Teaching in “Critical Perspectives on AI” at GCAS turned out to be an opportunity not just to think critically about dominant narratives of Big Tech, but a chance to reflect on teaching and what mode of learning still makes sense today.
This is the particularity of AI: the pervasive anthropomorphizing of it means that, more often than not, we end up reversing the initial question, ending up with a magnifying glass pointed right back at ourselves. One moment, we may be considering the ethical significance of a human-neuron-powered computer and asking, along with Safdari, “Are LLMs embodied?”. Next, we’re questioning human cognition and embodiment. As comparisons between humans and AI abound, we’re confronted with everything we still don’t, and may never, know about ourselves.
Similarly, considering the massive drive to integrate AI technology into teaching and universities makes it necessary to ask: what makes this space still make sense, and what kind of pedagogy is still irreplaceable? This is especially relevant as some higher education institutions seem to actively engineer their own demise: Harvard is reportedly selling a $699 course taught by AI clones of its faculty. In this case, the faculty members volunteered to have their likenesses and bodies of work used in this way. But what assumptions do they implicitly condone? Is it that the privilege of human interaction in the classroom is a premium product? Or that, with enough training data, we can predict a teacher’s next most probable word with a degree of certainty that makes their actual presence optional?
Teaching a course on AI critically, and teaching it at GCAS, offers particular insights. During the first two weeks of the course, we dispelled the myths of enchanted, immaterial “digital intelligence”; explored the material impacts of AI; and faced issues of coloniality and gender inequality reproduced by these technologies.
As we engaged with interdisciplinary perspectives from scholars representing diverse traditions, there seemed to be something many had in common. While they may not use these words, their work encourages resisting “algorithmic truth.” In a paper, “Automating Epistemology: How AI Reconfigures Truth, Authority, and Verification,” Donghee Shin notices an “epistemic shift” related to the increasing role AI plays in mediating “public knowledge and legitimacy.”
He says:
“The emergence of AI-powered truth claims represents a paradigmatic shift from traditional epistemologies—grounded in correspondence, coherence, and pragmatic theories of truth (Alston 1996; Blackburn 2005)—toward what this study terms algorithmic truth: an epistemic regime in which truth is constructed, classified, and operationalized through algorithmic procedures.”
Shin elaborates on the significance of this shift, explaining the distinction between verification and interpretation. Verification asks, “Did X occur?”, while interpretation explores contextual meaning-making: what does it imply, and how is it perceived? AI systems are proficient at the former while struggling with the latter, leading to “tensions between procedural certainty and epistemic depth.” That depth is exactly what critical perspectives attempt to cultivate and defend.
Shin explains how machine-learning models rely on “ground truth” datasets as benchmarks for training and evaluation. Yet, that ground truth is not an objective reality, but a socially constructed product shaped by human annotation, institutional labeling guidelines, and normative biases.
Further, the systems rely on “epistemic proxies,” scanning the surface level of data and “approximating credibility.” This reinforces an algorithmic rationality that favors detectability, speed, and scalability over deep contextual substance and nuance.
The process results in what the author calls epistemic centrality bias. Because search engines and verification algorithms weigh domain authority and institutional prominence heavily, they systematically favor mainstream, elite sources. This consolidates algorithmic metanarratives, marginalizing “informal, vernacular, or indigenous epistemologies”.
In simple terms, we could imagine such systems evaluating a “bell curve” of distributed perspectives and systematically assuming the middle of the dataset to be “ground truth,” leading to continuous homogenization of knowledge and what we could call “epistemic flattening”.
Perhaps it’s not incidental that, in “Critical Perspectives on AI,” I actively encouraged the opposite. That’s why my first question in class was not “What is AI?” but “What is it for whom?” While questioning the Big Tech promises of abundance powered by the “techno-capital machine,” we asked: will systemically invisibilized, underpaid data annotation workers in Dhaka be included in this future? And, importantly, do they have a say in whether this is the future they would choose?
Intuitively, rather than deliberately, the class was grounded in “epistemology from below”: an approach that asks what we can know when we take marginalized, embodied, local knowledge seriously.
In Data Feminism for AI, D’Ignazio and Klein state: “Feminists contend that outliers—in language as in life—tell us far more than data points at the center.”
With 78% of AI industry roles being held by men, we actively considered the position of women in the field who are plagued by conflicting expectations. They’re singled out for “lagging behind” when it comes to AI use, while suspected to be “incompetent” when they do use it.
We also discussed a report by a Kenyan data annotation worker turned advocate. In his research, James Mojez Oyange asks, “What does consent really mean when saying no could cost you your job?” This is an example of a question that people situated “below” dominant structures of power are uniquely equipped to answer. If an AI company CEO wants to keep their data private, a simple “no” suffices. Even more likely, an entire team can be tasked with ensuring digital security.
In Feminist AI: Critical Perspectives on Algorithms, Data, and Intelligent Machines, D’Ignazio and Klein comment:
“Rather than valorising the neutrality ideal, and trying to expunge all human traces from a data product because of their ‘bias’, feminist philosophers have offered alternative paths towards truth. Sandra Harding would posit a different kind of objectivity that strives for truth at the same time that it considers—and discloses—the standpoint of the designer.”
This prompts us to question not only algorithmic truth, but a kind of teaching that an AI-clone “professor” could offer. An assemblage and recombination of data can produce cost-effective course content, but not knowledge.
So why attend a GCAS seminar? It is not for “algorithmic truth,” which offers the illusion of “a view from nowhere”, and which today is always instantly within reach by prompting the algorithm. Perhaps it’s rather to engage with too-often-discarded outliers. GCAS gives us a space to experience knowledge as a process and a relationship, beyond binary fact-checking, and to engage with all of our diverse “views from somewhere”. The continued pressure on learning spaces to be optimized and automated means that, more than ever, acknowledging and celebrating our standpoints becomes even more of a strength.
This article reflects the author’s own views without aiming to represent GCAS’s position as a whole.