If we go back to 2020, most AI professionals were still discussing just how powerful GPT-3 was.
At that time, generative AI had not yet become a global focus; ChatGPT was still two years away from being released, and large models had not yet sparked the sweeping global investment frenzy we see today. Yet, in that very year, a top Google AI researcher clashed fiercely with the company over an unpublished paper and ultimately lost his job.
At the time, many believed this was just another Silicon Valley controversy over workplace management, academic publishing, and corporate culture; but looking back now, people realize that the warnings in that paper have almost all come true in the real world.
The researcher who was fired is Timnit Gebru, one of the most influential figures in the field of AI ethics.
A shocking firing event that shook the AI community
In December 2020, Timnit Gebru announced on social media that she had been fired by Google.
The news quickly ignited the entire AI research community. At the time, Gebru was not an ordinary researcher but a co-lead of Google’s Ethical AI Team and one of the world’s leading scholars in AI fairness and algorithmic bias.
Gebru, who was born in Ethiopia, has long focused on issues of racial bias, gender discrimination, and social equity in AI. Before joining Google, she conducted research at Stanford University. In 2018, a study on algorithmic bias that she co-authored was widely regarded as a pivotal moment in AI fairness research. Later that same year, Google recruited her and publicly highlighted the company’s commitment to “Responsible AI.”
Yet, just two years later, the two parties parted ways.
At the time, Google publicly stated that Gebru had resigned voluntarily, but Gebru offered a completely different account: she said she received an email from the company while on vacation, informing her that her departure was effective immediately, and that all access to internal systems and her email account was simultaneously revoked.
In her view, this was an unquestionable termination.
Subsequently, over 4,000 Google employees and industry professionals signed an open letter questioning the company’s handling of the situation and demanding Gebru’s reinstatement—all sparked by a 14-page academic paper.
A 14-page paper has sparked controversy.
The paper, titled "On the Dangers of Stochastic Parrots," authored by Timnit Gebru, University of Washington linguistics professor Emily Bender, and two other researchers, has been cited over 14,000 times to date.
Later, the term "stochastic parrot" also became widely known. (Paper link: https://s10251.pcdn.co/pdf/2021-bender-parrots.pdf)
The paper points out that large language models essentially reproduce language patterns based on statistical regularities: they can generate fluent, natural, and even logically coherent text, but do not truly understand the meaning of language—much like a parrot that has learned to mimic human speech, appearing intelligent but merely replicating patterns derived from vast amounts of internet text. Since the internet itself is filled with bias, discrimination, and hateful content, large models are highly likely to absorb these issues and amplify them in their generated outputs.
Keep in mind that this was 2020, when GPT-3 had just been released, ChatGPT had not yet been created, and the large model boom was still far off—this paper had already anticipated one of today’s industry’s most pressing challenges.
After the paper was submitted to a top AI ethics conference, Google management requested that it be withdrawn or that the names of Google researchers be removed. Gebru refused, demanding that the company specify its reasons and proposing further discussion between both parties.
At the same time, she sent a strongly worded email to Google’s internal employee group.
In the email, Gebru criticized Google for lacking concrete action in promoting diversity among underrepresented groups and addressing internal inequalities. She wrote, “When you start speaking up for marginalized groups, your situation only gets worse. You make other leaders uncomfortable.” She also stated that if the company continues to fail to explain why the paper was withdrawn, she would resign at the appropriate time.
The situation escalated far beyond her expectations. Gebru said Google subsequently responded that it would not meet her demands and immediately accepted her "resignation," revoking all of her access rights.
At the time, the incident quickly became one of the most controversial topics in the global AI community.
What once seemed like an radical view has now become reality.
What has kept this incident under discussion to this day is not the firing itself, but the content of the paper—because looking back today, nearly every concern it raised has become a real issue facing the AI industry.
(1) First alert: The model may hallucinate
In 2020, GPT-3 was just released. People marveled at the model’s ability to generate text, but few seriously discussed its reliability.
Gebru and Bender point out that as models grow larger, people will increasingly mistake fluent expression for genuine understanding. The models appear to be thinking, but they are merely predicting the next most likely word; thus, they will eventually generate information that seems plausible but is entirely incorrect.
Today, this issue has a name familiar to everyone: AI hallucination. Whether it’s ChatGPT, Gemini, Claude, or other advanced models, the problem of hallucination has yet to be fully resolved.
In a sense, the paper accurately anticipated the concept of “hallucination” before it became an industry buzzword.
(2) Second alert: Bias will not disappear—it will be amplified.
The paper also points out that the internet itself is not a neutral source of data, and training data inherently contains various racial, gender, cultural, and geographic biases. Models not only learn these biases but may also reinforce them further due to optimization mechanisms.
Later, various practical issues validated this concern:
Amazon once attempted to use AI to screen job resumes, but the system automatically lowered the scores of resumes containing keywords such as "women."
A medical risk assessment system used by several major U.S. hospitals was found to have consistently underestimated the healthcare needs of Black patients.
The Apple Card also drew regulatory attention for granting women significantly lower credit limits than men.
These cases illustrate that algorithms do not automatically ensure fairness; instead, they may inadvertently entrench real-world inequalities in more subtle ways.
(3) Third alert: AI's energy consumption will become a new issue
In 2020, computational costs were far from the focus they are today, but that paper had already begun discussing the environmental impact of training extremely large models. According to researchers’ calculations, the carbon emissions from training a single large language model were equivalent to the total lifecycle emissions of five cars—a claim that many at the time considered overly pessimistic.
However, as AI infrastructure development enters an arms race, problems have quickly emerged: according to Google’s publicly disclosed data, the company’s greenhouse gas emissions increased by 48% from 2019 to 2024; Microsoft saw a similar increase of approximately 29% over the same period. Both companies have explicitly identified AI data centers and computing infrastructure as significant contributing factors.
Ironically, these tech giants were loudly promoting carbon neutrality goals just a few years ago.
(4) Fourth alert: No one truly knows what’s in the training data
To many, training data seems like merely an engineering issue. But Gebru believes that as data scales up, conducting a complete audit of training data will become nearly impossible.
Her prediction proved correct again: in 2023, researchers discovered a large number of child abuse images within LAION-5B, a dataset widely used to train image generation models, including several mainstream models such as Stable Diffusion.
As expected, many developers were previously unaware of the existence of this content. In other words, even model developers themselves may not truly understand what the model is “fed”—which is precisely one of the key questions originally raised in the paper.
(5) Fifth Alert: The Internet Will Gradually Be Dominated by AI-Generated Content
To Google, this may be the most sensitive part of the entire paper. Gebru and Bender argue that the development of large models will ultimately concentrate control over language and culture in the hands of a handful of tech giants. The reason is simple: training ultra-large models requires massive funding, computational power, and data resources, leaving only a very limited number of companies truly capable of competing.
Over time, the dominant voice on the internet will gradually evolve into statistical averages trained by a few companies, then disseminated worldwide under the guise of a “neutral assistant.” Meanwhile, languages and cultures underrepresented in training data will become further marginalized.
More seriously, when AI-generated content re-enters the internet and becomes part of the next training dataset, the problem continuously reinforces itself—this is precisely what researchers today refer to as “model collapse.”
A 2024 study found that approximately 57% of new content added to the English internet is AI-generated or AI-assisted; research on low-resource languages has revealed that, due to training data increasingly coming from AI-generated content, translation quality for some languages has noticeably degraded.
In other words, this paper not only predicted the "model collapse" phenomenon but also identified its underlying mechanisms even before the concept was formally introduced.
After leaving Google, she chose to continue her research.
After the incident, many people later described Gebru as an “anti-AI” figure. But that’s not true—she never advocated for halting AI development. From the beginning, what she questioned was something else:
Who is really deciding the direction of AI?
In her view, researchers and executives driving the development of large models often share similar backgrounds, serve similar business goals, and are driven by the same competitive pressures. Under such incentive structures, releasing products faster, scaling user bases more quickly, and winning market competition often take higher priority than safety, fairness, and ethical concerns.
And anyone who tries to slow this process may be seen as an obstacle. Ironically, Gebru raised this very point within Google itself, and Google’s decision to fire her gave this perspective its most dramatic real-world validation.
More sadly, shortly after the incident, Margaret Mitchell, co-lead of the ethical AI team, was also fired—within just 90 days, Google’s once-proud ethical AI team was essentially dismantled.
After leaving Google, Gebru founded the Distributed AI Research Institute (DAIR) in 2021. Unlike large tech companies, this organization aims to conduct AI research beyond commercial interests, with a straightforward goal: to investigate issues that tech giants may be unwilling to address. Over the past few years, DAIR has consistently focused on topics such as data sourcing, algorithmic fairness, linguistic diversity, and the concentration of power in the AI industry.
As generative AI has experienced explosive growth, an increasing number of researchers have begun revisiting the paper “On the Dangers of Stochastic Parrots,” as they recognize that issues once considered overly cautious in the paper have now become everyday realities in the industry.
Perhaps she simply saw the issue earlier than others.
Six years later, the public may never arrive at a universally agreed-upon answer regarding the dispute between Timnit Gebru and Google.
Google viewed it as a normal academic review and departure; Gebru believed she was silenced for insisting on publishing her research. But one thing has become increasingly hard to deny:
The paper that led her to leave Google did not lose its significance when the controversy ended.
On the contrary, the issues it discusses—illusions, biases, data pollution, environmental costs, model collapse, and centralization of power—have now become unavoidable topics across the entire AI industry.
Sometimes, history offers its evaluation in unexpected ways.
In 2020, many people thought Timnit Gebru was too pessimistic;
In 2026, people began to realize that she may have simply seen the problem earlier than others.
Reference link: https://www.tumblr.com/dreaminginthedeepsouth/817865966907228160/darren-oconnor-timnit-gebru-was-fired-from
This article is from the WeChat public account "CSDN," compiled by Zheng Liyuan.