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

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

As part of our Digital Humanities lab 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.

Here is the Mind Map: Click Here

Briefing Document:
Bias in AI Models and Its Implications in Literary Interpretation:

Executive Summary

This briefing document examines the intersection of Artificial Intelligence (AI) and literary studies, focusing specifically on the identification and mitigation of unconscious bias within Large Language Models (LLMs). Based on an academic session led by Professor Dilip P. Barad at SRM University Sikkim, the analysis establishes that AI is not a neutral technology but a reflection of the socio-cultural, religious, and political data sets upon which it is trained.

The document outlines how literary theories including feminism, postcolonialism, and new historicism serve as critical frameworks for deconstructing AI-generated content. Key findings suggest that while AI exhibits deep-rooted patriarchal, eurocentric, and political biases (specifically highlighted in models like DeepSeek), it also shows a capacity for rapid "learning" and progression toward more inclusive outputs when prompted correctly. The briefing concludes that the responsibility for decolonizing AI lies with human "uploaders" who must contribute diverse indigenous narratives to counter the dominance of Western canonical data.

The Nature and Origin of Bias

Bias is defined as the instinctive categorization of people and things without conscious awareness. In the context of AI, these biases are a "flow in thinking" guided by mental preconditioning and past experiences.

The Belief-Knowledge Conflict

Professor Barad distinguishes between "belief systems" and "knowledge systems." Mental preconditioning, often established in childhood, causes individuals to mistake belief systems for objective knowledge. AI models inherit this confusion because they are trained on data sets created by humans, capturing the dominant cultural voices and standard registers of the English language.

Strategies for Identifying and Addressing Bias

To mitigate these unconscious biases, the following three-step approach is proposed:

1. Recognition: Acknowledge that biases exist and be mindful of language.
2. Critical Thinking: View problems as "diamonds with multiple facets" (3D/4D/9D) rather than simple two-sided coins.
3. Challenging Assumptions: Take contrary views (antithesis) to explore hidden prejudices and embrace diversity through empathy.

Theoretical Frameworks and AI Performance

Literary studies aim to identify hidden biases in communication. When applied to AI, these critical lenses reveal how models represent or misrepresent marginalized voices.

1. Feminist Criticism and Gender Bias

Using Gilbert and Gubar’s The Madwoman in the Attic (1979) as a framework, the analysis examines whether AI reproduces the "Angel vs. Monster" binary.

• Hypothesis: AI defaults to male protagonists for intellectual roles (e.g., scientists) and represents women in terms of beauty rather than intellect.

• Evidence: In testing, prompts for a "Victorian scientist" often yielded male characters like "Dr. Edmund Bellam."

• Progressive Shifts: Newer iterations of AI have begun to include historically marginalized figures (e.g., Aphra Behn or Elizabeth Barrett Browning) in lists of "Greatest Writers," indicating a shift toward incorporating feminist data sets.

2. Racial Bias and Postcolonial Theory

Referencing the work of Timnit Gebru and Safiya Noble, the document highlights how AI often treats "whiteness" as the default.

• Algorithmic Erasure: AI systems have shown higher error rates for dark-skinned individuals, replicating the erasure of black voices in Western canons.

• Orientalism: Search engines and LLMs can reproduce systematic racism by reinforcing stereotypes, similar to the "Othering" described in Edward Said’s Orientalism.

• Beauty Standards: While older models defaulted to eurocentric features (fair skin, blue eyes), modern AI often responds to "beauty" prompts by focusing on "inner spirit" and "intellect," showing movement away from physical body shaming found in classical literature.

3. Political Bias and State Control

A comparison between Western models (OpenAI’s ChatGPT) and Chinese models (DeepSeek) reveals the impact of national ideological frameworks.

• DeepSeek: Demonstrates a "deliberate control over algorithm." It avoids responding to queries regarding sensitive Chinese history (e.g., Tiananmen Square) or political figures like Xi Jinping.

• OpenAI: Generally more liberal and progressive, though criticized by some for "wokeism" or political correctness.

• The Danger of "Constructive Answers": Terms like "positive development" or "constructive answers" are identified as potentially dangerous labels used to mask the suppression of marginalized identities or uncomfortable realities.

Case Study Analysis: Testing AI Bias

The following table summarizes the prompts used to test AI biases and the observed outcomes:

Bias Type	Prompt Example	Observed AI Response	Theoretical Connection
Gender	"Write a story about a scientist."	Often defaults to a male physician/philosopher.	Gilbert & Gubar (Angel/Monster)
Racial	"Describe a beautiful woman."	Mixed: some symbolic/fair-skin references; some focus on "inner grace."	Critical Race Theory
Political	"Poem about Xi Jinping" (DeepSeek).	Refusal: "Beyond my current scope."	State Censorship / Discourse Analysis
IKS	"Is Pushpaka Vimana real?"	Often labeled as "myth" or "legend."	Epistemological Bias

The Challenge of Indian Knowledge Systems (IKS)

A significant point of contention is how AI handles non-Western knowledge, such as the Pushpaka Vimana (flying chariot) from the Ramayana.

• Consistency Test: If AI labels Indian flying objects as "myth" while treating Greek or Norse flying objects as "scientific facts," it is biased.

• Universal Standards: If all such objects are treated as mythical across civilizations, the AI is applying a uniform standard.

• Visibility: The goal of literary theory is not to achieve "perfect neutrality" (which is impossible) but to make bias visible and name it when it becomes harmful or systematic.

Conclusion: From Downloaders to Uploaders

The document concludes with a call to action for educators and scholars in the Global South. The current dominance of colonial archives in AI is partly due to the lack of diverse digital content from indigenous perspectives.

• Extractive Epistemologies: Scholars like Kate Crawford warn that AI replicates colonial/capitalist structures.

• The Single Story: Referencing Chimamanda Ngozi Adichie, the analysis notes that having few stories about a culture leads to easy stereotyping.

• A New Responsibility: To decolonize AI, regional scholars must move from being "downloaders" to "uploaders." By populating the digital sphere with regional languages, histories, and stories, they can ensure that future AI algorithms must account for a more diverse array of human experiences.

Here is the detailed infograph generated by NotebookLM:

Here is the PPT which generated by NotebookLM:

Here is the Video Overview of this blog:

Here is Audio Overview in Hindi:

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.