Apple’s Image Playground Has Racial Bias Issues

Apple’s new image generation app, Image Playground, has been found to have racial biases. Research by Jochem Gietema, a machine learning scientist, shows that the app struggles to accurately represent skin tones and hair textures, especially for people with darker skin. These issues exist even though the app was designed to avoid problems like deepfakes using illustrated face styles.

The biases go beyond just representation. The app changes facial features and skin tones based on the words used in descriptions. For example, terms related to wealth—like “rich” or “investment banker”—result in lighter skin tones in the generated images.

Conversely, words like “poor” or “destitute” lead to darker skin tones. Similar patterns were noted with different themes, where descriptions linked to basketball or rap music often resulted in darker skin tones, while those connected to classical music or ballet produced lighter skin tones. This reinforces negative racial stereotypes.

These problems are not just limited to Apple’s app. Another research paper found that biases in AI-generated images are widespread across the industry. Apple has promoted itself as a leader in responsible AI development and has put in place various safety measures in Image Playground, such as banning the use of celebrity names and negative keywords. Despite these measures, the app still shows these harmful biases.

Related: Mexico To Sue Google Over Gulf of America Renaming

These findings are particularly troubling for Apple, given its public image and commitment to ethical AI. The company has shown intent to tackle bias issues, as evidenced by recent research on gender bias in language models. However, this situation highlights the difficulties tech companies face in reducing bias in AI, even with careful design and precautions.

This incident adds to the growing concern about bias in AI, following similar issues with other major tech firms. While correcting algorithmic bias is easier than changing human prejudice, it requires continued effort and investment from developers.

Source: RPP