One afternoon in late November of last year, Timnit Gebru was sitting on the couch in her San Francisco Bay Area home, crying.
Gebru, a researcher at Google, had just clicked out of a last-minute video meeting with an executive named Megan Kacholia, who had issued a jarring command. Gebru was the coleader of a group at the company that studies the social and ethical ramifications of artificial intelligence, and Kacholia had ordered Gebru to retract her latest research paper—or else remove her name from its list of authors, along with those of several other members of her team.
The paper in question was, in Gebru’s mind, pretty unobjectionable. It surveyed the known pitfalls of so-called large language models, a type of AI software—most famously exemplified by a system called GPT-3—that was stoking excitement in the tech industry. Google’s own version of the technology was now helping to power the company’s search engine. Jeff Dean, Google’s revered head of research, had encouraged Gebru to think about the approach’s possible downsides. The paper had sailed through the company’s internal review process and had been submitted to a prominent conference. But Kacholia now said that a group of product leaders and others inside the company had deemed the work unacceptable, Gebru recalls. Kacholia was vague about their objections but gave Gebru a week to act. Her firm deadline was the day after Thanksgiving.
Gebru’s distress turned to anger as that date drew closer and the situation turned weirder. Kacholia gave Gebru’s manager, Samy Bengio, a document listing the paper’s supposed flaws, but told him not to send it to Gebru, only to read it to her. On Thanksgiving Day, Gebru skipped some festivities with her family to hear Bengio’s recital. According to Gebru’s recollection and contemporaneous notes, the document didn’t offer specific edits but complained that the paper handled topics “casually” and painted too bleak a picture of the new technology. It also claimed that all of Google’s uses of large language models were “engineered to avoid” the pitfalls that the paper described.