Introduction
The use of AI has risen across the world, and this makes it harder to detect what is real or
human created compared to something fake that AI created. Deepfakes are videos and images
that are made to look like real people, and they are made by using artificial intelligence. The study shows people using an AI tool called StyleGAN2 that can detect deep-fake images. The study also shows that even when given help, people are awful at being able to spot a deepfake. This relates heavily back to the study of social sciences because it shows behavior with technology and decision-making and human thought processes.
Relation to Social Science
This study relates to social science because it shows how individuals decide on something and how they use their knowledge to judge what is real and what is fake. The study of social science typically focuses on how people process information that they take in and how they interact with others. Throughout this research study, people tend to be very confident in something even when they are completely wrong; this shows that they are overconfident and biased. There are different methods of intervention that can help make deepfakes easier to detect, including practice images that show differences between a real image and a deepfake and giving advice on how to detect deepfakes. This can hopefully educate people and make it where they are less susceptible to falling for deepfakes. The study used controlled experiments, which relates back to research methods discussed in our class and helps researchers to observe actions and understand more about how people make their decisions. In Module 7 we read an article about cybersecurity and AI. The article explained how people interact with different forms of technology as well as AI, and it showed how misinformation and AI can affect individuals as well as society as a whole. Understanding or lack thereof can have real consequences, and in the deepfake research study, we learn that people tend to not be good judges of digital information in front of them. A person could think that they are right and that the image in front of them isn’t a deepfake, and if they are wrong, there can be
severe consequences, including information being stolen.
Hypotheses, Variables and Research Questions
Three main research questions were asked within this study. “Does advice or practice
improve detection? Is confidence linked to accuracy? Can people reliably detect deepfake images?” Researchers were able to predict that humans have a better chance at detecting deepfakes compared to random guessing, but even humans make mistakes and can be wrong about detecting them. If there were interventions like explaining how to detect a deepfake or showing real images versus deepfakes, detection could be easier. The independent variables in this experiment were the types of interventions that happened, including practice images, no advice (control group), reminders, and practice images. Dependent variables were “tell-tale.” features, confidence rating, and the accurate ability to detect deepfakes.
Research Method
In order to conduct an experiment, the researchers chose to do a controlled experiment
where they had a group of participants randomly placed into four different groups. Each of the four groups was given twenty images out of one hundred; half of these images were real, and the other half were deepfakes. Some of the groups received advice, some received no help, and the rest received practice images. This proved that if someone had no idea how to properly detect deep fakes, they would likely not be able to detect them. If someone was given advice they could
have a better chance of correctly identifying what images were fake. This method is used often in
social science so that way researchers can see how people act in a controlled environment.
Analysis and Data
Confidence ratings were collected on the ability to detect deepfakes, the accuracy of each person’s answers, and their explanations on why they chose deepfake or real images. They mentioned the tell-tale signs. The methods that researchers used were signal detection theory, correlation, analysis of variance and per image analysis in order to analyze the data from the experiment.
Relation to lessons
This research study related to multiple concepts that were covered in our lessons. Experimental designs were very common in social science, as they contain both dependent and independent variables and use random assignment. Module 7 relates back to this because not only can disruptions happen because of AI, but human errors and misinformation can also cause major disruptions.
Relation to Marginalized Groups
The participants in this study were almost all younger people who spoke English fluently.
However the elderly population and people with lack of exposure to technology would struggle
to identify a deepfake even more than a person who used technology daily. The elderly are a lot
more susceptible to fall for scams and would have no clue that an image or video could be fake.
Contributions made to Society
People tend to assume that they can identify a deepfake easily when in reality they are usually wrong; this could lead to them likely falling for and believing misinformation. When someone is given advice on detecting a deepfake, they still tend to get them wrong, but the likelihood of them being right tends to be slightly better. There needs to be programs and tools to educate people because technology changes each and every day and without proper education, people will keep falling for deepfakes every day.
Conclusion
This study showed that even with practice and advice, people still are not very good at
detecting a deepfake. Participants within this study were very confident in their decisions of
what was real and what was a deepfake, even when they got it wrong. This related heavily back
to our lesson in Module 7, where we learned that cybersecurity threats are so broad that humans
alone cannot protect themselves from the risks of AI. Social science research is able to help us
guide others and improve so that we are able to help protect those most vulnerable to attacks.
Safeguard, lessons and education are needed now more than ever to teach the population on how
to use AI responsibly.
Works Cited
Bray, S. D., Johnson, S. D., & Kleinberg, B. (2023). Testing human ability to detect ‘deepfake’
images of Human Faces | Journal of Cybersecurity | Oxford academic. Retrieved from
https://academic.oup.com/cybersecurity/article/9/1/tyad011/7205694
Wilner, A. S. (2018). Cybersecurity and its discontents: Artificial intelligence, the Internet of
Things, and digital misinformation. International Journal: Canada’s Journal of Global Policy
Analysis, 73(2), 308-316. Retrieved from
https://journals.sagepub.com/doi/abs/10.1177/0020702018782496