The warnings Google pushed Timnit Gebru out over have a receipt for almost every one

The paper, before the thread reacts to it

The post revisits "On the Dangers of Stochastic Parrots," the 2020 paper whose internal fight ended with Timnit Gebru leaving Google, and grades its predictions against the years since. The argument is that the paper called the shots before the vocabulary existed — "it predicted the hallucination problem before anyone had a word for it." The specific warnings and the receipts the post attaches:

• Fluency without understanding → the hallucination crisis, named years later.
• Bias amplification → Amazon's hiring algorithm penalizing resumes containing the word "women"; US hospital risk scores "systematically underestimat[ing] the medical needs of Black patients"; Apple Card credit lines reportedly 10x lower for wives than husbands.
• Environmental cost → the paper's estimate that training one large model emitted the lifetime output of 5 cars; by 2024 "Google's emissions were up 48% from 2019, and the company explicitly blamed AI infrastructure."
• Unauditable datasets → the LAION-5B set behind Stable Diffusion later found to contain thousands of CSAM images.
• Model collapse / centralized voice → the claim that a large share of new English web text is now AI-generated.

Where the thread pushes back

This is not a one-note "she was right" chorus. The most-engaged objection is about the framing, not the science. A reader who followed the saga closely insisted on the distinction Google itself made: Gebru quit rather than was fired, after an ultimatum over the paper. Others questioned the article's choices — one asked why it needed to mention Gebru is Black if race had no bearing on the content.

Simon Willison was struck enough to quote the receipts back — the Amazon, healthcare and Apple Card examples — as the part that aged hardest. And a systems engineer reframed the real hazard:

"The main issue with AI is often not what sci-fi or EA-adjacent prophets are trying to warn us about, but the insidious dangers of the failure modes." — a developer on the unglamorous risk

That's the through-line the paper anticipated: not a superintelligence, but a system that "seems capable but fails in surprising ways," deployed into hiring and healthcare faster than anyone audits it. Even sympathetic readers conceded the article never says what to actually do about it.

The uncomfortable scorecard

You can argue Gebru's exit was a resignation, and you can dislike the post's tone. What's harder to argue is the ledger: the four technical warnings each have a named, dated incident attached. The lesson the thread keeps circling isn't that one researcher was prescient — it's that the failure modes she flagged were boring, predictable, and shipped anyway.