AI Hyperrealism: Do AI Faces Look More Real Than Human Faces?

1. To do this project, you will need to find or make or find your own sets of both AI-generated and real pictures of human faces. Start with at least 10 of each type, but you can use more than that. Here are some things to consider:
   1. If you want to find images online, make sure you find them from a reputable source (such as a public image database from an academic institution) that clearly labels whether the images are real or AI-generated. General Google image search, social media, and public image hosting websites are not reliable places to find accurately labeled real and AI-generated images.
   2. The best way to guarantee that you know which images are real and which are AI-generated is to take or make them yourself.
      1. You can take pictures of people yourself using a smartphone or camera.
      2. Many AI image-generation sites and programs are available online. Some offer free versions and some may require subscriptions. Available tools are quickly changing and being updated all the time, so you will need to search to see what is currently available when you do this project. 
   3. Regardless of how you get or make your photos, make sure that there are no major systematic differences between your real and AI-generated photos. For example, if your AI-generated photos all have blank backgrounds (like Figures 1 and 2), then you should also use real photos with blank backgrounds. Conversely, if you take all your real pictures of people outside with trees or buildings visible in the background, you should prompt the AI to make the background of the image look like the person is outside. If your AI-generated photos are all of adults, then you should not take real pictures only of kids your own age, and so on. 
   4. Make sure that the real people used in your photos will not be easily recognizable by the volunteers in your experiment. For example, if you find pictures of popular (real) celebrities online, or take real pictures of your friends then also use your friends as volunteers in your experiment, they will be more likely to recognize the pictures of real people. The pictures should be of "random" or "anonymous" people that your volunteers will not recognize. 
2. Store your real and AI-generated photos on your computer, organized in folders or with clear filenames so you know which ones are real and which are AI-generated. 
3. Prepare your photos for your volunteers to view. 
   1. You should put them in a random order without an obvious pattern (for example, do not just alternate between real and AI-generated). You can print the photos, put them in an electronic document like Microsoft Word or Google Docs, or create an online survey for them using Google Forms, Survey Monkey, Qualtrics, etc.
   2. For each volunteer, you will need a way for them to mark whether they think the photo is real or AI-generated, and why they think so. How you do this will depend on how you present the photos (for example, writing their answers down on paper surveys vs. submitting an electronic form). 
4. Prepare three data tables like Tables 1, 2, and 3. The first table will help you record all the individual responses (add columns for more volunteers as needed). Use the second table to add up the total responses. Use the third table to categorize and record reasons people mention for identifying real or AI images.

Table 1. Data table for recording individual responses. 

Table 2. Data table for tallying total responses. 

Table 3. Table for tallying how many volunteers mention different features when determining if an image is real or AI-generated. Your entries for the rows may vary depending on your volunteer responses.

5. Show the images to each of your volunteers and record their responses. 
6. Analyze your data.
   1. Make a bar graph of your data from Table 2. Overall, how good were your volunteers at correctly identifying real and AI-generated faces?
   2. Did you observe "AI hyperrealism," where volunteers were more likely to identify AI-generated faces as real than real faces?
   3. From your dataset, identify the following:
      1. The real face that was most frequently correctly identified as real
      2. The real face that was most frequently incorrectly identified as AI
      3. The AI-generated face that was most frequently correctly identified as AI
      4. The AI-generated face that was most frequently incorrectly identified as real
   4. Compare your results from step c to your notes in Table 3. Do any of the features people listed help them correctly classify images? Do any tend to mislead or cause them to classify incorrectly, such as labeling an image as "too perfect" when it is actually of a real person?
   5. Based on your experiment, do you have any tips or recommendations for how people can identify AI-generated pictures?