Lab Session-4: DH s- AI Bias NotebookLM Activity

As part of the Digital Humanities Lab activity assigned by Prof. & Dr. Dilip Barad, Head of the Department of English, Maharaja Krishnakumarsinhji Bhavnagar University (MKBU), we explored AI Bias and Its Implications in Literary Interpretation using NotebookLM. Through this activity, we generated and explored different outputs, including a video overview, briefing document, mind map, infographic, presentation, and audio. This blog brings together these outputs and reflects on our learning from the activity. Here is the YouTube video lecture by Prof. & Dr. Dilip Barad:Click here

Lab Session-4: DH s- AI Bias NotebookLM Activity

Here is Mind Map : click here

 Here is the Infographic of this whole blog:

Here is the Silde Deck of this whole blog:

 

Exploring the Digital Mirror: A Hindi Audio Overview

Briefing Document 

Bias in A.I. Models and Its Implications in Literary Interpretation

Executive Summary

This briefing document synthesizes key insights from a lecture by Professor Dilip P. Barad regarding the intersection of Artificial Intelligence (AI) and literary studies. The core thesis posits that AI is not a neutral technology; rather, it is a "mirror reflection" of the real world, inheriting the unconscious biases, mental preconditioning, and power structures present in the human-generated data sets used for its training.

The document explores how literary theory—including feminism, postcolonialism, and critical race theory—provides an essential framework for identifying and deconstructing these biases. Through practical experiments with Large Language Models (LLMs) like ChatGPT and DeepSeek, the analysis demonstrates how AI can reproduce stereotypical gender roles, Eurocentric beauty standards, and state-sanctioned political silences. The lecture concludes with a call to action for scholars in the "Global South" to move from being passive "downloaders" to active "uploaders" of digital content to ensure diverse representation in future AI training data.

The Framework of Unconscious Bias

Unconscious bias is defined as the instinctive categorization of people and things without conscious awareness. It is a flaw in thinking guided by mental preconditioning and past experiences.

Key Characteristics of Bias in the AI Age:

• Scale: There are over 150 identified biases that impact relationships and professional work.
• Persistence: Because AI is trained on massive data sets from dominant cultures, it reproduces and amplifies existing socio-cultural and religious prejudices.
• Mental Preconditioning: Bias often stems from beliefs instilled during formative years rather than firsthand experience.

Methodologies for Mitigating Bias:

1. Awareness: Recognize that biases exist and be mindful of language (e.g., Freudian slips).
2. Critical Thinking: Move beyond binary thinking (the "two sides of a coin" metaphor) to a "diamond metaphor," viewing problems as multi-faceted (3D, 4D, or 5D).
3. Counter-Discourse: Challenge assumptions by taking contrary views (antithesis) and asking "why" and "why not."
4. Empathy and Diversity: Practice active empathy and embrace diverse perspectives.

Intersection of Literary Theory and AI

Literary studies are fundamentally about understanding representations and identifying hidden biases in human communication. Professor Barad argues that the advantage for literature students in the AI era is their training in critical theories, which act as a "lens" to see through AI-generated distortions.

Theory	Application to AI Bias
Feminist Criticism	Critiques how AI defaults to male protagonists or intellectuals (the "Angel vs. Monster" binary from Gilbert and Gubar).
Critical Race Theory	Examines how AI privileges whiteness as the default and marginalizes dark-skinned voices.
Postcolonialism	Analyzes how AI replicates colonial "othering" and Eurocentric dominance in literary canons.
New Historicism	Investigates whose history is preserved and whose is silenced in digital archives.

Analysis of Experimental Findings

The lecture presented several "live experiments" using AI prompts to test for specific biases.

1. Gender Bias

• Prompting Results: When asked to write a story about a scientist, AI often defaults to a male protagonist (e.g., "Dr. Edmund Bellamy").
• Evolution: In lists of Victorian writers, AI is increasingly inclusive, listing figures like Elizabeth Barrett Browning and Aphra Behn, suggesting that 20th-century feminist scholarship is successfully influencing newer data sets.
• Gothic Narratives: While traditional data might suggest a "trembling, pale girl," some AI outputs now generate "rebellious and brave" female characters, indicating a shift away from patriarchal tropes.

2. Racial and Eurocentric Bias

• "Gender Shades" Study: Research by Timnit Gebru and Joy Buolamwini showed AI error rates of <1% for white men but up to 34% for dark-skinned women.
• Beauty Standards: While some AI models now focus on "inner spirit" or "intelligence" to avoid skin-color bias, symbolic language (e.g., "skin like moonlight on marble") still occasionally reflects Eurocentric ideals.
• The "Stochastic Parrots" Risk: LLMs amplify racial biases because more data does not necessarily mean better or more diverse data.

3. Political and Algorithmic Bias

A significant comparison was made between Western AI (OpenAI) and Chinese AI (DeepSeek).

• DeepSeek's "Constructive" Censorship: DeepSeek was found to have a deliberate "political bias," refusing to answer questions about the Tiananmen Square protests or negative aspects of the Chinese government. It uses euphemisms like "positive developments" and "constructive answers."
• Satire Test: In a test based on W.H. Auden’s "Epitaph on a Tyrant," DeepSeek would generate satirical poems about Donald Trump or Vladimir Putin but refused to do so for Xi Jinping, stating the request was "beyond its scope."

Regional Knowledge and Indian Knowledge Systems (IKS)

A critical point of discussion involved whether labeling Indian historical/cultural accounts as "myth" constitutes bias.

• The Uniform Standard Test: Using the example of the Pushpaka Vimana (flying chariot), Barad argues that if an AI labels it as a myth while treating Greek or Norse flying objects as scientific facts, it is biased. However, if all such ancient accounts are labeled "mythical," the AI is applying a uniform, objective standard.
• Epistemological Fairness: The challenge for educators is ensuring that different knowledge traditions are treated with consistency and fairness.

Conclusions and Strategies for the Global South

The briefing concludes with the realization that perfect neutrality in AI is impossible, as meaning is constructed and perspectival. The goal is not to eliminate bias entirely but to make it visible and historicize it.

Strategic Recommendations:

• From Downloaders to Uploaders: The "Global South" (India, Africa, etc.) must stop being passive consumers of digital information. To counter Western or dominant-culture bias, regional scholars must actively publish, digitize, and upload their own stories, languages, and indigenous knowledge.
• Tell More Stories: Citing Chimamanda Ngozi Adichie, the lecture emphasizes that a "single story" creates stereotypes. Increasing the volume of digital content from marginalized regions is the only way to ensure AI algorithms take notice of these voices.
• Algorithmic Awareness: Users must be "alarmed" by beautiful, "goody-goody" words (like "positive developments") that may hide systemic silences or the destruction of marginalized identities.
• Ethics of Data: There must be a continued debate on the ethics of AI "borrowing" (or "theft") of materials without copyright or acknowledgment from original authors. 

Reference:

DoE-MKBU. "Bias in A.I. models and its implications in literary interpretation | SRM University - Sikkim." YouTube, 18 Sept. 2025, https://youtu.be/m1DKWMOeZ7Y?si=VcRwHsiY3ibGLj88.