AI ethics researchers Timnit Gebru and Abeba Birhane are mounting a systematic critique against Big Tech's foundation model approach, arguing it eliminates resource-efficient alternatives while generating fabricated outputs and environmental harm.
Gebru documented how Meta's No Language Left Behind model, announced as covering 200 languages including 55 African languages, directly killed funding for specialized startups. "Investors were like, 'Facebook has solved it, so your little puny startup is not going to be able to do anything,'" Gebru said in an AI Now Institute publication.
The pattern repeats across Big Tech releases. When OpenAI or Meta announces broad language coverage, investors pressure smaller organizations to shut down. OpenAI representatives have told language AI startups that OpenAI will make them obsolete and offered minimal payment for their data, according to Gebru's research.
"People came along and decided that they want to build a machine god," Gebru said. "They end up stealing data, killing the environment, exploiting labor in that process."
Birhane argues the 'AI for good' narrative serves as corporate deflection. "It allows companies to say 'Look, we're doing something good! Everything about AI is not bad. And you can't criticize us,'" she said, pointing to the framing as a response to grassroots resistance movements.
The critique arrives as mainstream AI development continues the resource-intensive trajectory. Recent developments include DeepSeek V4 and Nvidia's photonics investments for AI infrastructure, alongside enterprise efforts to engineer AI agent personalities. These advances follow the same paradigm Gebru and Birhane challenge: massive models requiring substantial computational resources rather than specialized, efficient alternatives.
The tension between ethical AI advocacy and industry direction centers on whether the benefits of general-purpose models justify their resource costs and market concentration effects. Gebru and Birhane argue that smaller, specialized models could serve specific communities and languages more effectively without the environmental and labor costs of training billion-parameter systems.
Their research suggests the dominant paradigm's market dynamics prevent alternatives from securing funding or partnerships, regardless of technical merit or efficiency advantages.