Microsoft purchased more than 20 million minutes of video from Be My Eyes, a nonprofit accessibility platform, to retrain its AI models after discovering they were producing offensive and inaccurate depictions of blind people.
Jenny Ray Fleury, the new head of Microsoft’s Trusted Technology Group, discovered the problem while overseeing accessibility efforts. The AI-generated images showed a stereotypical caricature of a visually impaired person who was completely blindfolded, drawn from biased training data that reflected societal biases rather than reality.
While Microsoft could have simply accepted this flaw, it decided to go a step further and collect videos of visually impaired people’s life experiences, such as walking around with a cane, working with a guide dog, and accessing objects in the home by removing visible faces, to train a model that realistically represents visual impairment.
Microsoft has chosen to centralize responsible technology into a top-down policy called the Trusted Technology Group, which was introduced in early 2025.
This model mirrors Bill Gates’ 2002 Trustworthy Computing memorandum, which philosophically prioritized reliability over speed, and is the complete antithesis to the Silicon Valley mantra of “move fast and break things.”
Nevertheless, Reliable CEO Annie Brown warns that data diversity alone is not enough to correct algorithmic bias. Metadata, the way data is labeled and categorized, presents another level of bias that may go unnoticed by a superficial audit.
While Microsoft’s investments in data sets are intended to improve data quality, there is no guarantee that the labeling process will be fixed for downstream users.
Early access to Copilot was provided to a group of employees with disabilities. Deaf employees received subtitles and sign language recognition. Nervous employees benefited from reduced cognitive load.
While Ray Fleury says AI is “leveling the playing field for previously marginalized workers,” disability advocacy groups emphasize including people with disabilities in decision-making roles, not just as users of solutions designed for them.
