Microsoft Tweaking AI Depictions Of People With Autism, Other Conditions

One of the nation’s biggest technology companies is taking steps to improve the accuracy of AI-generated images portraying people with disabilities.

Microsoft says it is introducing enhancements to depictions of dwarfism, blindness, low vision and limb difference in images generated using its Bing Image Creator and M365 Copilot. Similar changes affecting representations of autism and Down syndrome rolled out in May and the company said feedback has been “overwhelmingly” positive.

“AI systems hold immense potential for disabled people,” said Jenny Lay-Flurrie, vice president and chief accessibility officer at Microsoft in a posting about the updates. “However, AI can also get it wrong.”

AI models are often trained using publicly available content, meaning that they rely on more limited and sometimes inaccurate representations of disabilities, Lay-Flurrie said. As a result, she said that earlier versions of Microsoft’s image generation came up with images of people with dwarfism who looked like fantasy characters and blindness depicted as an individual wearing a blindfold. The system was not able to generate an image of limb difference.

The company worked with individuals with disabilities and other experts to train AI models on more authentic representation.

“When disability representation is misaligned or missing, it can inadvertently reinforce stigma, escalating harms by erasing lived experiences or disrespecting disabled identities,” Lay-Flurrie said. “Working to create accurate, respectful, and positive representations in image generation models will help to challenge stereotypes and enable people to see themselves reflected on the screen.”

Microsoft’s efforts come as concerns are being raised about AI-generated content depicting people with disabilities. Reports earlier this year indicated that AI-generated accounts impersonating people with Down syndrome were proliferating on social media.