The cycle asks what it would mean to see the machine clearly—without the narcotic of hype or the paralysis of fear. Gebru is the figure in the gallery who insists on seeing it clearly in the most literal sense: with numbers, with receipts, with disaggregated data that makes invisible harms visible and refutes the comfortable averages that conceal them. Where other thinkers in the cycle offer philosophical critiques of AI’s epistemic or social effects, she offers measurement science, and the measurements are often more devastating than the philosophy.
Her central contribution to the cycle’s argument is the refusal of inevitability. The hype, the concentration of power, the exploited labor, the epistemic crisis of large language models—all of these are presented, in the dominant discourse, as the natural and necessary consequences of a technology that is arriving regardless of what anyone thinks about it. Gebru’s response to each is the same: someone chose this. The skewed benchmark, the indiscriminate training corpus, the underpaid content moderator, the AGI theology that licenses indifference to present harms—none of these are laws of nature. They are decisions made by people who could decide otherwise, and the people affected retain the right and the capacity to decide differently. This is not optimism; it is a structural claim about agency that the cycle needs more than any individual diagnosis of failure.
The invisible labor critique relocates the AI narrative from the realm of autonomous systems to the realm of human work, and the relocation is permanently disorienting. The ghost in the training data—the writer, the content moderator, the data labeler in Nairobi who taught the model what is acceptable—does not appear in any product announcement. Gebru’s insistence on naming them, counting them, and attending to the conditions of their labor is not a sentimental addendum to a technical story. It is a correction of the technical story: what we call artificial intelligence is, in large part, human intelligence captured, compressed, and stripped of its attribution, and to understand what the systems are, you have to see who built them and at what cost.
Her TESCREAL critique—the argument that the pursuit of artificial general intelligence is organized by a coherent ideological bundle with roots in early-twentieth-century eugenics—is the most ambitious and most contested of her contributions to the cycle. It claims that the grand eschatology of existential risk and posthuman transcendence serves a social function regardless of the sincerity of its adherents: it directs enormous resources toward speculative futures while licensing indifference to documented present harms, and it happens to align with the commercial and personal interests of the wealthy men who most energetically promote it. Whether or not the genealogical argument is accepted in full, it performs a service the cycle requires: it denaturalizes the pursuit of AGI, exposing it as a contestable ideological project rather than the inevitable destiny of a technology.
Timnit Gebru was born in Addis Ababa in the early 1980s to Eritrean parents, her father an electrical engineer with a doctorate and her mother an economist. She fled the Eritrean-Ethiopian war as a teenager, was denied a United States visa, spent time in Ireland, and eventually obtained political asylum in America, finishing high school in Massachusetts. She watched teachers try to keep her out of advanced courses despite her evident ability, and she watched the police arrest her Black friend rather than take a report about an assault against her. These were early lessons in how institutions decide whose account of reality is credible, and they migrated, eventually, into her science.
She earned bachelor’s, master’s, and doctoral degrees at Stanford, completing her doctorate in computer vision under Fei-Fei Li. Her dissertation used deep learning on Google Street View images to infer the demographic and political makeup of American neighborhoods from the cars parked in them—a demonstration of the inference capacity of the technology that her subsequent career would spend documenting and constraining. After working at Apple and Microsoft Research, she co-authored “Gender Shades” with Joy Buolamwini, co-founded Black in AI, and became co-lead of Google’s Ethical AI team alongside Margaret Mitchell.
The events of late 2020 are well-documented: management demanded retraction or removal of authors from the “Stochastic Parrots” paper; Gebru asked the obvious questions about who had reviewed it and on what grounds; Google treated her conditional statement as a resignation and removed her access while she was on vacation. Thousands of Google employees and academics signed letters of protest. Mitchell was herself fired months later. Gebru founded DAIR—the Distributed Artificial Intelligence Research Institute—a year to the day after her ouster. The timing was not incidental: it was a demonstration, in organizational form, that the conditions of knowledge production are choices and that different conditions yield different knowledge.
Intersectional disaggregation as method. Aggregate accuracy hides harm. A face-recognition system that is ninety-eight percent accurate overall may be sixty-five percent accurate for darker-skinned women—a failure that vanishes in the headline number and only appears when the data is broken down by the intersection of race and gender. The methodological contribution of “Gender Shades” was not merely the finding but the frame: the design of an evaluation is itself a political act, and a benchmark that averages across groups it should distinguish is structurally designed to not see certain people. This principle generalizes: every aggregate metric is a choice about whose failures register and whose disappear into the mean.
Stochastic parrots and the gap between fluency and understanding. A language model learns the statistical regularities of how words co-occur in its training data and generates plausible continuations on that basis. Nothing in this process involves reference to meaning, intention, or a model of the world. The fluency is real; the understanding is imputed by us—and we cannot help but impute it, because humans are meaning-making creatures who read coherent text as evidence of a mind. The “stochastic parrot” framing names the danger precisely: confident, fluent assertion that is grounded in pattern rather than truth, generating belief in capabilities the system does not have, at the cost of documented harm to real people concentrated in the margins.
Invisible labor and the hidden substrate. The invisible labor of data annotation, content moderation, and filtering makes AI possible and is systematically erased from its presentation. Workers in Kenya, India, and the Philippines are paid a fraction of what their counterparts in wealthy countries earn, often to review material traumatizing enough to cause lasting psychological harm, so that a chatbot can appear clean. This is not a side effect to be ameliorated with better wages alone; it is structural, a continuation of long histories of extraction in which value flows from the margins toward the center, from the powerless toward the powerful. The “intelligence” of the system is, in large part, human intelligence captured and stripped of attribution.
Against the concentration of power. The capacity to build frontier AI systems rests with a small number of enormously wealthy corporations that answer to their shareholders and have no obligation to the billions of people whose lives their systems affect. Gebru’s diagnosis extends beyond corporate greed to the general condition of a tiny, unrepresentative group making world-altering decisions on behalf of everyone else, insulated from accountability by the resources that grant them power. The concentration is self-reinforcing: those who control the most powerful systems use them to accumulate more data, capital, and influence, funding the next generation of even more powerful systems. Against this she sets the unglamorous work of building counter-power: independent research institutions, worker organizing, regulation with teeth.
The TESCREAL critique and the theology of AGI. With philosopher Emile Torres, Gebru argued that the pursuit of artificial general intelligence is organized by an interconnected bundle of beliefs—transhumanism, extropianism, singularitarianism, cosmism, rationalism, effective altruism, longtermism—with roots traceable to early-twentieth-century eugenics and a shared structure: the promise of transcendence to a posthuman future licenses enormous expenditure and indifference to present harm. Ideas have social functions independent of the intentions of those who hold them, and this bundle, whatever its adherents sincerely believe, functions to direct resources toward speculative futures while making documented suffering look like a rounding error against the astronomical stakes of the AGI race.