Google Paper Cites Research at Center of Its Staff Firestorm

(Bloomberg) -- Google in 2020 and 2021 dismissed the two co-heads of its artificial intelligence ethics team in a dispute over research that was critical of the company's work. Now, Google Research has published another paper on the same topic that cites the original work that led to the removal of the two leaders, Timnit Gebru and Margaret Mitchell.

The company dismissed Gebru—she says she was fired, while Google says she resigned—in late 2020 after she refused to either retract the paper or remove the names of the Google authors. Jeff Dean, the senior vice president of Google AI, later said in an email he shared publicly that the original paper “didn't meet our bar for publication” and “ignored too much relevant research.” The episode stirred up a sustained controversy at the company's research division, a key area of investment for the tech giant.

Mitchell remained at Google a few months longer until she was fired, and the paper was published with her name styled as “Shmargaret Shmitchell.” The citation in Google Research's recent paper, shared on Twitter Monday by Dean, retains the Shmitchell nom de plume.

The new Google paper cites the 2020 research as a source. It is 83 pages and involved several dozen authors, including Dean. Like the earlier paper, it deals with large language models, AI programs that generate and predict human language by analyzing massive volumes of text. Gebru, Mitchell and co-authors from the University of Washington including Emily M. Bender had raised ethical concerns about the field in their 2020 paper.

“This shows me how little accountability these leaders have,” Gebru said. “They don't have to think about the fact that they fired me claiming our paper was ‘sub par.' They don't have to think about the consequences—it's a distant memory for them.”

Mitchell had no immediate comment on the paper.

Google referred requests for comment to Dean's earlier email about the paper, which, while commenting on what he saw as flaws in the paper, also noted that it “surveyed valid concerns with large language models, and in fact many teams at Google are actively working on these issues.”