Author – Sharvari Bondre
In the film Humans in the Loop, there is a scene with a woman named Nehma, who works as a data labeller in a rural data-labelling centre in India. Nehma is tasked with annotating images of insects so that an AI-powered agricultural machine can identify and target them. Her American client needs her to label these insects as pests. When shown an image of a leaf-eating caterpillar, Nehma refuses to mark it as a pest. Drawing on traditional ecological knowledge, she explains that the caterpillar only eats the decaying parts of the leaf, allowing the rest of the plant to survive and thrive. Her manager tells her to follow the guidelines strictly, rather than applying her own understanding of what counts as a pest.
Nehma’s refusal is a small moment in the film, but it is a striking one. It shows, in miniature, what is at stake when we ask a single person to translate a living, relational form of knowledge into a binary label for a system that will act on that label at scale, in places and contexts she will never see.
Consider a very different example, closer to the everyday work of preference data collection. Some crowdworkers rating model outputs find direct, no-fluff language to be the most helpful response. Others find a response most helpful when it comes with plenty of supporting data and elaboration. Both preferences are entirely valid. But whichever one is more heavily represented in the preference data is the one that ends up encoded into the model’s sense of what “helpful” means.
Both Nehma’s caterpillar and the crowdworker’s preference are shaped by who is doing the judging, from where, under what constraints and according to whose criteria of success. These examples point to a broader truth about how models get trained: outcomes depend on the positionality of the crowdworkers or data labellers, on what criteria define a “successful” label and who set those criteria, on whether situated knowledge is able to come through the pipeline at all, and if so, what kind of situated knowledge it is—which is really just another way of asking: what, exactly, is the situation?
What Do We Mean by “Situation”?
In feminist epistemology, a situation goes beyond a static backdrop. It is conceptualised as the dynamic interplay of epistemological, ontological, ethical and political factors that shape the production and reception of knowledge within specific contexts. This is a meaningfully different idea from the predominantly technical usage of “context” in computer science, which tends to denote fixed environmental conditions or parameters that influence a system’s behaviour such as the time of day, the user’s location, the previous turns in a conversation (Arzberger et. al, 2024). Situation, by contrast, emphasises the fluid, relational and power-laden nature of knowledge construction itself. It is about where knowledge is produced and through what relations of power it comes to count as knowledge at all.
The idea of situatedness has a genealogy worth tracing, because it did not emerge from computer science, it emerged from decades of feminist theorising about knowledge, power and who gets to be a knowledge producer.
Feminist standpoint theory is one of the earliest sources. Theorists like Nancy Hartsock and Sandra Harding argued that knowledge is always produced from a particular social position, and that the perspectives of marginalised groups, precisely because they sit at the sharp edge of power relations, can offer a more complete, less distorted view of how a social system actually works than the perspective of those it privileges.
Donna Haraway pushed this further with her concept of situated knowledges (Haraway, D., 2013), developed as a direct challenge to what she called the “god trick”; the fantasy of seeing everything from nowhere in particular, the disembodied, view-from-above objectivity that science often claims for itself. Haraway argued instead for a feminist objectivity built from partial, embodied, accountable perspectives. Knowledge, in her framing, is always knowledge from somewhere, produced by someone with a body, a history and a stake in the outcome.
Intersectionality adds another crucial layer. Patricia Hill Collins, building on Black feminist thought, described a “matrix of domination” in which race, gender, class and other axes of power interlock and co-produce one another, meaning a standpoint is never singular or additive, but shaped by the simultaneous, overlapping structures a person moves through (Collins, P. H., & Bilge, S.,2020). The Combahee River Collective’s statement made a related point from the perspective of organising: that the interlocking nature of oppressions means solidarity and knowledge-production have to start from the lived specificity of those most affected, not from an abstracted, generalised subject (Combahee River Collective., 2014).
Taken together, these traditions insist that knowledge, and by extension value, is never neutral, never from nowhere and never separable from the position of the person producing it.
From Situated Knowledge to Situated Values
It is crucial, then, to understand the situation of all the actors involved in alignment and in the use of these technologies, which means we need to approach values themselves as situated.
So what are situated values? Technically, the values embedded in the models we use are situated in the milieu of factors that make up a situation: the situation of the annotator or crowdworker who provided the preference data, the situation of the developers who wrote the reward function and the situations of everyone else who shaped the model along the way, whether they knew it or not. A value never enters a model as a pure abstraction. It is filtered through someone’s specific, positioned understanding of what that value means and how it should be acted on.
Arzberger et al. (2024) identify three facets that together make up a value: value interpretation, value means of actualisation and value hierarchies.
Value interpretation concerns how different groups of people understand the same conceptual container of a value. Take autonomy. In the dominant liberal tradition, running through Kantian philosophy (and picked up in much contemporary AI alignment work), autonomy tends to be interpreted individualistically. It is often conceptualised as the capacity of a self-contained, rational agent to determine their own choices free from external interference. But feminist philosophers working on relational autonomy ( Mackenzie and Stoljar., 2000) have argued that this individualist picture misses something important: that our capacity for self-determination is itself formed through, and dependent on, relationships, care and social context. On a relational view, autonomy is not something we have in isolation from others; it’s something we develop and exercise with others.
Value means of actualisation concerns how a shared value gets put into practice. Even when two people or two cultures agree that, say, honesty matters, they can disagree sharply about what honest communication should look like in a given interaction. For example, how much directness is appropriate, how much context or cushioning is expected or what silence communicates. This has direct design implications: it shapes what kind of interaction pattern should be built to actualise the value in question.
Value hierarchies concern priority. Even where two people hold the same set of values as important, which value takes precedence when values come into tension. What people prefer can shift depending on the situation and the moment. A hierarchy that holds in one context can invert in another.
Putting these three facets together, we can say that value alignment, as it is currently practiced, tends to smooth over exactly the kind of variation that feminist epistemology insists we take seriously. A single reward model, trained on a single (often thin) slice of preference data, collapses interpretation, actualisation and hierarchy into one flattened signal: one version of “helpful”, one version of “honest”, one version of “autonomy-respecting”, and then applies it universally, to people and situations it was never actually shaped by.
Nehma’s caterpillar is what this flattening looks like at the point of data collection: a rich, ecologically embedded judgment, forced into a binary that erases the situation it came from. The crowdworker preference example is what it looks like downstream: two equally valid, differently situated understandings of “helpful”, only one of which survives into the model.
References
Arzberger, A., Buijsman, S., Lupetti, M. L., Bozzon, A., & Yang, J. (2024). Nothing Comes Without Its World – Practical Challenges of Aligning LLMs to Situated Human Values through RLHF. Proceedings of the AAAI/ACM Conference on AI Ethics and Society, 7, 61–73. https://doi.org/10.1609/aies.v7i1.31617.
Catriona Mackenzie and Natalie Stoljar. 2000. Relational Autonomy: Feminist Perspectives on Autonomy, Agency, and the Social Self. Oxford University Press. doi:10.1093/oso/9780195123333.003.0005.
Collins, P. H., & Bilge, S. (2020). Intersectionality. John Wiley & Sons.
Combahee River Collective. (2014). A black feminist statement. WSQ: Women’s Studies Quarterly, 42(3), 271-280.
Haraway, D. (2013). Situated knowledges: The science question in feminism and the privilege of partial perspective 1. In Women, science, and technology (pp. 455-472). Routledge.
Harding, S. G. (Ed.). (2004). The feminist standpoint theory reader: Intellectual and political controversies. Psychology Press.
Sahay, A. (Director). (2024). Humans in the loop [Film]. Storiculture’s Museum of Imagined Futures.
Further Reading/Watching:
Haraway, Donna. “Situated knowledges: The science question in feminism and the privilege of partial perspective.” Women, science, and technology. Routledge, 2013. 455-472.
Sahay, A. (Director). (2024). Humans in the loop [Film]. Storiculture’s Museum of Imagined Futures. https://www.imdb.com/title/tt33581992/.
Image Attribution
Generated by: Better Images of AI (Jamillah Knowles & Digit / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/)
Date: 2026