"Timnit Gebru was fired from Google in December 2020 for refusing to retract a research paper, and every single warning that paper made about large language models has now happened at a scale the industry spent 4 years trying to make people forget about.
Her name is Timnit Gebru.
She co-led the Ethical AI team at Google. She co-wrote a paper called "On the Dangers of Stochastic Parrots" with Emily Bender at the University of Washington and two other researchers. The paper was 14 pages long. It was submitted to a top AI ethics conference. And it was the reason Google decided that one of the most senior Black women in AI research could no longer work there.
The story Google told publicly was that she resigned. The story she told, confirmed by 2,695 of her colleagues in an open letter, was that she was fired by email while on vacation because she refused to either retract the paper or remove her name from it.
The paper had not even been published yet.
Here is what she actually wrote, and why every prediction inside it has now come true.
The first warning was about scale itself. Bender and Gebru argued that training ever-larger models on ever-larger scrapes of the internet would produce systems that appeared fluent but had no actual understanding of language. They called these systems stochastic parrots because they would repeat patterns from training data with statistical confidence and zero comprehension. The paper predicted that this apparent intelligence would fool both users and developers into trusting outputs that were structurally incapable of being reliable.
This was 2020. GPT-3 had just come out. The paper predicted the hallucination problem before anyone had a word for it.
The second warning was about bias amplification. The paper documented in detail that internet-scale training data contains systematic overrepresentation of dominant viewpoints and underrepresentation of marginalized ones. The models would not just absorb this bias. They would amplify it, because the optimization process rewards confident outputs, and confidence in language patterns tracks frequency in the training set.
The prediction was that hiring tools built on these models would discriminate against women. That healthcare triage tools would underperform on Black patients. That loan approval systems would entrench inequality while presenting their decisions as neutral algorithmic judgment.
Every one of those things has now been documented in deployment.
Amazon's hiring algorithm penalized resumes that contained the word "women" in any context. Healthcare risk scoring algorithms used by major US hospitals were found to systematically underestimate the medical needs of Black patients. Apple Card's credit algorithm gave wives credit lines 10x lower than their husbands for the same financial profile.
The third warning was about environmental cost. The paper calculated that training a single large language model produced emissions equivalent to the lifetime output of 5 cars. The prediction was that the race to scale would create an environmental footprint that would eventually rival entire industries.
In 2024, Google's emissions were up 48% from 2019, and the company explicitly blamed AI infrastructure. Microsoft's were up 29%, same reason. Both companies have now quietly abandoned the climate commitments they were publicly celebrating the year Gebru was fired.
The fourth warning was about documentation. The paper argued that the training datasets being assembled were too large for anyone to actually audit. Nobody at Google, OpenAI, Meta, or any other lab could tell you with confidence what was in the data their models were trained on. This was not a temporary problem to be solved later. It was a permanent feature of the approach.
In 2023, researchers discovered that the LAION-5B dataset, used to train Stable Diffusion and other major image models, contained thousands of images of child sexual abuse material. The companies that had trained on the dataset had no way of knowing. The paper predicted that category of failure 3 years before it was found.
The fifth warning was the one Google cared about most.
Bender and Gebru argued that the deployment of these systems would centralize linguistic and cultural power in the hands of the small number of companies that could afford to train them. The internet would become a place where the dominant voice was a statistical average of dominant voices, presented as a neutral assistant. Languages underrepresented in the training data would degrade over time as more web content was generated by these systems and fed back into the next training run.
This is now happening in real time. A 2024 study found that 57% of new web content in English is AI-generated or AI-assisted. Researchers studying low-resource languages have documented active degradation in translation quality, because the synthetic content fed back into training is itself worse in those languages.
The paper Google fired her for predicted the model collapse problem before model collapse had a name.
The mechanism behind why this all happened is the part of her work that nobody quotes.
Gebru's argument was not that AI is dangerous in some abstract sci-fi sense. Her argument was that AI is dangerous in a very specific structural sense. The technology was being built by a small group of researchers who shared similar backgrounds, worked at similar companies, and were rewarded for shipping products faster than competitors. The incentive structure made it impossible for safety, ethics, and bias concerns to slow anything down. Anyone inside the system who raised those concerns was either ignored, sidelined, or removed.
She was making that argument from inside Google.
Then Google proved her right by removing her.
The team Google had built to make sure their AI was safe was dismantled in 90 days because they did the job they had been hired to do. Margaret Mitchell, the other co-lead of the Ethical AI team, was fired two months after Gebru for searching through her own emails for evidence of how Gebru had been treated.
Gebru did not stop. She founded DAIR, the Distributed AI Research Institute, in 2021. The mission is to do AI research outside the control of the companies that have a financial interest in not hearing the answers.
Every prediction in the Stochastic Parrots paper has now been validated by deployment. Hallucinations are an industry-wide problem the largest labs cannot solve. Bias amplification has been documented in hiring, healthcare, lending, and criminal justice. Environmental costs are larger than entire small countries. Training data audits remain impossible. Model collapse is an active research crisis at every major lab.
The question worth sitting with is the one almost no one in the industry will say out loud.
Every researcher with the technical credibility to call out these problems watched what happened to her in December 2020 and made a calculation about their own career. The number of people willing to speak publicly about safety and ethics issues inside the major AI labs collapsed after that firing and has not recovered.
The researcher Google fired for warning about exactly what is now happening was right.
The company that fired her is now the second-largest deployer of the technology she warned about.
And the people inside that company who agree with her are not allowed to say so."
+
Rebecca Solnit
Read this. The link to the paper discussed is here: https://dl.acm.org/doi/epdf/10.1145/3442188.3445922
Back in 2017 Long Island Ice Tea — known for its undistinguished, barely drinkable sugar-water — changed its name to “Long Blockchain Corp.” Its shares surged to a peak of 400% over their pre-announcement price. The company announced no specific integrations with any kind of blockchain, nor has it made any such integrations since.
If you’d like an essay-formatted version of this post to read or share, here’s a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
LBCC was subsequently delisted from NASDAQ after settling with the SEC over fraudulent investor statements. Today, the company trades over the counter and its market cap is $36m, down from $138m.
The most remarkable thing about this incredibly stupid story is that LBCC wasn’t the peak of the blockchain bubble — rather, it was the start of blockchain’s final pump-and-dump. By the standards of 2022’s blockchain grifters, LBCC was small potatoes, a mere $138m sugar-water grift.
They didn’t have any NFTs, no wash trades, no ICO. They didn’t have a Superbowl ad. They didn’t steal billions from mom-and-pop investors while proclaiming themselves to be “Effective Altruists.” They didn’t channel hundreds of millions to election campaigns through straw donations and other forms of campaing finance frauds. They didn’t even open a crypto-themed hamburger restaurant where you couldn’t buy hamburgers with crypto:
They were amateurs. Their attempt to “make fetch happen” only succeeded for a brief instant. By contrast, the superpredators of the crypto bubble were able to make fetch happen over an improbably long timescale, deploying the most powerful reality distortion fields since Pets.com.
Anything that can’t go on forever will eventually stop. We’re told that trillions of dollars’ worth of crypto has been wiped out over the past year, but these losses are nowhere to be seen in the real economy — because the “wealth” that was wiped out by the crypto bubble’s bursting never existed in the first place.
Like any Ponzi scheme, crypto was a way to separate normies from their savings through the pretense that they were “investing” in a vast enterprise — but the only real money (“fiat” in cryptospeak) in the system was the hardscrabble retirement savings of working people, which the bubble’s energetic inflaters swapped for illiquid, worthless shitcoins.
We’ve stopped believing in the illusory billions. Sam Bankman-Fried is under house arrest. But the people who gave him money — and the nimbler Ponzi artists who evaded arrest — are looking for new scams to separate the marks from their money.
Take Morganstanley, who spent 2021 and 2022 hyping cryptocurrency as a massive growth opportunity:
Today, Morganstanley wants you to know that AI is a $6 trillion opportunity.
They’re not alone. The CEOs of Endeavor, Buzzfeed, Microsoft, Spotify, Youtube, Snap, Sports Illustrated, and CAA are all out there, pumping up the AI bubble with every hour that god sends, declaring that the future is AI.
Google and Bing are locked in an arms-race to see whose search engine can attain the speediest, most profound enshittification via chatbot, replacing links to web-pages with florid paragraphs composed by fully automated, supremely confident liars:
Blockchain was a solution in search of a problem. So is AI. Yes, Buzzfeed will be able to reduce its wage-bill by automating its personality quiz vertical, and Spotify’s “AI DJ” will produce slightly less terrible playlists (at least, to the extent that Spotify doesn’t put its thumb on the scales by inserting tracks into the playlists whose only fitness factor is that someone paid to boost them).
But even if you add all of this up, double it, square it, and add a billion dollar confidence interval, it still doesn’t add up to what Bank Of America analysts called “a defining moment — like the internet in the ’90s.” For one thing, the most exciting part of the “internet in the ‘90s” was that it had incredibly low barriers to entry and wasn’t dominated by large companies — indeed, it had them running scared.
The AI bubble, by contrast, is being inflated by massive incumbents, whose excitement boils down to “This will let the biggest companies get much, much bigger and the rest of you can go fuck yourselves.” Some revolution.
AI has all the hallmarks of a classic pump-and-dump, starting with terminology. AI isn’t “artificial” and it’s not “intelligent.” “Machine learning” doesn’t learn. On this week’s Trashfuture podcast, they made an excellent (and profane and hilarious) case that ChatGPT is best understood as a sophisticated form of autocomplete — not our new robot overlord.
We all know that autocomplete is a decidedly mixed blessing. Like all statistical inference tools, autocomplete is profoundly conservative — it wants you to do the same thing tomorrow as you did yesterday (that’s why “sophisticated” ad retargeting ads show you ads for shoes in response to your search for shoes). If the word you type after “hey” is usually “hon” then the next time you type “hey,” autocomplete will be ready to fill in your typical following word — even if this time you want to type “hey stop texting me you freak”:
And when autocomplete encounters a new input — when you try to type something you’ve never typed before — it tries to get you to finish your sentence with the statistically median thing that everyone would type next, on average. Usually that produces something utterly bland, but sometimes the results can be hilarious. Back in 2018, I started to text our babysitter with “hey are you free to sit” only to have Android finish the sentence with “on my face” (not something I’d ever typed!):
Modern autocomplete can produce long passages of text in response to prompts, but it is every bit as unreliable as 2018 Android SMS autocomplete, as Alexander Hanff discovered when ChatGPT informed him that he was dead, even generating a plausible URL for a link to a nonexistent obit in The Guardian:
Of course, the carnival barkers of the AI pump-and-dump insist that this is all a feature, not a bug. If autocomplete says stupid, wrong things with total confidence, that’s because “AI” is becoming more human, because humans also say stupid, wrong things with total confidence.
Exhibit A is the billionaire AI grifter Sam Altman, CEO if OpenAI — a company whose products are not open, nor are they artificial, nor are they intelligent. Altman celebrated the release of ChatGPT by tweeting “i am a stochastic parrot, and so r u.”
This was a dig at the “stochastic parrots” paper, a comprehensive, measured roundup of criticisms of AI that led Google to fire Timnit Gebru, a respected AI researcher, for having the audacity to point out the Emperor’s New Clothes:
Gebru’s co-author on the Parrots paper was Emily M Bender, a computational linguistics specialist at UW, who is one of the best-informed and most damning critics of AI hype. You can get a good sense of her position from Elizabeth Weil’s New York Magazine profile:
Bender has made many important scholarly contributions to her field, but she is also famous for her rules of thumb, which caution her fellow scientists not to get high on their own supply:
Please do not conflate word form and meaning
Mind your own credulity
As Bender says, we’ve made “machines that can mindlessly generate text, but we haven’t learned how to stop imagining the mind behind it.” One potential tonic against this fallacy is to follow an Italian MP’s suggestion and replace “AI” with “SALAMI” (“Systematic Approaches to Learning Algorithms and Machine Inferences”). It’s a lot easier to keep a clear head when someone asks you, “Is this SALAMI intelligent? Can this SALAMI write a novel? Does this SALAMI deserve human rights?”
Bender’s most famous contribution is the “stochastic parrot,” a construct that “just probabilistically spits out words.” AI bros like Altman love the stochastic parrot, and are hellbent on reducing human beings to stochastic parrots, which will allow them to declare that their chatbots have feature-parity with human beings.
At the same time, Altman and Co are strangely afraid of their creations. It’s possible that this is just a shuck: “I have made something so powerful that it could destroy humanity! Luckily, I am a wise steward of this thing, so it’s fine. But boy, it sure is powerful!”
They’ve been playing this game for a long time. People like Elon Musk (an investor in OpenAI, who is hoping to convince the EU Commission and FTC that he can fire all of Twitter’s human moderators and replace them with chatbots without violating EU law or the FTC’s consent decree) keep warning us that AI will destroy us unless we tame it.
There’s a lot of credulous repetition of these claims, and not just by AI’s boosters. AI critics are also prone to engaging in what Lee Vinsel calls criti-hype: criticizing something by repeating its boosters’ claims without interrogating them to see if they’re true:
There are better ways to respond to Elon Musk warning us that AIs will emulsify the planet and use human beings for food than to shout, “Look at how irresponsible this wizard is being! He made a Frankenstein’s Monster that will kill us all!” Like, we could point out that of all the things Elon Musk is profoundly wrong about, he is most wrong about the philosophical meaning of Wachowksi movies:
But even if we take the bros at their word when they proclaim themselves to be terrified of “existential risk” from AI, we can find better explanations by seeking out other phenomena that might be triggering their dread. As Charlie Stross points out, corporations are Slow AIs, autonomous artificial lifeforms that consistently do the wrong thing even when the people who nominally run them try to steer them in better directions:
Imagine the existential horror of a ultra-rich manbaby who nominally leads a company, but can’t get it to follow: “everyone thinks I’m in charge, but I’m actually being driven by the Slow AI, serving as its sock puppet on some days, its golem on others.”
Ted Chiang nailed this back in 2017 (the same year of the Long Island Blockchain Company):
There’s a saying, popularized by Fredric Jameson, that it’s easier to imagine the end of the world than to imagine the end of capitalism. It’s no surprise that Silicon Valley capitalists don’t want to think about capitalism ending. What’s unexpected is that the way they envision the world ending is through a form of unchecked capitalism, disguised as a superintelligent AI. They have unconsciously created a devil in their own image, a boogeyman whose excesses are precisely their own.
Chiang is still writing some of the best critical work on “AI.” His February article in the New Yorker, “ChatGPT Is a Blurry JPEG of the Web,” was an instant classic:
[AI] hallucinations are compression artifacts, but — like the incorrect labels generated by the Xerox photocopier — they are plausible enough that identifying them requires comparing them against the originals, which in this case means either the Web or our own knowledge of the world.
“AI” is practically purpose-built for inflating another hype-bubble, excelling as it does at producing party-tricks — plausible essays, weird images, voice impersonations. But as Princeton’s Matthew Salganik writes, there’s a world of difference between “cool” and “tool”:
Nature can claim “conversational AI is a game-changer for science” but “there is a huge gap between writing funny instructions for removing food from home electronics and doing scientific research.” Salganik tried to get ChatGPT to help him with the most banal of scholarly tasks — aiding him in peer reviewing a colleague’s paper. The result? “ChatGPT didn’t help me do peer review at all; not one little bit.”
The criti-hype isn’t limited to ChatGPT, of course — there’s plenty of (justifiable) concern about image and voice generators and their impact on creative labor markets, but that concern is often expressed in ways that amplify the self-serving claims of the companies hoping to inflate the hype machine.
One of the best critical responses to the question of image- and voice-generators comes from Kirby Ferguson, whose final Everything Is a Remix video is a superb, visually stunning, brilliantly argued critique of these systems:
https://www.youtube.com/watch?v=rswxcDyotXA
One area where Ferguson shines is in thinking through the copyright question — is there any right to decide who can study the art you make? Except in some edge cases, these systems don’t store copies of the images they analyze, nor do they reproduce them:
For creators, the important material question raised by these systems is economic, not creative: will our bosses use them to erode our wages? That is a very important question, and as far as our bosses are concerned, the answer is a resounding yes.
Markets value automation primarily because automation allows capitalists to pay workers less. The textile factory owners who purchased automatic looms weren’t interested in giving their workers raises and shorting working days.
‘
They wanted to fire their skilled workers and replace them with small children kidnapped out of orphanages and indentured for a decade, starved and beaten and forced to work, even after they were mangled by the machines. Fun fact: Oliver Twist was based on the bestselling memoir of Robert Blincoe, a child who survived his decade of forced labor:
Today, voice actors sitting down to record for games companies are forced to begin each session with “My name is ______ and I hereby grant irrevocable permission to train an AI with my voice and use it any way you see fit.”
Let’s be clear here: there is — at present — no firmly established copyright over voiceprints. The “right” that voice actors are signing away as a non-negotiable condition of doing their jobs for giant, powerful monopolists doesn’t even exist. When a corporation makes a worker surrender this right, they are betting that this right will be created later in the name of “artists’ rights” — and that they will then be able to harvest this right and use it to fire the artists who fought so hard for it.
There are other approaches to this. We could support the US Copyright Office’s position that machine-generated works are not works of human creative authorship and are thus not eligible for copyright — so if corporations wanted to control their products, they’d have to hire humans to make them:
Or we could create collective rights that belong to all artists and can’t be signed away to a corporation. That’s how the right to record other musicians’ songs work — and it’s why Taylor Swift was able to re-record the masters that were sold out from under her by evil private-equity bros::
Whatever we do as creative workers and as humans entitled to a decent life, we can’t afford drink the Blockchain Iced Tea. That means that we have to be technically competent, to understand how the stochastic parrot works, and to make sure our criticism doesn’t just repeat the marketing copy of the latest pump-and-dump.
Today (Mar 9), you can catch me in person in Austin at the UT School of Design and Creative Technologies, and remotely at U Manitoba’s Ethics of Emerging Tech Lecture.
Tomorrow (Mar 10), Rebecca Giblin and I kick off the SXSW reading series.
CC BY 3.0
https://creativecommons.org/licenses/by/3.0/deed.en
[Image ID: A graph depicting the Gartner hype cycle. A pair of HAL 9000's glowing red eyes are chasing each other down the slope from the Peak of Inflated Expectations to join another one that is at rest in the Trough of Disillusionment. It, in turn, sits atop a vast cairn of HAL 9000 eyes that are piled in a rough pyramid that extends below the graph to a distance of several times its height.]
[ID: Google fired Timnit Gebru for a paper that predicted every AI failure the industry has spent five years documenting The Stochastic Parrots paper named hallucinations, bias amplification, environmental cost, unauditable training data, and model collapse as inevitable consequences of scaling, and the industry has spent four years confirming each one while trying to forget who said it first. The paper and the firing: She co-led Google's Ethical AI team and co-wrote "On the Dangers of Stochastic Parrots" with Emily Bender and two other researchers. Google publicly claimed she resigned. She said otherwise, backed by 2,695 colleagues in an open letter: she was fired by email while on vacation for refusing to retract the paper or remove her name from it. It had not been published yet. The predictions: The paper argued that internet-scale training produces systems mimicking fluency without comprehension, which it termed stochastic parrots, and predicted hallucinations before the industry had a word for them. It documented that training data bias would amplify in deployment, calculated that a single large model training emitted as much carbon as five cars over their lifetimes, and warned that datasets would grow too large to audit. It predicted model collapse before the phenomenon had a name. The receipts: Amazon's hiring algorithm penalized resumes containing the word "women." Healthcare scoring systems underestimated Black patients' medical needs. Apple Card gave wives credit lines ten times lower than their husbands' on identical profiles. Researchers found the LAION-5B dataset used to train Stable Diffusion contained thousands of images of child sexual abuse material. Google's emissions rose 48% from 2019 to 2024 and Microsoft's 29%, both blaming AI infrastructure; both have since abandoned the climate commitments they were publicly celebrating the year she was fired. The structure: Gebru's core argument concerned incentive structure. The technology was built by people rewarded for shipping fast, and anyone raising safety concerns was removed. She made that argument from inside Google. Google proved it by firing her, then fired Margaret Mitchell, the other co-lead of the Ethical AI team, two months later for searching her own emails for evidence of how Gebru had been treated. She founded DAIR in 2021 to do AI research outside the control of companies with a financial interest in not hearing the answers. https://dl.acm.org/doi/pdf/10.1145/3442188.3445922
— 🔻agitprop + absurdity🔻 (@agtprpnabsrdty) May 27, 2026]
Trailblazers like Aisha Bowe, a former NASA rocket scientist and CEO of STEMBoard, and Kimberly Bryant, founder of Black Girls CODE, are breaking barriers and inspiring the next generation. From creating accessible STEM education tools to bridging gender and racial gaps in tech, these women exemplify excellence. Here are 5 Black female CEOs leading the charge in tech.
Aisha Bowe
Aisha Bowe is a…
