Timnit Gebru: The Researcher Who Wouldn't Back Down

We need research that is rooted in the communities most affected by AI, not in the corporations that profit from it.

Addis Ababa, Ethiopia. The 1980s. A girl is growing up in a country in the grip of a brutal civil war, in a family that is educated and aspirational in the specific way of East African middle-class families — intensely focused on education as the path to a different kind of life. Her father is a physician. Her mother is a teacher. They have specific hopes for their daughter.

She is clever. She is also, from an early age, uncomfortable with authority — specifically uncomfortable with the kind of authority that demands compliance without explanation, that treats dissent as disloyalty, that requires you to agree with things you do not believe.

This quality will serve her in some ways and cost her in others.

Her name is Timnit Gebru. By 2021, she will be one of the most recognisable figures in AI research — not primarily for her technical contributions, though those are significant, but for the specific way she lost her job at Google, and for what she built from the wreckage.

---

Ethiopia, Stanford, and the Formation of a Researcher

Timnit Gebru was born in Addis Ababa, Ethiopia, in 1983. Her family fled the civil war in the early 1990s, eventually landing in the United States as refugees. She grew up partly in Virginia, navigating the specific experience of being an Ethiopian refugee in the American education system — the language difficulties, the cultural dislocation, the need to establish credibility in a context that did not automatically recognise the credentials she came from.

She was admitted to Stanford University, where she studied electrical engineering and moved into computer science. The trajectory was ambitious but not uncommon for the children of educated immigrants who had arrived in the United States with strong educational foundations and strong aspirations.

At Stanford, she developed an interest in machine learning — the mathematical foundations of pattern recognition in data — and in the specific problems at the intersection of machine learning and human welfare. The questions that drew her were not primarily the technical questions of how to build more accurate classifiers or more efficient learning algorithms. They were questions about what those classifiers were doing in the world, whose interests they served, and whose they damaged.

She completed her PhD in electrical engineering at Stanford in 2017, with a dissertation on fine-grained recognition methods — computer vision techniques for distinguishing between similar categories of objects. The technical work was solid and published in top venues. But even in this early work, her interest was not purely technical: the specific applications of fine-grained recognition that interested her included applications in domains where the social implications mattered — in healthcare, in accessibility technology, in the recognition of faces across demographic groups.

---

The Face Recognition Work: From Technical to Political

In 2018, Gebru co-authored, with Joy Buolamwini, the “Gender Shades” paper — the foundational study of demographic bias in commercial facial recognition systems that was discussed in the article on AI bias earlier in this series.

The Gender Shades collaboration was formative in several ways. It was a demonstration that technical AI research and social impact research could be combined — that the question “does this system work equitably for all demographic groups?” was a technical question that required technical methodology to answer. It was also a demonstration that the answer to that question, when the methodology was applied, was often alarming: the commercial facial recognition systems that were being deployed in real-world applications worked dramatically less well for darker-skinned women than for lighter-skinned men.

The reception of the Gender Shades work was itself instructive. The paper was technically rigorous, the findings were reproducible, and the methodology was straightforward. Three of the largest technology companies — Microsoft, IBM, and Face++ — were identified as having significant demographic disparities in their commercial facial recognition systems. The paper attracted extraordinary media attention and generated significant public debate.

The companies’ responses varied. Microsoft and IBM both updated their systems and reported improved performance after the publication. IBM eventually withdrew its facial recognition system from commercial availability. The willingness of major companies to respond substantively to documented evidence of bias was itself significant.

But the episode also taught Gebru something about the politics of AI ethics research. The attention her work attracted was welcome, but it also made her a target — a person whose research was capable of causing commercial and reputational harm to major technology companies. This attention was something she would continue to navigate, in increasingly difficult circumstances, for the next several years.

---

Google Brain and the Ethics Team

In 2018, Gebru joined Google as a researcher in the AI ethics organisation — a role that placed her inside one of the most powerful AI organisations in the world, with access to systems and resources that no independent researcher could access.

The role was an expression of Google’s specific response to the growing AI ethics conversation. By 2018, the company was under significant pressure from researchers, activists, and employees to demonstrate that it was taking the ethical dimensions of AI seriously. The AI ethics team — which Gebru joined and eventually co-led — was the institutional embodiment of that commitment.

The team’s work addressed the specific questions that were most actively discussed in AI ethics: bias and fairness, transparency and interpretability, the social and labour implications of AI deployment, and the governance of AI systems in high-stakes applications. These were genuine research questions, and the team produced genuine research.

But the team’s position inside Google also created specific tensions. Google was a commercial organisation with significant commercial interests in deploying AI systems that the ethics team was sometimes studying critically. The interests of the ethics team — in producing honest research about the social implications of AI — and the interests of Google’s product and commercial teams — in deploying AI systems that generated revenue — were not always aligned.

Gebru navigated these tensions for two years, producing research on facial analysis, on the social implications of NLP systems, and on the specific challenges of building AI systems for low-resource languages — languages with limited training data — that served populations outside the dominant English-speaking markets of AI development.

The tension became irresolvable in 2020, over a paper about large language models.

---

The Paper: Stochastic Parrots

The paper at the centre of Gebru’s firing from Google was titled “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” It was co-authored by Gebru, Emily Bender, Angelina McMillan-Major, and Margaret Mitchell (who was also on Google’s AI ethics team).

The paper was a critical analysis of large language models — the systems like GPT-3 that had become the dominant paradigm in NLP and that Google itself was heavily invested in developing. The paper made several specific arguments about the risks of large language models that were critical of the direction of the field.

Environmental costs. The paper documented the substantial energy consumption of training large language models, connecting the computational costs of state-of-the-art NLP to environmental concerns. The training of the largest models required computing equivalent to the lifetime carbon footprint of multiple cars; the paper argued that this environmental cost was inadequately acknowledged in the field’s celebration of scale.

Training data quality. The paper argued that the large text corpora used to train language models were not merely large but were specifically problematic — they encoded the specific perspectives, biases, and prejudices of the populations who had generated the text (primarily English-speaking, primarily Western, primarily more privileged users of the internet), and that training language models on this data encoded these perspectives in the resulting systems.

The illusion of understanding. The paper’s most provocative argument was the “stochastic parrots” framing: the claim that large language models, whatever their apparent linguistic capability, were not genuinely understanding language but were performing sophisticated statistical operations on large text corpora — generating text that looked like language understanding without the underlying comprehension. The “stochastic parrots” metaphor — birds that produce statistically plausible sequences of words without understanding what they mean — was a specific challenge to the field’s tendency to describe impressive language model performance as evidence of genuine intelligence.

The concentration of power. The paper argued that the trend toward ever-larger language models was concentrating the frontier of NLP research in a small number of very well-resourced organisations, marginalising researchers without access to the computing required to build and study the largest models.

The paper was submitted to the Fairness, Accountability, and Transparency (FAccT) conference — a venue specifically oriented toward the social and ethical dimensions of AI. It was accepted.

---

The Firing: What Happened

The sequence of events between the paper’s acceptance and Gebru’s firing from Google has been extensively documented, and it is worth recounting carefully because the details matter.

Google’s research policy required internal review and approval of papers before submission to external venues. Gebru and her collaborators had followed the standard process. But in November 2020, Gebru received an email from a manager indicating that the paper had not received approval from Google’s internal review process and that it would need to be revised or withdrawn before publication.

The specific objections to the paper that were communicated to Gebru and her co-authors were focused on the paper’s claims about environmental costs, about bias in training data, and about the limitations of large language models. The framing of these objections, as Gebru and her co-authors understood them, was that the paper made claims that Google’s research management considered scientifically unsupported — claims that, not coincidentally, were critical of the direction of Google’s most commercially important AI research programme.

Gebru objected to the review process and to the specific objections, which she did not find scientifically credible. She sent an internal email to the Brain Women and Allies group — a group focused on supporting women in Google Brain — describing her experience with the paper review and her frustrations with Google’s handling of diversity and inclusion issues more broadly.

The email was intended as an internal communication. It was seen by Google’s senior leadership and became the proximate cause of what happened next: Google informed Gebru that she had the choice of withdrawing the paper or resigning.

Gebru did not resign. She received an email from Google’s management stating that her employment had been terminated. The firing was announced publicly on December 2, 2020, and immediately attracted enormous attention.

---

The Aftermath: A Firestorm and Its Consequences

The reaction to Gebru’s firing was swift, intense, and revealing.

Thousands of Google employees signed a petition demanding that Google explain its decision and reverse the firing. The petition was not just from junior employees; it included senior researchers and engineers who had significant standing at the company. The breadth and forcefulness of the internal dissent was unusual and showed how significant the firing had been for the company’s internal culture.

External reactions were even more intense. The AI ethics community, the broader research community, and technology journalists all responded with a combination of outrage and analysis. The specific details of the firing — the paper about the risks of large language models, the internal email, the management’s response — were interpreted, by most observers, as a company silencing a researcher who had made uncomfortable arguments about technology that was commercially central to the company.

Google’s official account — that the firing was about Gebru’s behaviour in sending the internal email, not about the content of the paper — was widely received as insufficient. The timing, the specific nature of the conflict, and the context of the paper’s criticism of Google’s own AI research made it difficult to separate the content dispute from the employment decision.

Margaret Mitchell, the other senior Google AI ethics researcher who had co-authored the paper, was fired in February 2021, two months after Gebru, after Google alleged that she had taken proprietary data in a personal security key. Mitchell and her supporters disputed the characterisation.

The loss of both Gebru and Mitchell in quick succession effectively ended the original version of Google’s AI ethics team. The team was reorganised, and the research programme that Gebru and Mitchell had been building was substantially changed. Several other team members left in the months following the firings.

---

The Significance: What the Episode Revealed

The Gebru firing was significant not primarily because of what happened to one researcher but because of what it revealed about the structural tensions in corporate AI ethics.

Google had established an AI ethics team — had made a specific institutional commitment to taking the ethical dimensions of AI seriously. This commitment was genuine in some respects: real researchers were hired, real research was done, real resources were allocated. But it was also conditional: the commitment extended to AI ethics research that did not challenge the commercial and strategic interests of the company, and it did not extend to research that did.

The Stochastic Parrots paper challenged Google’s interests directly. Large language models were central to Google’s AI strategy. Research arguing that large language models had significant costs and limitations that the field was inadequately acknowledging was research that Google had a commercial interest in not publishing.

The firing demonstrated, to the AI ethics research community, that corporate AI ethics teams were in a structurally compromised position. They existed within organisations whose interests were sometimes in conflict with the critical research that AI ethics required. The independence that critical research required was available only up to the point where the research threatened the organisation’s core interests; beyond that point, the interests of the organisation prevailed.

This was not a surprising observation in principle — anyone familiar with the history of corporate research had reason to expect this dynamic. But the Gebru episode made it undeniable and public in a way that significantly changed how the AI ethics community thought about the relationship between independent research and corporate research.

---

DAIR: Building the Institution the Field Needed

After leaving Google, Gebru did not go to another technology company or return to academia in the conventional sense. She founded the Distributed AI Research Institute — DAIR — as an independent, community-rooted AI research organisation.

The name was deliberate. “Distributed” opposed the centralisation that characterises frontier AI research, in which the most significant work happens in a small number of very large, very resource-rich organisations. DAIR was designed to be something different: a research organisation that was not beholden to any corporate interest, that was rooted in communities affected by AI, and that could pursue the kind of critical, independent research that the AI field needed but that corporate ethics teams could not provide.

The founding of DAIR reflected a specific theory of what was wrong with AI research and what would fix it. The problem was not just that corporate AI ethics teams were in structurally compromised positions — it was that the entire ecosystem of AI research was organised in ways that produced specific systematic biases. The research happened in organisations and communities that were not representative of the people most affected by AI. The questions the research addressed were shaped by the interests and concerns of those organisations and communities. The result was a field that was simultaneously very capable technically and very limited in its understanding of its own social implications.

DAIR’s research programme addressed this problem by starting from different questions — not “how do we make AI more accurate or more efficient?” but “what are the conditions in which AI systems cause harm, and how can those conditions be changed?” The research was explicitly oriented toward the communities most affected by AI deployment: Black communities affected by facial recognition and predictive policing, workers affected by algorithmic management, immigrants affected by AI-enabled border control.

DAIR also committed to a different organisational model. Unlike corporate research organisations, DAIR maintained no commercial interests that could compromise its independence. Unlike academic research organisations, DAIR was explicitly oriented toward practical impact — toward producing research that could be used by the communities it studied. Unlike advocacy organisations, DAIR maintained the methodological rigour of academic research.

---

The Research Programme: What DAIR Has Done

In the years since its founding, DAIR has produced research that reflects the organisation’s specific vision of what AI research should be.

Community-centred AI evaluation. DAIR has developed and promoted evaluation methodologies for AI systems that centre the experiences of affected communities — that assess AI system performance not just on benchmark datasets but on the specific tasks and in the specific contexts that matter to the people who are affected by those systems.

AI and immigration. DAIR has done research on the use of AI in immigration enforcement and border control — a domain where the people most affected by AI decisions have the least power to challenge them and where the consequences of error are severe. The research has documented specific AI systems used in asylum decisions, in visa processing, and in border surveillance, and has analysed the specific failure modes that affect people who are already in vulnerable situations.

The environmental dimension. Following the Stochastic Parrots paper’s attention to the environmental costs of large language models, DAIR has continued to develop methodology for measuring and communicating the environmental impact of AI systems. The research connects the environmental costs of AI to the specific communities that bear them — the communities near data centres and the mining sites that provide the minerals for AI hardware.

AI and labour. DAIR’s research on AI and labour has focused on the specific workers who are most affected by AI automation and AI-assisted management — the warehouse workers monitored by AI management systems, the content moderators whose mental health is damaged by their exposure to harmful content, the platform workers whose income is determined by AI allocation algorithms.

The common thread across DAIR’s research programme is the emphasis on the specific and the concrete — on actual people in actual situations, rather than abstract principles or hypothetical scenarios. This is in deliberate contrast to the approach that Gebru had found inadequate in the mainstream AI ethics conversation, which often engaged with the abstract questions of fairness and transparency without connecting them to the specific harms experienced by specific people.

---

The Stochastic Parrots Legacy: The Field Responds

The Stochastic Parrots paper was eventually published at FAccT 2021, after Gebru’s and Mitchell’s departures from Google. It became one of the most cited papers in AI ethics and one of the most cited papers in NLP.

The paper’s influence has been broad and continuing. Its “stochastic parrots” framing — the characterisation of large language models as systems that produce statistically plausible text without understanding — entered the mainstream AI discourse and became a reference point in ongoing debates about what language models are and what they can do.

The environmental argument in the paper contributed to increased attention to the energy consumption of AI training. The documentation of AI’s carbon footprint became a standard element of responsible AI practice, with major AI research organisations publishing estimates of the energy consumption of their training runs.

The training data argument contributed to the broader conversation about the quality, representativeness, and provenance of AI training data — a conversation that has only become more important as language models have been trained on ever larger and less carefully curated datasets.

The power concentration argument contributed to the debate about the distribution of AI research capabilities and the importance of investing in research outside the major AI companies. The founding of DAIR itself was one institutional response to this argument.

What the paper did not do — what no single paper could do — was change the direction of the NLP field. The trend toward larger and larger language models that the paper criticised has continued and accelerated. The environmental costs have grown. The training data issues have not been resolved. The power concentration has intensified.

The Stochastic Parrots paper was a warning. The field acknowledged the warning and continued on its course. Whether the warning will be vindicated by the development of AI — whether the costs it identified will prove, in the long run, to have been worth it — is still to be determined.

---

Gebru and Diversity: The Connection Between Who and What

One of the most important aspects of Gebru’s work is the explicit connection she draws between the diversity of who does AI research and the quality and direction of what AI research does.

This argument — that diversity is not just an ethical goal but a technical necessity — runs through Gebru’s work from the Gender Shades paper through the founding of DAIR. The specific argument is: if the people who build AI systems are not diverse, the systems they build will encode the specific perspectives, assumptions, and blind spots of the people who built them. A field dominated by men will build systems that work better for men. A field dominated by people from high-income countries will build systems that work better for people from high-income countries. A field dominated by people without experience of discrimination will build systems that discriminate.

The argument is not primarily moral — it is not primarily about the justice of including underrepresented groups. It is epistemic — it is about what knowledge is available to the people building AI systems and how the absence of diverse perspectives creates specific technical failures.

The Gender Shades work was the clearest demonstration of this argument: the demographic bias in facial recognition systems was partly a consequence of the demographic homogeneity of the training datasets and partly a consequence of the demographic homogeneity of the researchers who built and tested the systems. Researchers who had personal experience of the surveillance risks of inaccurate facial recognition — who were themselves members of communities most likely to be harmed by biased systems — would have asked different questions and applied different standards of evaluation.

This argument has been influential in the AI community, and the field has made some progress on diversity over the past decade. But the progress has been limited, and the field remains significantly less diverse than the broader population it affects.

---

The Political Dimension: Gebru as Advocate

Gebru has not limited herself to academic research and institution-building. She has also been a public advocate — using her platform to argue for specific policy changes and to hold specific institutions accountable for specific actions.

Her public advocacy has focused particularly on:

The use of AI in surveillance and law enforcement. She has been a consistent and vocal critic of facial recognition technology in law enforcement, arguing that the specific harms — the wrongful arrests, the discriminatory deployment, the chilling effects on free expression and assembly — are not acceptable even if the technology were accurate, and that the technology is not in fact accurate for the populations most likely to be surveilled.

The accountability of AI companies. She has argued consistently for greater accountability of AI companies — for requirements that companies publish evaluations of their systems’ social impacts, for legal liability for AI-caused harms, and for governance frameworks that give affected communities meaningful power over AI systems that affect them.

The treatment of AI workers. She has been an advocate for better conditions for the invisible workers who make AI systems work — the data labellers, the content moderators, the platform workers whose labour is essential to AI development and deployment but who are often paid poorly and protected inadequately.

Representation in AI research. She has been an advocate for the specific institutions — like DAIR — that support AI research from underrepresented communities, and for changes in the funding and evaluation structures that currently advantage the large, well-resourced organisations that dominate the frontier.

The advocacy has made her a central figure in the AI accountability conversation. It has also made her a target — for criticism from AI researchers who find her views too political, from technology companies whose practices she criticises, and from advocates who believe her positions do not go far enough.

---

The Critics: What Gebru Gets Wrong

Any honest account of Gebru’s contributions must acknowledge the specific criticisms her work has attracted — criticisms that come from both sympathetic and unsympathetic directions.

The technical critique. Some AI researchers have argued that the Stochastic Parrots paper overstates the limitations of large language models, that the “parrots” framing is philosophically loaded in ways that it does not adequately defend, and that the paper’s influence on the AI governance conversation has been disproportionate to its technical rigour. The paper engages primarily with the social and political dimensions of large language models rather than with the technical questions about their capabilities and limitations, and critics argue that this emphasis produces an incomplete analysis.

This critique has some merit. The Stochastic Parrots paper is not primarily a technical paper, and its characterisation of large language models’ capabilities is contested. The subsequent development of GPT-4 and its successors has demonstrated capabilities that the parrots framing does not fully capture — capabilities that go beyond statistical text generation in ways that deserve more nuanced characterisation.

The advocacy critique. Some researchers have argued that Gebru’s combination of technical research and political advocacy creates a confusion of roles — that advocacy positions should be distinguished from research findings, and that the conflation of the two can lead to research that is organised to support prior conclusions rather than to investigate questions with genuine openness.

This critique reflects a genuine tension in what Gebru does. Her research is explicitly oriented toward social impact and is informed by prior commitments about whose interests should be served. This is a defensible approach — all research is shaped by values, and being explicit about values is preferable to pretending to value-free objectivity. But it does create a specific risk of motivated reasoning that the research programme needs to actively resist.

The constructive critique. Some supporters of Gebru’s general approach have argued that the focus on criticism — on documenting what is wrong with AI systems and AI research — needs to be complemented by more constructive engagement with what AI systems and AI research should look like. The critique without the alternative can leave audiences without a clear direction for improvement.

This critique is being addressed, to some extent, by DAIR’s research programme, which includes constructive work on evaluation methodologies, on community-centred design, and on governance frameworks. But the constructive dimension is less developed and less visible than the critical dimension.

---

The Legacy: What Gebru Has Built

Gebru’s legacy in the AI field is still being written — she is in the middle of her career, DAIR is young, and the full impact of the Stochastic Parrots paper and the broader intellectual programme she has developed is still unfolding.

But several contributions are already clear.

The institutionalisation of independent AI ethics. The founding of DAIR demonstrated that independent, community-rooted AI research was viable — that research organisations outside the major AI companies could do rigorous, impactful work on the social dimensions of AI. DAIR has inspired similar organisations — the Algorithmic Justice League, the Data & Society Research Institute, the AI Now Institute — and together these organisations constitute a nascent civil society infrastructure for AI accountability.

The discourse on training data. The attention to the quality, representativeness, and provenance of AI training data that the Stochastic Parrots paper helped generate has become a standard element of responsible AI practice. The development of data sheets for datasets, of model cards for trained models, of nutrition labels for AI systems — all of these practices reflect the specific concerns about training data quality that Gebru’s work articulated.

The connection of diversity to quality. The argument that diverse AI research produces better AI — that the inclusion of underrepresented perspectives is not just an ethical goal but an epistemic necessity — has become more widely accepted in the field. Whether this acceptance translates into meaningful change in who does AI research remains to be seen.

The accountability norm. The expectation that AI companies should be accountable for the social impacts of their systems — that they should produce systematic evaluations of those impacts, disclose the limitations of their systems, and be subject to meaningful oversight — has become more widely accepted since the Gebru episode. The specific governance frameworks that have been developed — the EU AI Act’s transparency requirements, the US AI Safety Institute’s evaluation frameworks — reflect norms that Gebru’s work helped establish.

---

The Person: Who Timnit Gebru Is

Behind the public figure — the fired Google researcher, the DAIR founder, the AI accountability advocate — is a specific person whose character and history illuminate why she has done what she has done.

She is someone who has experienced, directly and personally, the specific kinds of harm that AI systems can cause and that unjust social structures produce. Her experience as an Ethiopian refugee, as a Black woman in a field dominated by white and Asian men, as a researcher who was fired by one of the world’s most powerful companies for doing her job — these are not abstract matters. They are her life.

This personal experience informs her research in ways that are both strengths and potential limitations. The strength: she asks questions that researchers without her experience do not think to ask, she sees risks that researchers without her experience do not see. The limitation: the risk that personal experience becomes a substitute for evidence, that the certainty of personal knowledge forecloses the uncertainty that good research requires.

She has navigated this tension, in her work and in her public advocacy, with varying degrees of success. At her best, she combines personal experience with rigorous methodology, using the former to identify the questions and the latter to answer them. At her worst, the conviction that she is right about important things — a conviction that is often justified — can make her less receptive to the kinds of evidence and argument that might qualify her conclusions.

What is not in doubt is the commitment. She has paid a real price — in her career, in her reputation within certain communities, in the personal costs of sustained public controversy — for the positions she holds. The commitment she has demonstrated is not the commitment of someone who is performing a role. It is the commitment of someone who believes that the things she is arguing for are genuinely important and who is willing to bear significant costs to argue for them.

In a field that often rewards caution, deference, and the avoidance of controversy, this willingness to argue for things that are inconvenient and to accept the costs of doing so is itself a significant contribution.

In a field that often rewards caution, deference, and the avoidance of controversy, the willingness to argue for things that are inconvenient and to accept the costs of doing so is itself a significant contribution. Gebru has paid a real price — in her career, in her reputation within certain communities, in the personal costs of sustained public controversy — for the positions she holds.

---

Further Reading

---

---