The Question of Diversity of Data in AI Development

The paper proposes a critique of universal intelligence (represented by what we call artificial intelligence or machine learning models) that presupposes a stable and context-free epistemology of AI.The author argues that, instead, we should consider epistemic uncertainty as a key element of AI's functioning.This epistemic uncertainty stems "from gaps in model knowledge or exposure to novel contexts" (Iman, 2025, p. 6).This aligns with Derrida's concept of meaning and textuality, which are defined as always context-dependent and never universal.If we claim that meaning is always context-dependent, we must propose an alternative to, for example, the datacentric approach in AI, which presupposes that the larger the dataset, the closer we get to objective truth.Iman argues that "reality, data are always contextually situated, collected, labelled, and interpreted within specific socio-technical frameworks.Thus, no dataset, no matter how vast, fully captures future contexts" (ibid., p. 7).The next step is to consider the nature of data and its colonial and geopolitical background.In addressing this issue, Iman calls for data diversification.However, we should also inquire into the origin and nature of the data used in model development.Is it necessary to consider data quality, and based on what criteria?This aspect is mentioned in the paper, but would benefit from more detailed attention.First, it would be useful to clarify what exactly is meant by "diversity."Does this refer to the sociopolitical concept of diversity (racial, gender, etc.), and what is the relationship between these dimensions and data diversity?For example, when we refer to data quality in large language models (LLMs), we often mean the diversity of languages used in training.Linguistic diversity encompasses idioms and dialects within a language, a variety of text types (scientific writing, literary genres), and the diversity of human authors (including their political perspectives), as well as different