A white paper from The Noetic Feminism Institute. Impact Series. 2026

Executive summary

Artificial intelligence is the most powerful knowledge system being built in the twenty-first century. It shapes who gets hired, who gets insured, who gets paroled, who gets seen by a doctor, whose content gets amplified, and whose reality gets defined as statistically normal. It does this at a scale and speed that no previous institution of knowledge production has approached. And it is being built, overwhelmingly, on data generated by patriarchal societies, by teams that remain disproportionately male, within corporate structures that treat profit as the primary measure of intelligence.

This white paper argues that the feminist critique of artificial intelligence cannot stop at bias audits and diversity hiring. These interventions are necessary and insufficient. The deeper problem is epistemological: AI systems inherit and automate the same hierarchy of knowledge that feminist scholars have spent decades exposing in law, medicine, education, and science. They encode whose experience counts as data, whose language patterns define coherence, whose pain is legible to the system, and whose ways of knowing are classified as noise. When a large language model is trained on text produced overwhelmingly by men in positions of institutional authority, the resulting system does not merely reflect existing bias. It reproduces a worldview and calls it neutral.

The Noetic Feminism Institute situates this analysis at the intersection of feminist epistemology, critical technology studies, and the Institute’s founding commitment to recovering forms of knowledge that patriarchal systems have suppressed. We ask: what would artificial intelligence look like if it were built on feminist epistemological principles, on situated knowledge, embodied perception, relational accountability, and the honest admission that no system trained on patriarchal data can claim objectivity? The question is not rhetorical. It is a design problem, and feminist thought has been generating answers for decades.

1. The epistemological inheritance: what AI systems learn before they learn

“Feminist objectivity means quite simply situated knowledges.”

— Donna Haraway, Situated Knowledges (1988)

Every artificial intelligence system begins with data. The data is not raw material waiting to be processed. It is already a product of the societies that generated it: their assumptions about who matters, whose voice carries authority, whose experience constitutes a valid observation, and whose reality is too marginal to measure. Training data is not a mirror of the world. It is a selective archive, shaped by every hierarchy of power that determined who got to write, publish, record, and be counted.

Feminist epistemologists have made this argument about knowledge institutions for decades. Sandra Harding’s standpoint theory demonstrates that all knowledge is produced from a social position, and that positions of dominance produce characteristic blind spots precisely because they can afford not to notice what they exclude. Donna Haraway’s insistence on situated knowledges argues that objectivity is not the view from nowhere but the accountable view from somewhere, a claim that requires naming one’s position rather than pretending to transcend it. Patricia Hill Collins’s Black Feminist Thought shows how the intersection of race and gender produces epistemological insight that dominant frameworks systematically miss.

These insights, developed in relation to universities, legal systems, medical establishments, and scientific institutions, apply to AI systems with particular urgency. When a machine learning model is trained on data that reflects decades of gendered, racialized, and class-stratified knowledge production, it does not start from a clean slate. It starts from a position, the position of the data it was given. And that position, in most existing systems, is the position of patriarchal, Western, English-dominant, institutionally credentialed knowledge, presented as though it were universal.

The result is not a neutral tool. It is a knowledge system that has inherited the epistemological commitments of the societies that produced its training data, including the commitment to treating certain forms of knowledge as legitimate and others as irrational, emotional, subjective, or feminine.

2. The gendered architecture of AI: bias as structure, not accident

The evidence that AI systems reproduce gender bias is extensive and growing. Safiya Umoja Noble’s Algorithms of Oppression (2018) demonstrated how search engines reinforce racist and sexist stereotypes, not through intentional programming but through the logic of systems trained on data saturated with structural inequality. Cathy O’Neil’s Weapons of Math Destruction (2016) showed how algorithmic decision-making in hiring, lending, policing, and education systematically disadvantages women and people of color, often while claiming mathematical neutrality. Virginia Eubanks’s Automating Inequality (2018) documented how automated systems used in public services reproduce and deepen the surveillance and punishment of poor women. Ruha Benjamin’s Race After Technology (2019) named the phenomenon the New Jim Code: the use of apparently neutral technologies to reinforce existing racial and gender hierarchies. Kate Crawford’s Atlas of AI (2021) mapped the material infrastructure of artificial intelligence, showing how the industry depends on exploited labor, extracted resources, and concentrated corporate power.

These scholars have collectively demonstrated that AI bias is not a malfunction. It is a feature of systems built on biased data within biased institutions by teams whose composition reflects the very hierarchies the systems then reproduce. When Amazon’s hiring algorithm taught itself to downgrade resumes containing the word “women’s,” it was not making an error. It was faithfully reproducing the pattern its training data contained: that successful hires at the company had historically been men. The algorithm performed exactly as designed. The design was the problem.

The feminist analysis goes further than identifying individual biases and demanding corrections. It asks a structural question: why is the architecture of AI systems organized around the assumption that intelligence means pattern-matching on existing data? Whose definition of intelligence is this? And what forms of intelligence does this architecture exclude by design?

3. What counts as intelligence: the epistemic question AI has not asked

“The master’s tools will never dismantle the master’s house.”

— Audre Lorde, Sister Outsider (1984)

The dominant paradigm in AI development defines intelligence as the capacity to identify patterns in large datasets, generate statistically probable outputs, and optimize for measurable objectives. This definition is not neutral. It privileges a specific epistemological tradition: one that values quantification over qualitative understanding, prediction over perception, scalability over situated judgment, and outputs that can be evaluated by metrics over knowledge that resists measurement.

Feminist epistemology has spent decades questioning precisely this hierarchy. When Sandra Harding argues that knowledge produced from marginalized standpoints reveals truths invisible from dominant positions, she is making a claim about what counts as evidence. When Patricia Hill Collins insists that lived experience is a legitimate form of knowledge, she is challenging the assumption that detachment produces better understanding than engagement. When Audre Lorde writes that the master’s tools will never dismantle the master’s house, she is naming a structural constraint: systems designed within a particular logic cannot produce outcomes that logic was built to prevent.

Applied to AI, the implications are significant. If artificial intelligence is built exclusively on the epistemological assumptions of patriarchal, Western, rationalist knowledge traditions, then even a “debiased” version of that system will remain structurally limited. It will be a more polite version of the same narrow intelligence: capable of processing the kinds of knowledge that fit its architecture and incapable of recognizing the kinds that do not.

The forms of knowledge that AI systems currently cannot process are precisely those that feminist, Indigenous, contemplative, and noetic traditions have insisted are essential: embodied perception, relational awareness, emotional intelligence, intuitive recognition, ethical sensitivity, and the kind of knowing that arises through deep listening rather than data extraction. These are not soft supplements to real intelligence. They are dimensions of intelligence that patriarchal epistemology has dismissed because they resist the forms of control and measurement that patriarchal institutions depend on.

The question the Noetic Feminism Institute poses is not simply how to make AI fairer within its existing framework. It is whether the framework itself is adequate to the kind of intelligence human societies actually need.

4. The body in the machine: what AI cannot feel and why it matters

One of the most consequential absences in artificial intelligence is embodiment. AI systems process text, images, and numerical data. They do not inhabit bodies. They do not experience fatigue, pain, fear, desire, intuition, or the felt sense of being in a room with someone who is lying. This absence is treated, within the dominant AI paradigm, as a feature: intelligence purified of the body’s messiness, objectivity freed from the distortions of flesh.

Feminist philosophy has a name for this aspiration. It is the view from nowhere: the pretense that knowledge improves as it moves further from the body, from social location, from the particularities of lived experience. Haraway called it the god trick, the claim to see everything from no position at all. It is the epistemological stance that patriarchal science has treated as the gold standard for centuries, and it is built into the architecture of artificial intelligence at the foundational level.

The consequences are not abstract. In medicine, AI diagnostic systems trained on data that historically underrepresented women’s symptoms and women of color’s presentations reproduce those gaps at scale. Caroline Criado Perez documented in Invisible Women (2019) that women’s heart attack symptoms, women’s pain thresholds, and women’s drug reactions have been systematically understudied because the default research subject has been a 70-kilogram white man. When AI systems are trained on this data, they do not correct the bias. They automate it, at a speed and confidence level that makes it harder, not easier, to challenge.

In criminal justice, risk assessment algorithms used in sentencing and parole decisions encode historical patterns of racially and economically biased policing and prosecution. Eubanks showed that automated systems in welfare and child protective services subject poor women, disproportionately women of color, to levels of surveillance and punishment that wealthier families never encounter. The system does not see the structural conditions that produced the data. It sees the data and calls it probability.

What a body knows, a disembodied system cannot replicate: the gut recognition that something is wrong before the evidence assembles, the somatic awareness that a situation is unsafe, the felt knowledge that a patient’s chart does not capture what is actually happening in the room. These are not imprecisions to be eliminated from intelligence. They are forms of intelligence that the current AI paradigm is structurally incapable of including, and their absence has material consequences for the people whose lives are shaped by algorithmic decisions.

5. Language, power, and the training corpus

“If you can control the meaning of words, you can control the people who must use them.”

— Philip K. Dick

Large language models are trained on text. The text they are trained on is not a representative sample of human expression. It is a sample of what has been written down, published, digitized, and made available for scraping, which is to say it is a sample shaped by every form of privilege that determines who gets to write and be read. English dominates. Institutional and academic language dominates. Male authorship dominates historically. The voices of people who were denied literacy, publication, or digital access are absent, not because they had nothing to say but because the systems that produced the training data were never designed to include them.

Feminist linguists have long argued that language is not a neutral medium. It encodes power. It reflects whose experience has been treated as worth naming and whose has been left without vocabulary. The fact that “hysteria” derives from the Greek word for uterus is not a historical curiosity. It is a trace of the epistemic architecture that coded women’s distress as pathology and men’s authority as reason. When AI systems learn language from corpora saturated with these patterns, they learn the patterns, not as bias but as the statistical norm.

This has implications beyond offensive outputs. It shapes what AI systems treat as coherent, relevant, authoritative, and true. If the training data privileges analytical, detached, institutionally credentialed prose, then the system will reproduce that register as the default voice of intelligence. Relational language, emotional precision, narrative ways of knowing, the kinds of expression that feminist and Indigenous and contemplative traditions have cultivated for centuries, will be treated as deviations from the norm rather than as legitimate modes of thought.

The Noetic Feminism Institute notes that this is the same epistemic hierarchy that operates in universities, in medical consultations, in legal proceedings, and in newsrooms: the hierarchy that treats certain voices as authoritative and others as anecdotal. AI does not invent this hierarchy. It inherits it, scales it, and presents it as the output of mathematical neutrality. The math is real. The neutrality is not.

6. Surveillance, care, and the gendered economy of data

The data that feeds AI systems is not freely given. It is extracted. It is harvested from social media interactions, search histories, medical records, purchasing behavior, location tracking, and the countless small digital transactions that constitute contemporary life. The people who generate this data rarely control how it is used, who profits from it, or what decisions it informs.

Feminist scholars have noted that this extractive relationship mirrors older patriarchal and colonial patterns. Just as women’s unpaid care labor subsidizes economies while remaining invisible in GDP calculations, users’ data labor subsidizes the AI industry while remaining uncompensated and largely unacknowledged. The parallel is not metaphorical. It is structural. The same logic that treats women’s domestic work as a free resource, endlessly available and not worth measuring, treats users’ data as raw material, endlessly harvestable and not worth compensating.

The surveillance dimension is equally gendered. Eubanks demonstrated that automated decision-making systems are disproportionately deployed against poor women and women of color, in welfare administration, child protective services, healthcare rationing, and housing allocation. These systems do not empower their subjects. They watch them, score them, and make consequential decisions about their lives without meaningful transparency or recourse. The promise of AI efficiency, in these contexts, is a promise of more efficient control over populations that patriarchal and racial capitalism has always surveilled most aggressively.

Meanwhile, the AI industry itself reproduces internal gender inequalities. Women remain underrepresented in technical roles, in leadership, and on the research teams that determine what problems AI is built to solve. When Timnit Gebru, a leading AI ethics researcher, was pushed out of Google after co-authoring a paper on the risks of large language models, the incident illuminated a structural truth: the institutions building the most powerful knowledge systems in human history are not accountable to the communities most affected by those systems. Feminist analysis asks whose interests are served by this arrangement and whose safety depends on changing it.

7. Ghost work: the invisible labor that trains the machine

Behind every AI system is human labor that the industry prefers not to discuss. Large language models do not train themselves. Their training data must be collected, cleaned, labeled, and filtered by human workers, many of them women, many of them in the Global South, many of them paid poverty wages for work that is psychologically demanding, ethically fraught, and structurally invisible.

Mary L. Gray and Siddharth Suri documented this phenomenon in Ghost Work (2019), showing how the AI industry depends on an enormous, largely hidden workforce of contract laborers who annotate data, moderate content, and perform the countless micro-tasks that machine learning systems require. These workers are classified as independent contractors, denied benefits, and rendered invisible in the corporate narratives that present AI as an autonomous technology. The term “ghost work” is precise: the labor disappears into the machine, which then presents its outputs as though they emerged from pure computation.

Content moderators, many of them women in Kenya, the Philippines, and other countries where labor costs are low, spend their working hours reviewing and labeling violent, sexually explicit, and psychologically disturbing content so that AI systems can learn to filter it. The psychological toll is documented and severe. The compensation is minimal. The workers are invisible to the users who benefit from the system they maintain.

Feminist economists will recognize this pattern immediately. It is the same structure that has always undergirded patriarchal capitalism: essential labor performed by women and marginalized people, rendered invisible by the systems it supports, compensated at rates that reflect the social status of the workers rather than the value of the work. Silvia Federici’s analysis of how capitalism depends on the devaluation of reproductive labor applies with disturbing precision to the AI industry. The machine learns. The women who train it disappear.

A feminist analysis of AI that addresses only bias in outputs without examining the labor conditions that produce those outputs is incomplete. The epistemological critique and the labor critique are connected: both reveal a system that extracts value from people it refuses to see, and calls the result intelligence.

8. Feminist AI: what would it mean to build differently?

If the problem is epistemological, the solution cannot be limited to technical patches. Debiasing individual datasets, while necessary, does not address the structural question: what kind of intelligence are we building, and whose definition of knowledge does it serve? Feminist epistemology offers a framework for asking these questions with the depth they require.

A feminist approach to AI development would begin with accountability. Haraway’s situated knowledges demand that every knowledge system name its position: who built it, with what data, for what purpose, and with what limitations. Applied to AI, this means radical transparency about training data, model architecture, intended use, and known gaps. It means refusing the pretense that a system trained on English-language, Western, institutionally dominant text can speak for humanity.

Standpoint theory suggests that AI systems should be deliberately designed to amplify knowledge produced from marginalized positions, not as a corrective afterthought but as a foundational principle. If knowledge from the margins reveals truths that dominant perspectives cannot access, then an AI system that excludes marginalized voices is not merely unfair. It is epistemologically impoverished. It is, in the most precise sense, less intelligent than it could be.

A feminist AI ethics would also take seriously the forms of knowledge that current systems cannot process. Embodied perception, relational awareness, intuitive recognition, and ethical sensitivity are not luxuries. They are dimensions of intelligence that feminist, Indigenous, contemplative, and noetic traditions have cultivated with rigor and care. The question is whether AI development will continue to define intelligence as what machines can currently measure, or whether it will expand its definition to include what human consciousness, at its most perceptive, actually knows.

The Noetic Feminism Institute argues for the second. This does not mean demanding that machines develop intuition. It means insisting that the systems we build acknowledge the existence and value of forms of intelligence they cannot replicate, and that human decision-making processes are never fully delegated to systems incapable of perceiving what the body, the community, and the contemplative mind can perceive.

Concretely, this means designing AI systems with what we might call epistemic guardrails: built-in acknowledgments of the system’s perceptual limits. A medical AI that flags its own inability to account for patient self-report. A hiring algorithm that discloses the demographic composition of its training data. A risk-assessment tool that names, in every output, the forms of contextual knowledge it cannot access. These are not technical impossibilities. They are design choices that the current incentive structure of the AI industry does not reward, precisely because epistemic humility is less marketable than the promise of omniscient efficiency.

Feminist community-based AI design offers another pathway. If standpoint theory is correct that knowledge from marginalized positions reveals what dominant positions cannot see, then the communities most affected by AI systems should be involved in their design, not as test subjects but as knowledge-holders. Participatory design processes led by women, by communities of color, by disabled people, by those who have experienced the sharp end of algorithmic decision-making, would produce fundamentally different systems. They would produce systems that know what they do not know, because the people who built them have lived the consequences of being unseen.

9. The oracle and the algorithm: two models of knowing

The title of this paper names a tension. The algorithm processes data, identifies patterns, and generates outputs optimized for measurable objectives. It is powerful, scalable, and fast. It operates on the epistemological assumptions of the traditions that built it: that intelligence is pattern recognition, that objectivity means detachment, that better data produces better knowledge.

The oracle, in its oldest sense, is a different kind of intelligence. The oracle at Delphi did not process data. She listened. She entered an altered state of consciousness and spoke from a position that was embodied, situated, and irreducible to the information available. Her knowledge was relational. It depended on presence, on the specific question asked by the specific person asking it, on the conditions of the encounter itself. It could not be scaled or automated. It was, by every metric the algorithm values, inefficient.

Feminist epistemology does not need to choose between these models. It needs to insist that both exist, that both have value, and that a society that delegates all consequential decision-making to the algorithm while dismissing the oracle has not achieved superior intelligence. It has achieved a very fast, very confident form of partial knowledge, one that is structurally incapable of recognizing what it cannot see.

The oracle, in this framing, is not a mystical figure to be revived literally. She is a placeholder for every form of intelligence that the algorithmic paradigm excludes: the grandmother’s knowledge of her family’s medical history that the intake form does not capture, the social worker’s felt sense that a situation is dangerous in ways the risk score cannot measure, the patient who knows something is wrong before the lab results confirm it, the community that understands its own needs better than any model trained on national averages. These are forms of situated, embodied, relational knowing that feminist epistemology has always insisted are real. The question is whether the most powerful knowledge systems of the twenty-first century will be designed to honor them or to override them.

It is worth noting that the historical oracle was not anti-rational. She operated within a sophisticated institutional framework: the temple at Delphi had protocols, priestesses, interpreters, and political relationships. The oracle’s knowledge was taken seriously by heads of state. Her authority was institutional as well as spiritual. The modern dismissal of intuitive, embodied, and relational knowledge as the opposite of real intelligence is itself a historical product, a consequence of the Enlightenment’s decision to define reason narrowly and then to build institutions, including now AI systems, on that narrow definition. Feminist epistemology recovers what was lost in that narrowing: not the abandonment of reason but the recognition that reason alone is a partial intelligence claiming to be the whole.

10. Toward a feminist epistemology of technology

The Noetic Feminism Institute’s 2025 IMPACT paper, “Making Patriarchy Visible,” argued that patriarchy operates through institutional structure and epistemic hierarchy, determining who gets to know, whose perception counts, and whose experience is treated as evidence. The present paper extends that analysis into the technological domain. Artificial intelligence is the newest and most scalable institution of knowledge production human societies have built. If feminist analysis does not engage with it at the epistemological level, the epistemic hierarchies that feminists have spent decades exposing will be encoded into systems that operate faster, more confidently, and with less accountability than any institution before them.

This is not a prediction. It is already happening. Hiring algorithms that reproduce gender discrimination. Medical AI that underdiagnoses women. Predictive policing systems that intensify the surveillance of communities of color. Content moderation systems that suppress feminist, queer, and Indigenous voices while amplifying dominant narratives. Welfare automation that punishes poor women for the conditions poverty produces. Each of these systems presents itself as neutral. Each reproduces the epistemological commitments of the data it was trained on. Each makes consequential decisions about human lives without the forms of intelligence, embodied, relational, intuitive, that feminist traditions insist are essential to ethical judgment.

A feminist epistemology of technology would insist on several commitments. Transparency: every AI system should be required to name its data sources, its known limitations, and the populations whose experience it cannot represent. Accountability: the people affected by algorithmic decisions should have meaningful recourse, not after the harm is done but in the design process itself. Epistemic humility: AI systems should be built with the explicit acknowledgment that they represent one form of intelligence among many, and that the forms of knowledge they cannot process, embodied, intuitive, relational, spiritual, are not inferior. They are beyond its reach.

Most importantly, a feminist epistemology of technology would refuse the assumption that faster, larger, and more automated necessarily means more intelligent. Intelligence, as feminist and noetic traditions understand it, includes the capacity to perceive what dominant systems have rendered invisible, to sit with complexity that resists optimization, to attend to the particular when the system demands the aggregate, and to trust forms of knowing that cannot be scaled. These capacities are not obstacles to good decision-making. They are conditions of it.

The question this paper leaves open

Artificial intelligence will continue to grow in power and reach. The question is not whether it will shape human life but whose intelligence it will encode and whose it will erase.

Feminist epistemology offers something the current AI paradigm desperately needs: a tradition of asking whose knowledge counts, whose experience is treated as evidence, and whose ways of knowing have been systematically excluded from the institutions that define reality. These are not secondary concerns to be addressed after the technology is built. They are foundational design questions. Every choice about training data, model architecture, optimization metrics, and deployment context is an epistemological choice, a decision about what kind of intelligence the system will embody and what kind it will discard.

The algorithm is powerful. The oracle is necessary. A feminist future will not be built by choosing between them. It will be built by societies willing to ask, at every stage of technological development, what forms of intelligence are missing from the room, and whose lives depend on their inclusion.

The most sophisticated pattern-recognition system in existence remains partial knowledge if it cannot perceive what a woman’s body knows before the data confirms it, what a community understands about its own conditions, or what the silence in a room is actually saying. Feminist epistemology has always known this. The question is whether the builders of the most powerful knowledge systems in human history are willing to learn it.

References

Benjamin, R. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press.

Collins, P. H. (2000). Black Feminist Thought: Knowledge, Consciousness, and the Politics of Empowerment. Routledge.

Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press.

Criado Perez, C. (2019). Invisible Women: Data Bias in a World Designed for Men. Abrams Press.

Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press.

Gray, M. L., and Suri, S. (2019). Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt.

Haraway, D. (1988). Situated Knowledges: The Science Question in Feminism and the Privilege of Partial Perspective. Feminist Studies, 14(3), 575–599.

Harding, S. (1991). Whose Science? Whose Knowledge? Thinking from Women’s Lives. Cornell University Press.

Lorde, A. (1984). Sister Outsider: Essays and Speeches. Crossing Press.

Noble, S. U. (2018). Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press.

O’Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown Books.

Share this: