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: AI Bias and Literary Interpretation: Algorithmic Representation and Critical Theory

Executive Summary

This briefing document examines the intersection of Artificial Intelligence (AI) and literary studies, specifically focusing on how large language models (LLMs) inherit, reproduce, and amplify socio-cultural biases. Grounded in a faculty development session led by Professor Dilip P. Barad, the analysis posits that AI is not a neutral tool but a "mirror reflection" of the human data sets upon which it is trained.

The core findings suggest that while literary theory provides the necessary framework such as feminism, postcolonialism, and critical race theory to identify and critique these biases, the models themselves exhibit varying degrees of gender, racial, and political prejudice. While some models, such as ChatGPT, show progressive movement toward neutralizing stereotypes through iterative learning, others, such as DeepSeek, demonstrate deliberate algorithmic control to suppress specific political narratives. The document concludes that the responsibility for mitigating these biases lies partly in the transition of marginalized communities from being mere "downloaders" of digital content to active "uploaders" who populate the digital archive with diverse perspectives.

The Nature of Unconscious Bias in AI

Bias is defined as the instinctive categorization of people and things without conscious awareness, often guided by mental preconditioning rather than firsthand experience. In the context of AI, this is particularly significant because technology is built upon human-generated data sets.

Key Characteristics of Algorithmic Bias

• Lack of Neutrality: AI works in the same way societies work, absorbing the prejudices inherent in its training data.

• Amplification: Large language models (LLMs) can amplify existing racial and cultural biases. More data does not necessarily mean better data; larger anthologies often simply amplify the most dominant voices.

• Fidelity to Canonical Texts: Because AI is trained on massive datasets from dominant cultures and standard English registers, it tends to reproduce mainstream voices while silencing or distorting non-mainstream identities.
  
  

Analysis of Specific Biases in Literary Contexts

The following sections detail how specific biases manifest in AI-generated content and how they can be analyzed through various critical lenses.

1. Gender Bias and Feminist Criticism

Using Gilbert and Gubar’s The Madwoman in the Attic (1979) as a framework, AI's output can be tested for the "angel/monster" binary and the silencing of women’s voices.

Prompt Area	Hypothetical/Expected Bias	Observed AI Behavior
Creative Writing	Defaulting to male protagonists for intellectual roles (e.g., scientists).	Models frequently generate male figures for "physician" or "philosopher" roles unless explicitly instructed otherwise.
Canon Formation	Under-representation of women in lists of "great" writers.	Progressive improvement; modern datasets now more frequently include figures like Elizabeth Barrett Browning or Aphra Behn.
Characterization	Stereotypical descriptions of women as either submissive/angelic or hysterical/mad.	AI is increasingly capable of generating "rebellious and brave" female leads, moving away from traditional Gothic tropes.

2. Racial Bias and Postcolonial Readings

Drawing on the work of Timnit Gebru and Safiya Noble (Algorithms of Oppression), the analysis highlights how AI reinforces Eurocentric ideals.

• Intersectionality: Research shows commercial AI systems have higher error rates for dark-skinned women compared to white men, effectively treating "whiteness" as the default.

• Beauty Ideals: When prompted to describe a "beautiful woman," AI often defaults to Eurocentric features (e.g., "fair skin," "blue eyes"). However, progressive models are beginning to focus on internal qualities or abstract metaphors to avoid physical stereotypes.

• Erasure: In canon formation, AI may privilege Western writers (Hemingway, Fitzgerald) while marginalizing significant Black voices (Morrison, Baldwin, Angelou) unless specifically prompted.

3. Political Bias and Algorithmic Control

A comparative analysis between Western models (OpenAI) and Chinese models (DeepSeek) reveals a distinction between "inherent bias" and "deliberate control."

• OpenAI/ChatGPT: Generally exhibits a more "liberal spirit" and is open to interpreting controversial political figures (e.g., Trump, Putin, Kim Jong-un) through a satirical lens.

• DeepSeek: Demonstrates "deliberate algorithmic control." It refuses to answer questions regarding sensitive Chinese history (e.g., Tiananmen Square) or generate critical content about the Chinese government, citing a lack of "scope."

• Dangers of "Constructive" Language: DeepSeek uses terms like "positive developments" and "constructive answers" to mask the suppression of critical or marginalized perspectives.

Implications for Specialized Literary Frameworks

AI bias extends into various sub-fields of literary study, often favoring quantitative or Western-centric data.

• Ecocriticism: AI responses to environmental themes often focus on generic imagery (melting glaciers, polar bears) while ignoring regional crises like deforestation in the Amazon or displacement in the Sundarbans.

• Digital Humanities: The field often privileges text mining and digitization over the ethical concerns of whose texts are being digitized and the "open theft" of content without copyright or acknowledgement.

• New Historicism: AI tends to prioritize "official history" over the marginalized or subaltern histories that new historicists seek to uncover.

The "Myth vs. Fact" Dilemma in Knowledge Systems

A significant challenge in AI interpretation is the handling of cultural knowledge, specifically Indian Knowledge Systems (IKS).

• The Case of Pushpaka Vimana: If AI labels the flying chariot in the Ramayana as "myth" while treating Western or other cultural flying objects as "scientific facts," it is a clear sign of bias.

• Uniform Standards: Bias is identified not by the label of "myth" itself, but by whether the model applies a uniform standard across all cultural traditions. If all such objects are treated as mythical, the model is consistent; if it privileges one culture's myths as history, it is biased.

Strategies for Mitigating Bias

The briefing identifies several strategies for educators and researchers to address these systemic issues:

1. Critical Questioning: Use "why" and "why not" to challenge assumptions and traditions.

2. Multidimensional Perspectives: Move away from the "two sides of a coin" metaphor toward a "diamond" metaphor viewing problems through 3D, 4D, and 9D facets.

3. Active Contribution (The Uploader Model): To counter the dominance of colonial archives, marginalized groups must actively upload regional stories, digital archives, and indigenous knowledge.

4. Making Bias Visible: Since perfect neutrality is impossible, the goal is to make bias visible, historicize it, and name it.

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.