Jess Smith, a former Australian Paralympic swimmer, discovered that AI image generators like ChatGPT initially failed to accurately create images of her with one arm, despite her specific prompts.
The AI either depicted her with two arms or replaced her missing limb with a prosthetic device.
When she inquired, the AI explained it lacked sufficient data to generate such images.
This experience highlighted for Smith how AI reflects existing societal inequalities and discrimination.
Recently, however, ChatGPT’s image generation capabilities have improved, allowing it to produce accurate images of people with disabilities like Smith, marking a significant step forward in representation.
Smith emphasizes that this progress is more than just a technological advancement; it represents a broader movement toward inclusion and humanity in AI development.
OpenAI, the company behind ChatGPT, acknowledged making meaningful improvements to its image generation model and is actively working to reduce bias by refining training methods and incorporating more diverse examples.
Despite these advances, other users with disabilities, such as Naomi Bowman who has sight in only one eye, continue to face challenges.
Bowman reported that AI altered her facial features when asked to blur the background of a photo, demonstrating ongoing issues with bias and representation.
Critics also raise concerns about the environmental impact of AI technologies like ChatGPT, noting the high energy consumption of data centers powering these models.
Experts agree that AI bias often mirrors societal blind spots, extending beyond disabilities to other underrepresented groups.
Abran Maldonado, CEO of Create Labs, stresses the importance of cultural representation during the data training and labeling process to ensure AI systems are inclusive.
He points out that without consulting people with lived experiences, AI will continue to miss or misrepresent diverse populations.
Historical examples, such as a 2019 US government study revealing facial recognition algorithms’ lower accuracy for African-American and Asian faces, illustrate these challenges.
Smith does not identify as disabled, viewing the barriers she encounters as societal rather than personal limitations.
She highlights everyday design oversights, like public taps that require holding down, which hinder accessibility.
Smith warns that similar oversights in AI development risk excluding people with disabilities from the digital world.
When she shared her experience on LinkedIn, a developer claimed his AI app could generate images of a woman with one arm, but Smith found it failed similarly.
She notes that conversations about disability often become uncomfortable, causing people to avoid engagement, which further perpetuates exclusion.
Overall, the article underscores the evolving nature of AI and the critical need for inclusive design and diverse data to ensure fair representation.
It calls attention to the societal implications of AI bias and the importance of involving people with disabilities and other marginalized groups in the creation process.
The recent improvements in AI image generation represent progress, but ongoing efforts are necessary to address persistent challenges and build technology that truly reflects the diversity of human experiences.