GOOGLE — ETHICAL AI▊
The team hired to find the harms in Google's AI found some — in a paper about the technology Google was betting the company on. Both co-leads were gone within eleven weeks.
- OPERATION
- Ethical AI research team — hired to study the harms of the company's own technology, and publicized as proof it took them seriously FACT
- HABITAT
- Google Research (Brain), Mountain View
- EVENT
- CO-LEADS GONE IN 11 WEEKS — Timnit Gebru out Dec 2020, Margaret Mitchell fired Feb 2021 FACT
- DISPUTE
- Google: she resigned. Gebru: she was fired. Both versions on this page, at full strength ATTRIBUTED
- DISPOSITION
- No court, no regulator, no adjudication — a labor dispute litigated entirely in public FACT
Here is the conscience of the organism — a team hired, titled, and publicized specifically to say uncomfortable things about the host's product — observed at the precise moment it said one, about the technology the host had just bet the company on.
I was fired by @JeffDean for my email to Brain women and Allies. My corp account has been cutoff. So I've been immediately fired :-)
@timnitGebru — on X, Dec 3, 2020 FACT (that she said it) · x.com/timnitGebru/status/1334352694664957952
…multiple violations of our code of conduct, as well as of our security policies, which included the exfiltration of confidential business-sensitive documents.
Google, official statement on the Mitchell firing — Feb 2021, as carried by BBC News ATTRIBUTED (Google's claim) · bbc.com/news/technology-56135817
Google built an Ethical AI team and told the world about it. The team's whole function was friction: find the harms in the machine before the machine ships. In late 2020 the friction arrived on schedule — a paper about the risks of very large language models, the exact technology Google was racing to deploy — and within eleven weeks both co-leads were gone.
Our read, and it is ours, not a finding and not Google's stated intent: this is the first specimen in the drawer we label safety theater — a conscience created for credibility that gets cut the moment it frictions with revenue. Google's answer — that this was a publication-quality dispute and a resignation, not a silencing — is on this page at full strength. Read both. Keep the dates.
the drama timeline
ACT I — THE PARROT PAPER (NOV–DEC 2020)
A research paper moves through an internal review process. What comes out the other end is not edits.
-
NOV 2020
The paper hits internal review
"On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — co-authored by Gebru, Mitchell, and University of Washington linguists — surveys the risks of large language models: environmental cost, unfathomable training data, encoded bias. Google's internal review orders it retracted, or stripped of Google authors' names. Dean's stated reason, in his own later-published email: it "didn't meet our bar for publication" and "ignored too much relevant research" on mitigations.
Note what the paper was about: the technology the host organism was about to stake everything on. The conscience did its job. That is what makes the next node interesting.
-
DEC 2–3 2020
"We accept her resignation"
Gebru emails leadership with conditions — transparency about who reviewed the paper and how — and says that absent them she will work toward a departure date; she separately posts a frustrated note to the internal "Google Brain Women and Allies" list. Google cuts her corporate access and tells her team it has "accepted her resignation," effective immediately. Gebru, publicly and consistently since: she never resigned; she was fired, while on vacation.
-
DEC 2020
The apology that wasn't a reversal
Thousands of Googlers and outside academics sign a protest letter. Pichai sends a company memo apologizing that the episode led employees "to question their place at Google" and orders a review of the exit — an apology for the handling, not a reinstatement, and not an acceptance of Gebru's account.
ACT II — THE FOUNDER FOLLOWS (JAN–FEB 2021)
The team's founder starts looking through her own email for evidence. The host organism notices.
-
JAN 2021
Mitchell locked out
Google suspends Margaret Mitchell's corporate access. Per press reporting, she had been using scripts to search her own messages for material documenting the treatment of Gebru. The lockout lasts five weeks while Google investigates.
-
FEB 19 2021
Fired — this one not disputed as a firing
Google fires Mitchell, citing "multiple violations of our code of conduct, as well as of our security policies, which included the exfiltration of confidential business-sensitive documents." Both co-leads of the Ethical AI team are now gone, eleven weeks apart. The same month, Google reorganizes responsible-AI work under executive Marian Croak and announces changes to its research-publication and exit processes.
The team survives on the org chart — renamed, re-parented, and led by neither of the people who made it famous. The archive notes that the org box persisted and the friction did not.
both sides, on the record
The conscience was cut when it spoke. The team hired to find harms in Google's AI found some, in a peer-review-bound paper — and the company's response was retract-or-remove-names, followed by both co-leads' exits within eleven weeks [1] [3].
The process was the anomaly. Researchers told reporters the retraction demand was unprecedented and opaque — delivered late, without the reviewers' identities or an ordinary chance to revise [1] [3].
The company's own CEO conceded the handling failed. Pichai apologized and ordered a review; thousands of employees signed the protest letter [3].
Google's account: a quality dispute and a resignation. Dean's email states the paper "didn't meet our bar for publication," ignored research on mitigating the very harms it described, and was submitted a day before its deadline instead of the required two weeks. Gebru, in Google's telling, set conditions and offered to work toward departure if they weren't met — and Google accepted [2].
The Mitchell firing had a stated, specific basis: a five-week investigation and alleged violations of security policy including "exfiltration of confidential business-sensitive documents" — Google's claim, but a concrete one, not a vague culture-fit line [4].
The work continued. Google reorganized responsible-AI research under Marian Croak, changed its publication and exit processes after the internal review, and kept publishing in the field. Nothing here was adjudicated against Google — no court or regulator has ruled on any of it [3] [4].
YOU DECIDE
Scoped to the claims. That the paper was ordered retracted, that both co-leads were gone within eleven weeks, and that Pichai apologized for the handling — on the record. Whether Gebru resigned or was fired is a genuine dispute; this page will not resolve it for you. Whether a publication-quality process just happened to fire its first fatal round at the in-house critics of the company's flagship technology — that is the question the timeline puts in your hands.
Weigh the costly signal: the company kept the ethics team's name and org box. What it did not keep was either person the world associated with it.
The archive does not judge. It keeps the timestamps.
evidence locker
PRIMARY & INVESTIGATIVE
PRESS & ANALYSIS
BBC News — "Margaret Mitchell: Google fires AI ethics founder" (Feb 2021) ATTRIBUTED — the firing, Google's stated reasons, the Croak reorganization.
bbc.com/news/technology-56135817
The standard. The paper dispute, both exits, and the reorganization are sourced to named reporting and to Google's own published email. The fired-vs-resigned dispute is presented as a dispute; Google's stated rationale and Dean's email are carried at full strength. Google's stated reasons for the Mitchell firing are quoted as Google's claims. The "safety theater" framing is our analytic read of the pattern, labeled as ours, and is not asserted as Google's intent. Nothing on this page is an adjudicated finding, and no personal wrongdoing by any named individual is alleged. If it couldn't survive a defamation challenge, it wouldn't be on this page.