The Platform Question: Power, Accountability and Global South The Platform Question: Power, Accountability and Global South experiences, the backend repeats existing, often reductive, classifications. The consequences of this narrow perspective in classification have implications across the board, like the ability of a young woman on social media being able to report an image of herself being posted online without her consent. When annotators label what is considered “sexual content,” they may focus solely on nudity. But in contexts like India, intimacy or what is sexual does not need to be in the form of nudity, but can be of two people holding hands in public. These nuances are lost when the annotation logic is framed and designed through a Western lens and then becomes embedded into the functioning of AI systems. This is even more concerning when “some AI scientists have stated their desire to capture the world and to supersede other forms of knowing” (Crawford 2021). When selective stories are compressed into data, and that data is shaped by certain biases which produce AI outputs of “beautiful bodies” that users come to accept as reality, alternative stories and representations are effectively overwritten. Annotators are provided parameters and guidelines for labeling by developers. Studies have shown that annotators’ of socio-demographic backgrounds and lived experiences impact how they label. 126 The data is then labeled by low-wage workers from the Global South who determine what is “beautiful, and which “bodies” count. While the labour is global, the power to determine or influence is not, because those who monopolise resources monopolise imagination.” (Benjamin, 2019). Annotators are provided parameters and guidelines for labeling by developers. Studies have shown that annotators’ of socio-demographic backgrounds and lived experiences impact how they label. For example, in a study of large language models, researchers found that female annotators were more likely to label texts related to gender, religion, or cultural insensitivity as hateful than male annotators (Das et al, 2024). However, when speaking with AI/ML practitioners, annotators were generally treated as “an apparatus for achieving a representation of the world” (Kapania et al, 2023). The subjective nature of the annotator would result in noting disagreement and labeling ambiguous content on the basis of their interpretations, informed by their own assumptions and biases. In a model, any disagreements and inconsistencies by an annotator are averaged across others and against a The Missing Blocks of AI “golden data set”.4 The annotator is treated as an objective tool who would simply need to follow the parameters provided to them. Yet, like the example of ImageNet, where bias was attempted to be neutralised by crowdsourcing annotations from Amazon Turk workers’ (Deng et al, 2009), the error rate was 6% (Northcutt et al, 2021), demonstrating that subjectivity matters even if notations were averaged out. The Sociotechnical Machine If we begin with the understanding that neither data nor bodies are neutral, then we must also accept that AI is not an abstract force - it is built, shaped, and sustained by human choices. However, if social bias continues to be treated as a bug and not a feature, as argued by several feminist authors, the redressal mechanisms will often be piecemeal and limited to one model or narrow benchmarks or ‘fairness’ or ‘transparency’. Take, for example, Responsible AI. Rather than a universal framework, it is an overarching term that encompasses the intention of differing companies. Implementation varies as they are principles rather than accountability mechanisms. More importantly, existing within the same structures, assumptions, and biases of the architects of major AI systems, it cannot be transformative in changing the way that AI can be developed. Bias and unequal power structures are built into the system, so “how can one be “fair” in an unjust society?” (Hampton 2021). Therefore, there is a need to interrogate and go beyond frameworks like Responsible AI. To include the lived worlds of the structurally marginalised people, questions around consent and ownership of data, performance and exploitation of labour, and co-opting social justice or feminist ideologies have to be addressed. Feminist AI principles remind us that the design of technology is never separate from systems of power (Costanza-Chock, 2023). They ask us to center care, context, and co-creation; to ask not just what data is collected, but who gets to define it, label it, and benefit from it. This is where imagination becomes essential. As Ruha Benjamin (2024) urges, imagination is a tool of resistance, especially in the face of systems designed to exclude, flatten, or erase. And so reclaiming the right to imagine, especially for those whose stories have long been left out of technological futures, is a deeply political act. The Missing Blocks of AI To include the lived worlds of the structurally marginalised people, questions around consent and ownership of data, performance and exploitation of labour, and co-opting social justice or feminist ideologies have to be addressed. 127