OpenAI's Whisper speech recognition model fabricates medical transcriptions, according to AI ethics researchers challenging the dominant scaling paradigm pursued by Big Tech companies.
Timnit Gebru and Abeba Birhane of the AI Now Institute argue the 'one giant model for everything' approach lacks empirical evidence of benefits while causing real-world harm. Gebru documented how Meta's 200-language No Language Left Behind model announcement prompted investors to pressure small African language NLP startups to close. "Facebook has solved it, so your little puny startup is not going to be able to do anything," investors told these organizations.
OpenAI representatives threatened similar startups, warning the company would make them obsolete and offering minimal payment for their data. "OpenAI is going to put you out of business soon because we're going to make our models better in your language," Gebru reported them saying.
The Whisper medical transcription errors represent a different failure mode. The model generates fabricated content when processing healthcare audio, creating safety risks in clinical settings where accuracy is critical.
Gebru characterizes the scaling paradigm as "stealing data, killing the environment, and exploiting labor" to "build a machine god." The criticism comes as Nvidia invests $4 billion in photonics partnerships and DeepSeek releases its V4 model, continuing the race toward larger systems.
Birhane identifies "AI for good" initiatives as deflection tactics. "It's a way to paint a positive image of AI technologies, especially in light of the backlash from the resist or refuse AI grassroots movement," she said. "AI for good allows companies to say 'Look, we're doing something good! Everything about AI is not bad. And you can't criticize us.'"
The researchers argue governments are making massive infrastructure investments based on promises rather than evidence. The dominant paradigm prioritizes scaling over specialized solutions that serve underrepresented communities.
Small language organizations face existential threats when Big Tech announces coverage of their languages, regardless of actual model quality. The competitive dynamic eliminates alternatives before users can evaluate performance differences.
The crisis highlights tensions between centralized scaling approaches and community-specific AI development. As major labs continue expanding model sizes, critics question whether the paradigm serves genuine needs or corporate growth imperatives.