The Platform Question: Power, Accountability and Global South The Platform Question: Power, Accountability and Global South into this process and how it creates the erasure of stories, knowledge systems, and representation caused by systemic biases, and how these absences are replicated and amplified in the digital world. Data, data everywhere but whose stories are told? Historically, the process of data collection, who gets counted, and what the counting signifies has always been deeply political. The “outcast, minoritised bodies, and subaltern” are often excluded or its inclusion minimised against the dominant colonial narrative. 124 Type in “beautiful bodies” on Google, ChatGPT, Gemini or Midjourney and you will see mostly homogenous, eurocentric, slim bodies. These outputs reflect a system built on data about humans, but only the dominant beauty standards that see whiteness, thinness, non-disabled, heteronormative as desirable. When identified, this bias is often presented as an anomaly or oversight. However, AI, often assumed to have sociotechnical blindness (Johnson and Verdicchio, 2017), is a system “shaping and shaped by what is social” (Çetin, 2021). And one of its foundational elements, training data2, is not just a technical input but is inherently social, created and shaped by experiences and narratives that are recorded, especially in large volumes. Crawford (2021) cites the example of National Institute of Standards and Technology (NIST), which consists of mug shots of arrested individuals from the United States, as a popular training dataset for image-related AI models. The context of when these images were taken or the lack of consent of the individuals (or their families) to be included in the database becomes irrelevant; individual experiences become reduced to datapoints in the form of an image which are to be extracted, circulated and repetitively used. Stories of lived experiences, therefore, get flattened into data. Moreover, historically, the process of data collection, who gets counted, and what the counting signifies has always been deeply political. The “outcast, minoritised bodies, and subaltern” are often excluded or its inclusion minimised against the dominant colonial (Çetin, 2021), patriarchal (D’Ignazio and Klein 2020) and white (Noble, 2018) narrative. For example, during colonial times in India, nomadic people were labeled as criminal tribes if they did not fit into certain parameters and were recorded in public databases as such (Kovacs, 2020). In more recent times, Ruha Benjamin (2019) highlights how crime data, a public record, reflects the skew of over-policing in Black and Brown communities. As a result, The Missing Blocks of AI foundational training datasets in machine learning (ML), often taken from such public records or scraped from the internet, become unneutral archives. Furthermore, in the Indian context, shared or surveilled phones of young women or the digitally excluded, due to the caste, socio-economic, or other barriers, leave minimal digital footprint; being absent in datasets used to train ML models altogether. These create models that reflect social biases, like in the case of Amazon’s hiring bias3 or predictive policing tools over-policing in communities of colour (Ruha Benjamin, Race After Technology). The bias can also be compounded, as demonstrated by Joy Buolamwini and Timnit Gebru (2018) analysis of facial recognition algorithms and datasets. Their findings showed that consistently, the highest error rate was among dark-skinned women, demonstrating both a gender and racial bias. When these already biased sets are applied in the Global South context, it becomes even more exclusionary, missing out the sociopolitical and cultural lived realities of marginalised communities. To therefore have “beautiful bodies” reflect queer, Brown, Dalit, disabled bodies in AI systems, additional modifiers and labels have to be included. Type these in another language, and the results may be sexualised or irrelevant images. This reinforces the notion that marginalised bodies are the “other”, deviations from the norm or the default representation, where one has to constantly assert and insert themselves into systems not inclusively designed for them. The Data Doesn’t Label Itself Bias is not limited to the data that is being used. While training data determines what AI systems learn, annotations give that data its meaning by shaping how it is interpreted through classification. The data is given labels, defining what is and what isn’t. This binarisation further reduces complex, nuanced realities to fit clean boxes of labeling. Consider gender. Facebook at one point had many choices of genders; however, on the backend, these identities were collapsed into the binary of male/female for Facebook advertisers (Beltrán & Ranganathan 2020). Therefore, classification systems, or the ways that people are sorted, are systems of power, determined and shaped by the decisions of those designing the AI (D’Ignazio and Klein; Wajcman and Young 2023). Regardless of bandaid frontend solutions that seem to include diverse The Missing Blocks of AI When these already biased sets are applied in the Global South context, it becomes even more exclusionary, missing out the sociopolitical and cultural lived realities of marginalised communities. 125