Exploring AI Bias and Literary Interpretation through Notebook LM

As part of our Digital Humanities laboratory activity, we used Notebook LM to analyse the YouTube lecture "Bias in A.I. Models and Its Implications in Literary Interpretation" Using Notebook LM, we generated a video overview, briefing document, mind map, infographic, presentation, and an audio. This blog presents all the outputs created during this activity.

Bias in AI Models and Implications for Literary Interpretation

Executive Summary

This briefing document examines the critical intersection of Artificial Intelligence (AI) and literary studies, focusing on the identification and mitigation of algorithmic bias. Based on the insights of Professor Dilip P. Barad, the document asserts that AI is not a neutral technology but a reflection of the sociocultural, religious, and political prejudices inherent in its training data. By applying literary theories such as feminism, postcolonialism, and critical race theory scholars can unmask the "unconscious biases" within AI outputs. Key findings indicate that while some models (e.g., Open AI) are evolving toward more progressive interpretations, others (e.g., Deep Seek) exhibit localized political censorship. Ultimately, the document argues that the goal is not to achieve impossible neutrality but to make biases visible and to challenge the "extractive epistemologies" of dominant cultures by actively contributing diverse digital content.

1. Defining Bias in the Human and Artificial Context

Bias is defined as an instinctive categorization of people and things without conscious awareness. It represents a "flow in thinking" guided by past experiences and, more dangerously, mental preconditioning where belief systems are often confused with knowledge systems.

The Nature of AI Bias

AI models are trained on massive, non-neutral datasets. Consequently, they reproduce the perspectives of:

Dominant Cultures: Mainstream voices often overshadow marginalized perspectives.
Standard Registers: Standard English dominates the linguistic data, potentially erasing regional or non-standard variations.
Human Interactions: AI functions as a "mirror reflection" of the real world; if a bias exists in society, it is likely to manifest in the virtual world.

The "Diamond" Metaphor for Critical Inquiry

To counter bias, researchers are encouraged to abandon the "two sides of a coin" metaphor in favour of seeing problems as a diamond with multiple facets (3D, 4D, 5D). This requires:

Critical attention to data and evidence.
Challenging assumptions and traditions.
Embracing diversity and practicing empathy.

2. Literary Theory as a Diagnostic Tool

Literary studies provide the necessary framework (hermeneutics and interpretive theories) to identify unconscious biases. Professor Barad highlights several key theoretical intersections:

Literary Theory	Application to AI Bias
Feminism	Critiquing the "Angel vs. Monster" binary and the default to male protagonists in scientific or intellectual narratives.
Postcolonialism	Identifying Eurocentric beauty ideals and the erasure of indigenous or "Global South" knowledge.
Critical Race Theory	Analysing how AI reinforces "whiteness as default" and marginalizes dark-skinned individuals in recognition and description.
New Historicism	Comparing official, "big data" histories with marginalized or subaltern narratives.
Ecocriticism	Examining how AI prioritizes generic Western environmental imagery (e.g., polar bears) over localized crises in the Global South.

3. Case Studies and Experimental Evidence

The document identifies specific instances where AI models demonstrate or have begun to overcome inherent biases.

Gender Bias: The "Madwoman in the Attic" Framework

Using Gilbert and Gubar’s 1979 framework, AI outputs were tested for patriarchal defaults.

Male as Default: In creative prompts for "a scientist," AI frequently defaults to male characters (e.g., "Dr. Edmund Bellamy").
Progressive Shifts: Newer iterations of models like Chat GPT show improvement by including female writers (e.g., Aphra Behn) in lists of restoration dramatists, indicating that 20th-century history datasets are influencing AI learning.

Racial and Algorithmic Bias

Reference is made to the work of Timnit Gebru and Safiya Noble (Algorithms of Oppression):

Whiteness as Default: Research shows AI systems had error rates of less than 1% for white men but up to 34% for dark-skinned women.
Stochastic Parrots: Large Language Models (LLMs) often amplify existing racial biases simply because they operate on a scale where "more data doesn't mean better data."

Political Bias: Open AI vs. Deep Seek

A significant distinction is made between the "liberal spirit" of American-based models and the controlled algorithms of others.

Deep Seek (China): Experiments revealed that Deep Seek would refuse to answer prompts regarding sensitive Chinese political history (e.g., Tiananmen Square) or generate satirical poetry about contemporary Chinese leadership, responding instead with: "That's beyond my current scope."
Open AI (USA): Conversely, Chat GPT was found to be more "progressive" and open to political satire across various world leaders (Trump, Putin, Kim Jong-un), though it faces criticism from the right wing for "wokeism."

4. The Challenge of Indian Knowledge Systems (IKS)

A specific point of tension exists regarding how AI handles non-Western myth and history.

The Myth vs. Fact Dilemma: When AI labels the Pushpaka Vimana (flying chariot) as a "myth" while potentially treating other cultural concepts differently, it raises questions of fairness.
Uniform Standards: Bias is confirmed only if AI treats Western myths as scientific facts while dismissing Indian narratives. If all such objects across civilizations are labeled "mythical," the model is applying a uniform standard rather than a cultural prejudice.

5. Strategic Recommendations for Educators and Scholars

To navigate the era of AI-driven interpretation, the following strategies are proposed:

Move from "Downloaders" to "Uploaders": The Global South must combat its "laziness" in digital contribution. Bias persists because colonial archives are digitized while indigenous knowledge is not. Scholars must actively upload regional stories, languages, and histories to digital spaces like Wikipedia and Project Gutenberg.
Counter-Prompting: Use "anti-capitalist" or "subaltern" prompts to test if the AI can provide a balanced observation.
Critical Awareness of Tools: Users must remain mindful of where a tool was designed (e.g., the "Global North") and what "deliberate controls" may be placed on its algorithm.
Visible Bias: Neutrality is impossible. The goal of critical theory is to make bias visible, historicized, and questioned so it does not become enforced as a "universal truth."

6. Concluding Insights

"Bias itself is not the problem. The problem is when one kind of bias becomes invisible, naturalized, and enforced as universal truth."

As AI models evolve, their capacity for "algorithmic consciousness" (understanding user needs and awareness of data gaps) may improve faster than human preconditioning. However, the responsibility remains with human scholars to provide the critical lens through which these outputs are evaluated, ensuring that "positive developments" and "constructive answers" do not mask the erasure of marginalized realities.

Here is Infographic