As development on machine learning reaches a critical stage where decisions made on current research may have vast, unchecked effects for lifetimes, Google finds itself under ethical fire after having dismissed a researcher from its Brain unit who wanted to challenge the findings of a paper that other Googlers authored on the benefits of A.I.-assisted chip design.
The background to this story, as The New York Times reports from its sources, involves an earlier paper on the same technology published in 2020. Google asked Dr. Satrajit Chatterjee to work on bringing the technology to chipmakers for sale or licensing. Chatterjee, however, expressed concerns about some of the claims made and whether they had been tested thoroughly. As the complaints were being discussed, the names of two of the paper’s co-authors who had similar concerns were removed in a later revision that was made as part of Google’s pitch to Nature for what would be the 2021 paper in question. Some observers said that the revision was not properly vetted through an approval process the company created after it fired Dr. Timnit Gebru (via Wired) and Margaret Mitchell (via The New York Times), leads at Google’s A.I. ethics branch who also had misgivings about the company’s direction on research.
After the Nature paper (PDF as obtained by The Verge) was published, Google allowed Chatterjee and other opposing researchers the opportunity to draft a rebuttal paper. The rebuttal was submitted to an internal arbitration committee and was rejected months later. The opposing researchers asked to escalate the issue to Alphabet’s board of directors, saying that the decision not to publish the rebuttal violated the company’s own principles. It was shortly afterwards that Chatterjee was fired.
Google confirmed to the Times it terminated Chatterjee with cause and supports the findings of its papers. Google Research vice president Zoubin Ghahramani also mentioned that the company investigated claims brought up in a “subsequent submission” that ultimately “did not meet our standards for publication.”
Anna Goldie, one of the two lead authors of the Nature paper, said revisions to the 2020 paper did not need to be vetted under the full approval process. She also claims that Chatterjee approached her and the co-lead author about managing the project in 2019 and that they had refused and that she addressed his concerns about her research with evidence but was ignored. Goldie accused Chatterjee of a “campaign of misinformation.”
Chatterjee’s lawyer called out Jeff Dean, a senior vice president at Google Research, for his “actions to repress the release of all relevant experimental data, not just data that supports his favored hypothesis […]”
The methods described in the Nature paper are touted to generate chip floorplans in under 6 hours as opposed to the months it takes for humans to do the same. But while the results have yet to speak for themselves in wider practice, it’s pretty clear that there’s plenty to fight about when it comes to how the chips — and any wares derived from artificial intelligence itself — are made.
In a statement to Android Police, Google said it had nothing more to share at the moment, but directed us to a tweet from one of the Times story’s co-authors, Daisuke Wakabayashi, summarizing the story.