Project
AI and Accessibility
AI still faces hurdles in making the world more accessible to the disabled.
When AI transcription tools fail those who need them most
Next time you go to the doctor, your conversation with the physician might be automatically transcribed by an AI-powered tool, saving the practice the time spent typing and record-keeping.
Transcription tools are much more widely used in medical settings than you might expect. But as Pulitzer Center AI Accountability Network Fellow Hilke Schellmann reported, in partnership with AP investigative journalist Garance Burke, the tools are not ready to accurately interpret sensitive conversations related to your medical needs.
In fact, Schellmann and Burke found that popular AI tools like Open AI’s Whisper are prone to inventing and inserting chunks of text in the transcripts they produce, including “racial commentary, violent rhetoric, and even imagined medical treatments.”
Several independent researchers are sounding the alarm. Open AI itself cautions clients about using Whisper in high-risk settings, but that hasn’t stopped medical facilities across the U.S. and the world from rushing to adopt the tool. Some hospitals use a Whisper-based tool built by a company called Nabla that deletes the original recording, making it impossible to go back and check for accuracy.
While some of these applications are fairly recent, fueled by the expansion of generative AI technologies, transcription tools have been used in the workplace since the 1990s, in part to address the needs of people with disabilities. In her article for the Financial Times, AI Accountability Fellow Joanna S. Kao explains that while transcription tools have helped deaf employees navigate work when human interpreters are not present, flaws in the technology design often make their lives harder.
One of the reasons why transcription tools fail to recognize accented or irregular speech, Kao writes, is because AI companies historically have failed to seek input from disabled individuals—the very communities that could benefit most from the technology.
To help readers understand what disabled people face when transcription tools fail to accurately capture their speech, Kao embedded in her article audio snippets from deaf individuals along with the transcripts generated by providers such as Google, Amazon, and Speechmatics. Don’t forget to check those out.
At the Pulitzer Center, we will continue to confront AI hype with in-depth, nuanced reporting that informs and helps audiences become critical users of AI technologies, while holding powerful AI interests to account.
Best,
Impact
Grantee Alec Kuhn’s Pulitzer Center-supported project Why Are Alaska’s Rivers Rusting? has won the prestigious Kavli Science Journalism Award from the American Association for the Advancement of Science (AAAS). The project, which investigates the effects of warming temperatures on nature and infrastructure in the state, is part of the Pulitzer Center’s Connected Coastlines initiative, focusing on the impacts of climate change on U.S. coastal communities. The Kavli Award is the oldest science journalism award.
Photo of the Week
This message first appeared in the November 22, 2024, edition of the Pulitzer Center's weekly newsletter. Subscribe today.
Click here to read the full newsletter.
If the speech-to-text transcription process produces a significantly higher “Word Error Rate” (WER)...
