Setting the Record Straight #ISupportTimnit #BelieveBlackWomen
We’re setting the record straight on Dr. Timnit Gebru’s firing.
Did Dr. Gebru resign?
Dr. Gebru did not resign, despite what Jeff Dean (Senior Vice President and head of Google Research), has publicly stated. Dr. Gebru has stated this plainly, and others have meticulously documented it. Dr.Gebru detailed conditions she hoped could be met. Those conditions were for 1) transparency around who was involved in calling for the retraction of the paper, 2) having a series of meetings with the Ethical AI team, and 3) understanding the parameters of what would be acceptable research at Google. She then requested a longer conversation regarding the details to occur post-vacation. In response, she was met with immediate dismissal, as she details in this tweet. Dr. Gebru’s dismissal has been framed as a resignation, but in Dr. Gebru’s own words, she did not resign. All reports under her management received a letter from Megan Kacholia (Vice President of Engineering for the Google Brain organization), stating that Megan had accepted Timnit’s resignation. Megan went around Dr. Gebru’s own manager, Samy Bengio (lead of Google Brain) in sending these emails, which he has stated publicly.
What’s this about a controversial paper?
Dr. Gebru and colleagues at Google and the University of Washington wrote a paper detailing the ethical considerations of large language models. Large models (like BERT and ELMo) are used in many different places within the tech industry and AI research, including in products like Google’s search engine. Details of the paper can be found in this article from MIT Technology Review.
So was the paper submitted to the conference without approval?
No. On October 8, shortly after midnight, Dr. Gebru et al.’s paper received approval for publication through the standard processes, which is called PubApprove internally. The deadline for the conference they submitted it to (FAccT) was on the same day.
But didn’t Jeff Dean say Google’s research approval process takes two weeks?
Jeff Dean’s message indicated that the approvers “require two weeks”. But there is no hard requirement for papers to actually go through this review with two weeks’ notice. Numerous papers are approved for publication submission without meeting this “requirement”: an internal analysis shows that just under half of the papers submitted to PubApprove are done so with a day or less notice to approvers. So it is clear that this is a standard which was applied unevenly and discriminatorily.
How rigorously was the paper reviewed?
The PubApprove process is not a replacement for peer review, but is usually a check that no sensitive material has been released. The paper received a “looks good to me” review from a subject matter expert, in addition to formal approval from Dr. Gebru’s manager, Samy Bengio. In addition to going through standard publication approval, the authors further opted to circulate the paper amongst 28 internal and external colleagues for feedback in the weeks preceding the submission to PubApprove. To have this many eyes on a paper during the process of peer review is unusually high for ML and AI research, and is a testament to the due diligence of the authors. They also took the additional step of looping in PR, giving them a heads up about the upcoming paper on September 18, and then adding them as PubApprove reviewers on October 7.
What happened after the paper was approved and submitted?
On November 18, Dr. Gebru and co-authors were invited, last minute, to a meeting with Megan Kacholia and Samy Bengio where they were given a verbal directive to either retract the paper or pull their names. No written feedback was provided from leadership, the authors were not given an opportunity to communicate about the verbalized concerns to anyone involved, and the authors were not provided with an opportunity to revise the paper in light of the feedback, despite the camera-ready deadline being over a month away. In a followup meeting on November 27, Dr. Gebru was provided with anonymous feedback, which was read to her by her manager from a Privileged and Confidential document. Her manager was prohibited from sharing it with her. The remaining Google co-authors were not included in this exchange. Moreover, this feedback was not provided as the first step of a process of revision.
And what about this email?
On December 1, Dr. Gebru sent an email internally to the Brain and Women Allies listserv. This list was created and is maintained as a space to foster an inclusive work environment for women in Google Brain. In this email, she shared her experiences of discrimination and her frustration regarding inaction around DEI efforts. She also shared her experience regarding the process of submitting, gaining approval, and then being forced to retract a research paper. This email was cited as cause “to move her resignation date forward,” effectively immediately.