Why AI-Generated Art Isn't Catering to the Needs of Individuals with Disabilities

Artificial intelligence (AI), especially within the field of instructional design, has transformed the way we obtain and produce educational materials. The emergence of image generators allows for the rapid generation of visuals, enhancing our educational content. However, similar to any nascent tool, AI-based image generation comes with its challenges, particularly when it comes to ensuring inclusivity.

The Challenge

In a community that is already grappling with the complexities of achieving adequate representation, it can be disheartening to observe that AI-generated images frequently encounter difficulties when attempting to faithfully depict individuals with disabilities. The fidelity of these generated images falls behind when compared to their able-bodied counterparts, possibly because of the scarcity of diverse training images of individuals with disabilities that AI can draw upon for learning.

For instance, when requesting images of deaf individuals, one may receive a wide range of inaccurate and potentially offensive representations. These can include depictions of individuals with exaggerated and perplexed facial expressions, appearing as if they are mimicking sign language. Likewise, when asking for images of people with hearing aids or cochlear implants, the results might include individuals with overly futuristic cybernetic attachments or, equally puzzling, individuals simply wearing regular headphones. Although cochlear implants can technically be categorized as cybernetic devices, the embellishment in AI-generated images is both incorrect and misleading.

Similarly, creating images of visually impaired individuals comes with its unique set of difficulties. Rather than producing precise portrayals, AI applications tend to excessively incorporate sunglasses, thus reinforcing the stereotype that all blind individuals wear them. Additionally, AI-generated visuals often feature unconventional "seeing-eye" canes that seem to be integrated into the person's body, underscoring the AI's lack of comprehension. The same issue arises in AI-generated depictions of people in wheelchairs. Frequently, these images depict a peculiar fusion of human and wheelchair elements, resulting in hybrids that appear half-human and half-wheelchair.

Even apparently simple requests, such as creating images of younger individuals using canes, yield inconsistent results. While older individuals with canes are portrayed with slightly better accuracy, it is evident that the AI's training data predominantly leans towards favoring older individuals over their younger counterparts.

The resolution

The key to solving this challenge lies in comprehending the mechanics behind the generation of images, where a pivotal role is played by what we call "tokens." These tokens serve as the governing units that direct the AI's image creation, and each request is associated with a specific token count. Excessive usage of keywords can lead to confusion within the AI, resulting in the creation of what is sometimes referred to as "mishmash monstrosities." This stands in sharp contrast to text-based AI models such as ChatGPT, which thrive when provided with detailed instructions. In the realm of AI-generated images, it is often the case that simplicity is more effective. Certain image generators, like MidJourney, even offer commands such as "/shorten" to enhance the efficiency of the input prompt.

In my personal journey, grasping the concept of tokens has proven to be of utmost importance. Through the generation of over 10,000 varied clipart images, I came to realize the significance of segregating character creation from the image backgrounds. It was around the 6,000th image mark that I transitioned to generating characters on a simple white backdrop. Devoting all available tokens exclusively to crafting the characters led to a marked improvement in the quality of images, especially those featuring individuals with disabilities. Subsequently, these images could be effortlessly merged with backgrounds generated independently.

While the current portrayal of individuals with disabilities in AI has significant room for improvement, gaining insight into its underlying mechanisms can lead us towards achieving greater accuracy and inclusivity. As the field continues to advance, it becomes crucial to champion the development of more comprehensive and diverse training datasets. In the meantime, refining our utilization of existing tools, as demonstrated through the token approach, can serve as a means to narrow this disparity.

For instructional designers and educators, ensuring the precision of content, whether in text or visuals, holds utmost importance. AI, despite its potential, should be employed with a thorough comprehension and a conscientious approach. As we progress, let us make certain that each person, regardless of their capabilities, is depicted with the dignity and consideration they are entitled to.