AI Image Detection through Human-Centered Training Methods

As AI-generated images become increasingly prevalent in digital media, the ability to distinguish between real and manipulated content is essential for combating misinformation. Our study investigates whether targeted training can significantly improve individuals’ ability to detect AI-generated images. Somoray and Miller (2023) found that deepfake detection accuracy remained low, averaging around 55%, regardless of whether participants were given a list of detection strategies [1]. Our study builds on this by implementing a structured training program, which includes video demonstrations and interactive practice with feedback. We investigate whether detection accuracy improves after participants view videos explaining how to identify deepfakes for AI-image identification instead of just a list of strategies. The experiment began with a pre-test to assess participants’ baseline ability to distinguish between real and AI-generated images. Next, two-thirds of the participants received targeted training on identifying inconsistencies, while one-third served as a control group with no training. Finally, we administered a post-test to measure any improvements in their detection skills after training. Demographic and experiential factors such as age, sleep, AI experience, and screen time did not significantly impact detection accuracy. A paired t-test was performed to evaluate the impact of training on detection accuracy, and the results show a statistically significant improvement in detection accuracy post-training (p=0.009). A statistically significant positive correlation was found between the time spent analyzing images and detection accuracy (p < 0.0001), indicating that more thorough analysis improves performance.