Who speaks when the machine speaks?

A reflective brief on artificial intelligence, power, and the people the social sector exists to serve — written for practitioners, executives, and anyone who suspects that "efficiency" is doing too much work in the room.

Intelligence has always been a contested title. For most of recorded history, the question of who possessed it — and who, by structural arrangement, did not — was a political instrument before it was a scientific one. Women were not unintelligent because they were proven to be; they were unintelligent because the institutions that defined the term had reasons not to extend it. The same logic excluded the colonised, the enslaved, the disabled, the poor. The history of intelligence as a concept is, before anything else, a history of authorisation: who is permitted to know, to speak, to be cited, to be believed.

It is worth holding this history in mind as the social sector enters its first serious encounter with what is now called artificial intelligence. The phrase carries enormous unexamined weight. It assumes that the thing being built is intelligent in some uncomplicated sense, that its outputs are knowledge, that the relationship between user and system is one of consultation rather than transaction. None of these assumptions survives close inspection. What we are calling intelligence is, materially, a statistical compression of inherited language — language produced inside unequal societies, scraped without consent, processed by underpaid annotators in Nairobi and Manila, and returned to us in the cadence of authority.

The social sector is being asked to adopt this technology at a speed that should give any serious institution pause. Stanford's Institute for Human-Centered AI reports that approximately half of surveyed nonprofits were already deploying AI tools by 2024, while three in four believed they should be deploying more. The same study identified bias — not cost, not capability, not infrastructure — as the principal barrier to adoption. The sector is therefore in a remarkable position: accelerating toward a technology whose primary acknowledged risk is the harm it inflicts on the populations the sector exists to protect. This is not a paradox to be resolved through better implementation. It is a governance failure dressed in the vocabulary of progress.

This essay argues that a feminist analysis is not an ornamental supplement to AI governance in the social sector but its operational core. Feminism, as a tradition of inquiry, has spent a century interrogating the words neutral, objective, efficient, and natural — the precise terms now being used to introduce algorithmic decision-making into family violence services, settlement programs, child protection systems, and public housing. The tradition is not borrowed; it is exactly fitted to the question.

I. The neutrality story, and why it never held

The first thing a feminist reading of artificial intelligence does is refuse the premise that the technology is neutral. This refusal is older than ChatGPT. It is older than the modern internet. Donna Haraway wrote in 1988 that there is no view from nowhere, only views from somewhere — a position she called situated knowledge. Every gaze, including the machine's, comes from a body, a budget, a history. To pretend otherwise is itself a political act.

What contemporary AI does, and does superbly, is launder situated knowledge into apparent objectivity. Safiya Umoja Noble, in Algorithms of Oppression, traced how search engines made racism appear neutral by encoding it in rankings. Ruha Benjamin, in Race After Technology, named this phenomenon the New Jim Code — the use of new technologies to reinscribe old hierarchies under the banner of progress. Joy Buolamwini's landmark 2018 Gender Shades study showed that commercial facial recognition systems failed to recognise dark-skinned women at error rates up to 34.7%, while recognising light-skinned men at near-perfect accuracy. The systems were not broken. They were working exactly as their training data had taught them.

"The most dangerous technologies are the ones that look clean." Ruha Benjamin, Race After Technology (2019)

UNESCO's 2024 study of large language models, replicated across multiple frontier systems, confirmed that the bias problem has not been engineered out — it has been laundered into fluency. In one open model, women were described in domestic roles four times more often than men. Men were given more diverse, higher-status occupations. Women were given homes to clean and children to raise. The model did not invent these patterns. It absorbed them, then handed them back to us at scale, with the air of statistical authority.

Kate Crawford, in Atlas of AI, calls these systems registers of power: not just tools that reflect inequality, but infrastructures that produce it — through the rare-earth mining that powers their hardware, the underpaid annotation labour that trains them, the data extracted without consent, and the institutional decisions they then automate. Abeba Birhane, an Ethiopian cognitive scientist whose audits of LAION and other foundation datasets have rewritten the field, has shown that the datasets behind the most celebrated generative models contain images and captions of staggering racism, misogyny, and child exploitation. We are training the future on the worst of the past.

The contradiction is itself the diagnosis. The sector is reaching for an instrument whose central acknowledged risk — bias against the populations it serves — is precisely the property that should disqualify it from deployment without rigorous governance. The deployment continues regardless, because budgets are constrained, because donors expect visible innovation, and because a generation of consultants has become exceptionally fluent in selling automation as a form of care. The market has produced a vocabulary in which adopting the tool is moral and questioning it is reactionary. This vocabulary is not accidental. It is what allows the contradiction to persist.

II. What gets repeated, gets believed

The most consequential finding in the UNESCO research is not the magnitude of any single output but the principle of repetition. Generative AI does not need to be wrong dramatically; it only needs to be wrong consistently. When language models repeatedly cast women in domestic settings, men in executive ones, certain accents as suspect, certain names as risk, the bias does not arrive as an error to be flagged. It arrives as a default. It arrives, more precisely, as a tone — the ambient register against which all subsequent claims are measured.

This matters acutely in the social sector because language is the medium through which the work travels. Risk assessments become evidence. Case notes become eligibility decisions. Grant applications become funded programs. Communications become the institutional voice that reaches communities. If the instruments generating that language are quietly tilted, the institutions built atop them tilt with them — invisibly, but cumulatively. Consider the practitioner drafting a risk assessment with AI assistance: the system that nudges, almost imperceptibly, toward describing a Latin American mother as volatile when it would have called a white Anglo-Australian mother stressed; the proposal that, when generated, defaults to passive constructions about migrant women — they were brought, they were assisted, they were resettled — rather than constructions that grant agency. Las palabras que nos eligen también nos forman. The words that choose us also shape us.

The mechanism here is older than artificial intelligence. Catherine D'Ignazio and Lauren Klein, in Data Feminism, observe that "data is never raw" — it is always cooked by someone, in some kitchen, for some meal. The proposition extends, with devastating accuracy, to generated text. Someone wrote the recipe. The question that ought to govern every adoption decision in the social sector is whether they were cooking for us, or for an audience we have never met.

III. The labour that nobody costed

Productivity decks have a habit of describing AI adoption as a subtraction — fewer hours spent on email, fewer hours spent on summaries, fewer hours spent on first drafts. The arithmetic is incomplete. Labour does not vanish when a tool is introduced; it migrates. In the social sector, the migration follows a familiar pattern. Time savings accrue at the top of the organisation, where senior staff pass through generated outputs at speed. The repair work — checking, correcting, softening, recontextualising, culturally translating, refusing — accrues at the bottom, where frontline workers absorb the gap between what the machine produced and what the situation actually required.

This is the same redistribution feminist political economy has been documenting for half a century, in domestic work, in care work, in clerical work: the labour that holds a system together by repairing its mistakes is rarely the labour that gets counted. Silvia Federici theorised it as reproductive labour — the unwaged work that makes waged work possible. Arlie Hochschild called it the second shift. Tressie McMillan Cottom, writing more recently on the political economy of inclusion, names the mechanism by which marginal workers are absorbed into systems that extract more than they return: predatory inclusion. The names differ; the structure is constant. Efficiency at one level of the organisation is purchased by invisibility at another.

The pattern repeats wherever the technology lands. A family violence service adopts a tool that summarises client intakes; the summaries are 80% accurate. The number sounds high until one notices that the 20% inaccuracies are not randomly distributed. They concentrate around clients with non-Anglo names, accented English, complex migration histories, and trauma narratives that do not follow a Western linear arc. The practitioner now spends her saved time correcting the summary so that it does not misrepresent her client to the next worker who reads it. She is performing two jobs. Only one is paid. Only one is recognised. Only one appears in the productivity report.

Sasha Costanza-Chock, in Design Justice, formulates the question that ought to govern every procurement decision in the sector: "Whose labour is the system designed to save, and whose labour is it designed to extract?" Until that question is on every contract, every business case, every implementation plan, the social sector is not adopting AI ethically. It is adopting it on the backs of the workforce it already underpays, and on behalf of the communities it claims to serve.

Whose labour is the system designed to save, and whose labour is it designed to extract? Sasha Costanza-Chock, Design Justice (2020)

IV. The Australian frame, and why it sharpens everything

Australia has chosen, for the present, to govern artificial intelligence by extending existing frameworks rather than constructing new ones. The choice is sometimes described as regulatory pragmatism. It is, more accurately, a deferral — the consequence of which is that organisations in the social sector are now expected to invent their own guardrails inside vendor relationships they did not design and cannot meaningfully renegotiate. The legislative vacuum is being filled, by default, with private terms of service.

The shape of this transition, as it touches Australian women, is increasingly legible in the data. Good Things Australia reported in 2025 that one in three Australian women struggle to distinguish AI-generated content from human-made content. The number rises to 39% among mothers, 49% among grandmothers, 40% among women with disability, and 35% among culturally and linguistically diverse women. Deloitte's 2025 TMT report, cited by the Australian Marketing Institute, found that 50% of Australian women were using generative AI at work, compared with 70% of men. The disparity is not merely one of access. It is a disparity of confidence, trust, and time — three resources the gender pay gap has been quietly eroding for decades.

For organisations operating in Victorian family violence and migration services, the implication is not abstract. The communities served sit precisely at the intersections where the divide is widest. The Victorian Women's Trust has been articulating these concerns for some time, drawing on the work of Tania Farha at Safe & Equal and Dr Jessica Lake at Melbourne Law School on the misuse of generative AI in domestic and family violence contexts — including the deployment of deepfakes as instruments of coercive control, and the use of impersonation accounts to circumvent intervention orders.

The argument is not that artificial intelligence should be excluded from this work. The argument is that it cannot be deployed as if its harms were evenly distributed. They are not. They concentrate, with mathematical reliability, on the populations the sector already struggles to keep safe.

V. The questions before the procurement

The single most useful contribution a feminist analysis makes to AI governance in the social sector is a revised set of questions for the procurement table. Standard procurement frameworks ask whether a tool works, whether it scales, whether it saves money. A feminist framework adds another register, and refuses to skip it.

1. What problem is this tool solving, and for whom?

If the underlying answer is insufficient staffing, the question beneath the question is whether automation or funding is the appropriate intervention. AI is increasingly being used to absorb the consequences of underinvestment on a permanent basis. Sectors should be honest, internally and externally, when this is what is occurring.

2. Whose knowledge shaped this system, and whose is missing?

Foundation models are trained predominantly on English-language internet text — a corpus whose demographic composition is well-documented and whose absences are not random. The communities most absent from the training data are, with painful regularity, the same communities the social sector exists to serve. This is not a marginal gap to be patched. It is the constitutive shape of the technology.

3. Who carries the risk if the model is wrong?

In a marketing department, a hallucinated output is an embarrassment. In a triage system, it is a person. The risk surface of the social sector is human, and the threshold of acceptable error must be calibrated accordingly. A model that is good enough for a technology company's internal workflows is not, by transitive logic, good enough for ours.

4. Does this redistribute power downward, or pull it upward?

Tools that grant frontline workers better information, more time, and more authority to challenge institutional decisions are categorically different from tools that aggregate worker activity into management dashboards and increase surveillance over those same workers. Both are sold under the same heading. Only the first is liberatory. The procurement question is not is this AI? but which kind?

5. What would refusal look like?

The most under-used word in the AI conversation is no. A feminist organisation requires the institutional capacity to refuse a tool — not as an act of contrarianism, but because refusal is itself a governance instrument. Without the structural option of no, every yes is, in some measure, coerced.

VI. What feminists are already building

It would be intellectually dishonest to mount a critique without acknowledging the construction work already underway. The feminist AI field is not a complaint department; it is a building site, and the structures going up in it are worth examining seriously.

Chayn, founded by Hera Hussain, develops survivor-centred digital infrastructure across 195 countries, including assisted letter-writing systems for victims of intimate-image abuse seeking the removal of non-consensual content. The A+ Alliance, led jointly by Women at the Table and Code for Africa, has incubated SafeHer; SOF+IA, an Argentinian system for flagging gender-based political violence online; and AymurAI, which structures judicial data on gender violence cases for use by court administrators. Caroline Sinders's Feminist Data Set is a multi-year participatory project examining what it means to construct a dataset by consent rather than by extraction. Pollicy's Afro-Feminist AI principles, developed across the African continent, articulate the operational difference between symbolic inclusion and material participation.

None of these projects claims to have resolved the contradictions of the field. All are working through the same governance, labour, and consent questions the wider sector now faces. What they share is a refusal to treat artificial intelligence as a finished product to be received from a vendor, and a sustained commitment to building tools that remain accountable to the people they affect. The shift is not from AI for women to a feminised version of the same architecture. The shift is from AI as product to AI as accountable infrastructure.

VII. A practical agenda, for the people in the meeting

This essay is reflective, not prescriptive, but readers — practitioners, executives, board members, feminist workers — will reasonably expect to leave with operational instruments. What follows is a short discipline: five commitments offered in the order an organisation might attempt them. They are insufficient on their own. They are the beginning of a longer practice.

One. Begin every AI conversation with mission, not tools. If a use case cannot be defended in the language of organisational purpose without resorting to efficiency as a proxy virtue, it should be paused. Efficiency is not a value; it is a measurement, and measurements without values are instruments of drift.

Two. Map risk by community, not by department. The relevant question is not which team uses this tool but which clients does it touch, and who among them is most likely to absorb its errors? Risk assessment must be built around the second.

Three. Treat human review as a design requirement, not a compliance gesture. A practitioner without time to disagree with the machine is not reviewing it. They are ratifying it. Either resource the review properly, or stop claiming it as a safeguard.

Four. Make invisible labour visible. If an AI tool generates output that someone must repair, recognise that someone in job descriptions, workloads, and remuneration. If the productivity case cannot survive that transparency, the productivity case was never real.

Five. Include affected communities in governance, not in feedback forms. There is a structural difference between consulting people about a system and granting them authority over it. The first is a public relations exercise. The second is a feminist commitment.

VIII. Closing — on speed, and what it costs

It is necessary to name, in closing, the conviction that haunts every conversation about artificial intelligence in this sector and that few participants in those conversations will state plainly. Speed is not progress. Adoption is not transformation. Efficiency, considered alone, is not a value but a measurement, and the political question is always the same: efficient at what, for whom, at whose expense, and over what time horizon?

The social sector exists, structurally, because certain forms of human need cannot be automated, scaled, or optimised without becoming something other than what they were. Care is one of these forms. Trust is another. Witness is a third. The relationships that sustain this work — between practitioners and communities, between institutions and the people they have historically failed, between workers and the cultural literacies they accumulate over decades — are not in the training data. They cannot be retrieved by prompt. They must be built, and they must be defended.

The feminist tradition offers something more useful, in this moment, than warning. It offers a method: a refusal of the false binary between innovation and values; a discipline of asking who benefits, who bears the cost, whose labour is counted, and whose voice carries when an automated system speaks with institutional authority. The method is older than computing. It will outlast the present technological enthusiasm. It applies, with surgical precision, to the question now in front of every social sector leader.

The trajectory of the sector over the next decade will be determined less by what the technology becomes than by what its institutions choose to govern. Artificial intelligence will not become equitable through good intention. It will become equitable, if it does at all, through governance — through the unromantic, slow, expensive work of building accountability infrastructures that survive procurement pressure, vendor capture, and the seductions of speed. That work is a feminist undertaking, whether or not it is named one.

There is a phrase that travels across feminist movements in different forms and different languages: lo que se construye sin nosotras, se construye contra nosotras. What is built without us is built against us. The sentence predates artificial intelligence by a century. It will outlast the present iteration. The remaining question — and it is genuinely open — is whether the social sector remembers it in time, or remembers it after the fact, in the language of regret.