive recipients of protection, but as learners capable of making decisions and taking action.

However, the study demonstrates a starting point for fraud prevention, not its completion. Personalized shock can make a threat feel like “something that could happen to me,” but that alone cannot prevent an actual scam. The success of the education depends on what learners are asked to practice after the shock.

Experiential learning becomes genuinely protective when it does not teach older adults to distrust every video, but instead helps them build the habit of pausing when faced with an urgent demand for money and verifying the situation through a familiar and trusted channel.