There's a full draft copy of the disputed paper on Reddit: https://old.reddit.com/r/MachineLearning/comments/k77sxz/d_timnit_gebru_and_google_megathread/gepcliq. For the file itself:...
I found it quite funny that the "redaction" of authors names etc. is just pink boxes placed over the text. I opened up the PDF in LibreOffice Draw and deleted the pink boxes - the original text is plain to see. You can achieve something similar by just selecting the text and copying it.
I speculate that other teams doing related work saw a draft and didn't think it was accurate, and instead of figuring out how to use their data to make it into a better paper, the conflict somehow...
I speculate that other teams doing related work saw a draft and didn't think it was accurate, and instead of figuring out how to use their data to make it into a better paper, the conflict somehow escalated. For example, the data center folks at Google probably have better data than anyone about their CO2 emissions.
You might compare with another paper that was published recently. There are about 40 authors, which makes sense because it was using data from lots of different teams. That's a huge collaborative effort on a paper that makes AI look pretty bad. Someone had to be good at interacting with other teams to make that happen.
My guess has been that something odd, and probably drama-causing in itself, happened with the approval from the internal review process (I think some Google researchers are referring to it as...
the conflict somehow escalated.
My guess has been that something odd, and probably drama-causing in itself, happened with the approval from the internal review process (I think some Google researchers are referring to it as PubApprove?). If you notice, in Jeff Dean's email, his standard writing is almost exclusively active voice, with agentive subjects. Then, suddenly, for one sentence, he switches to the passive voice, and has no agents (emphasis mine):
Unfortunately, this particular paper was only shared with a day’s notice before its deadline — we require two weeks for this sort of review — and then instead of awaiting reviewer feedback, it was approved for submission and submitted.
In other words, in line with the Walkout letter, he seems to be saying that the review process approved the paper for publication, and that the person in charge of approvals (presumably not the authors) approved it without waiting for responses from some or all reviewers. As Gebru has pointed out that the two weeks requirement doesn't show up anywhere in the policy for the process, I suspect he is referring to what the reviewers needed, not what the policy required. He seems to be trying to pass the blame for all this to the authors by covering up the actors in this part, and not even disclosing their roles.
My best guess is this:
The authors submitted the paper to the internal review process, probably the day before the conference submission deadline, or at least shortly before it (Gebru is trying to argue that the process didn't strongly require a set time, which suggests this is true). Whether this is common at Google seems under debate.
Regardless of the policy, the person in charge of the review process sent it to reviewers with one day's notice, and then, hearing no objection or ignoring them, approved it for submission, and told the authors. Whether or not this low level of scrutiny was standard, and what the usual purpose of the internal review process is (internal peer review, or cursory IP/legal/etc check?), seems under debate. It doesn't seem unlikely to me that the latter type of review would typically be of a "Hey, do you need time to look over this closely, or should I just approve it?" nature, and maybe some reviewers missed the email, or didn't make clear that they needed time, or simply didn't think it merited forwarding to those who ended up having the problems.
The authors, confident that the paper was approved by the internal review, submitted it to the conference.
Some reviewers, however, disagreed with the paper, and also disagreed with it having been approved. This presumably led to an argument with the people who approved it about whether they should have done so, and what to do now.
Presumably after that battle was fought (five weeks later, according to the Walkout letter), higher-ups at Google tell Gebru that while the paper was approved for submission, after further consideration, Google was retracting the approval, and demanding that the paper be retracted or Google be disconnected from it.
Gebru, understandably, is incensed—at least for peer review, a demand for retraction after approval is the sort of action reserved for extreme cases like fundamental flaws or research dishonesty—and demands to know who called for this retraction, and why.
Since the problem here stemmed from the reviewers not having their reviews in the initial review process, Google then tells her what sound like rather standard peer review comments, which probably would have been fine if they had come in the course of a review, but coming at this point, infuriate her even more.
There's a full draft copy of the disputed paper on Reddit: https://old.reddit.com/r/MachineLearning/comments/k77sxz/d_timnit_gebru_and_google_megathread/gepcliq.
For the file itself: https://gofile.io/d/WfcxoF.
I found it quite funny that the "redaction" of authors names etc. is just pink boxes placed over the text. I opened up the PDF in LibreOffice Draw and deleted the pink boxes - the original text is plain to see. You can achieve something similar by just selecting the text and copying it.
I speculate that other teams doing related work saw a draft and didn't think it was accurate, and instead of figuring out how to use their data to make it into a better paper, the conflict somehow escalated. For example, the data center folks at Google probably have better data than anyone about their CO2 emissions.
You might compare with another paper that was published recently. There are about 40 authors, which makes sense because it was using data from lots of different teams. That's a huge collaborative effort on a paper that makes AI look pretty bad. Someone had to be good at interacting with other teams to make that happen.
My guess has been that something odd, and probably drama-causing in itself, happened with the approval from the internal review process (I think some Google researchers are referring to it as PubApprove?). If you notice, in Jeff Dean's email, his standard writing is almost exclusively active voice, with agentive subjects. Then, suddenly, for one sentence, he switches to the passive voice, and has no agents (emphasis mine):
In other words, in line with the Walkout letter, he seems to be saying that the review process approved the paper for publication, and that the person in charge of approvals (presumably not the authors) approved it without waiting for responses from some or all reviewers. As Gebru has pointed out that the two weeks requirement doesn't show up anywhere in the policy for the process, I suspect he is referring to what the reviewers needed, not what the policy required. He seems to be trying to pass the blame for all this to the authors by covering up the actors in this part, and not even disclosing their roles.
My best guess is this: