Identifying bias in AI generated images

Sammanfattning

Introduction: The goal of this study is to investigate and identify potential biases of generative AI models when generating images. The increased availability of generative AI models capable of generating images poses a risk of images being posted, used and shared without a proper analysis in bias. When consuming media one can, to some degree, identify the bias of the media source and effort is often put on being aware of which biases are at play. However, the rapid spread of generative AI model lead to a potential risk of missing or foregoing bias awareness when consuming media generated said models. Research Question: The primary research question is: “What are the potential biases identified when analyzing images of people generated using generative AI models?” Method: The study used a sample of 440 images generated by 3 different popular generative AI models which was analyzed to identify possible bias when using different prompts as well as identifying any potential biases between the different models themselves. Results: Multiple biases affecting all generative AI models were identified, such as a bias against people having a heavyset body type and bias against disabilities. Moreover multiple biases affecting each model individually were identified, such as Grok generating less West Asians when generating successful people, Gemini favoring women when generating successful people and ChatGPT favoring generating white people. Discussion: The biases identified range from harmless biases like a bias against generating images with accessories, to harmful such as the sexualization of women generated by Grok. Since AI generated images are not fully regulated, the identified biases suggest that users need to be aware of potentially harmful biases included in the images they generate. Knowing which biases are shared across all investigated models as well as the model specific biases might empower users to add specific parameters or to tailor their prompts based on model to mitigate specific biases directly.