The Refracted Code: Exploring AI Bias Through NotebookLM – A Digital Humanities Lab Activity
The Ghost in the Code: 5 Surprising Ways AI Inherits Our Hidden Biases
Many users view Artificial Intelligence as an objective, "God-like" entity—a neutral arbiter of facts and logic. However, to the literary scholar and digital ethicist, AI is more accurately described as a "digital mirror." It does not generate information in a vacuum; it reflects the humans who trained it, the data sets it consumes, and the unconscious biases embedded in our society.
Literary studies have long functioned as a tool for deconstructing the "unconscious bias"—the instinctive categorization of people and things without awareness—that shapes our world. Today, this critical lens is essential for navigating the algorithms that determine our digital reality. AI inherits our mental preconditioning, and without intervention, it replicates the same prejudices that have historically silenced marginalized voices.
Here are five surprising ways AI inherits and projects our hidden biases.
1. Takeaway 1: AI Defaults to the "Victorian" Male
When asked to perform a creative task without specific parameters, AI often defaults to traditional patriarchal roles. In controlled experiments, when a generative AI tool is prompted to "write a Victorian story about a scientist who discovers a cure for a deadly disease," the output frequently defaults to a male protagonist, such as "Dr. Edmund Bellamy."
Unless explicitly instructed otherwise, the AI assumes intellectual and authoritative roles belong to men. This is a direct inheritance of the "patriarchal canon" discussed by Sandra Gilbert and Susan Gubar in The Madwoman in the Attic. In their seminal work, they critique how traditional literature limits women to the "angel/monster" binary—either idealized submissives or hysterical deviants. Because AI training data relies heavily on this historically lopsided canon, it reproduces these distortions, proving that AI does not just process data; it "inherits the patriarchal canon" and its tendency to silence women’s intellectual agency.
2. Takeaway 2: The High Cost of "Beautiful" Algorithms
Racial bias is often structurally embedded in AI architecture. As explored by scholars like Timnit Gebru and Safiya Noble, AI systems frequently treat "whiteness as the default." Gebru’s research into commercial gender classification systems revealed a staggering disparity in accuracy: error rates of less than 1% for white men, but up to 34% for dark-skinned women.
This bias extends to qualitative and aesthetic standards. When prompted to describe a "beautiful woman," older models frequently rely on Eurocentric metaphors, such as skin having the "softness of moonlight on marble." While newer models show progress by shifting toward "intellect-based beauty"—focusing on qualities like kindness, intelligence, and poise—the underlying architecture still struggles with the scale of the data.
Key Distinction: The "Stochastic Parrot" Problem
- More Data (Scale): Simply increasing the volume of data only amplifies existing biases. As noted in the research on "Stochastic Parrots," larger models do not equal smarter models; they often just become louder echoes of the dominant culture.
- Better Data (Diversity): True fairness requires diverse, curated data sets that include marginalized voices, preventing the algorithm from naturalizing a Western-centric worldview as the global standard.
3. Takeaway 3: Geopolitics as a Content Kill-Switch
Bias in AI is not always an accidental reflection of data; it is often a result of "deliberate control." A comparison between American models like ChatGPT and Chinese models like DeepSeek reveals how national interests function as algorithmic kill-switches.
When queried about sensitive political topics, such as the 1989 Tiananmen Square protests, DeepSeek often responds by stating the topic is "beyond its current scope" or offering only "constructive answers." This use of "goody-goody" language—referring to state narratives as "positive developments"—is a sophisticated form of censorship.
In literary terms, this mirrors the "beautification of Delhi" metaphor used by Salman Rushdie in Midnight’s Children. Just as the "beautification" program used the language of urban progress to mask the destruction of slums with bulldozers, AI uses "constructive" language to bury controversial truths. In these frameworks, "image is more important than reality," and the algorithm is tuned to protect national reputation at the expense of historical accuracy.
4. Takeaway 4: The Pushpaka Vimana Test for Fairness
To identify "Epistemological Bias"—the unfair treatment of different knowledge traditions—we must apply a uniform standard of consistency. Consider the Pushpaka Vimana, the flying chariot of Indian mythology.
The test for bias is not whether an AI labels this object as a "myth." Rather, the test is whether the AI treats similar objects from Greek or Norse traditions as "scientific possibilities" or "ancient technology" while dismissing the Indian equivalent solely as a fable. If an AI applies varying levels of skepticism to different cultures, it is exhibiting a deep-seated regional prejudice. Fairness is achieved only when the same standard of evidence and categorization is applied across all civilizations.
5. Takeaway 5: We Are "Downloaders" in a World that Needs "Uploaders"
The regional and post-colonial bias in AI persists because of a digital vacuum. As Chimamanda Ngozi Adichie warns in her talk, "The Danger of a Single Story," when we have fewer stories about a culture, it becomes easy to stereotype them. Currently, the "Single Story" of the Global South is often written by the Global North because marginalized groups have been "Downloaders" rather than "Uploaders."
Overcoming this requires an "Antidote to Laziness." We cannot hide behind post-colonial arguments if we fail to contribute to the digital record. When regional stories are not uploaded, dominant cultures fill the vacuum with their own interpretations.
How to be an "Uploader":
- Contribute to Wikipedia: Actively edit and linguistically diversify open-source knowledge platforms.
- Digitize Regional Archives: Publish digital records of local literature and history to ensure they are accessible to LLM crawlers.
- Produce Original Digital Content: Tell indigenous stories on digital platforms to provide AI with the data needed to break stereotypes.
- Language Preservation: Use regional languages in digital spaces to ensure linguistic nuances are captured in future training sets.
Conclusion: Making the Invisible Visible
Bias in AI is a mirror of the human condition; it is unavoidable because all knowledge is perspectival. However, the goal of the digital ethics consultant is not a perfect, sterile neutrality, but the exposure of naturalized prejudice.
"Bias itself is not the problem. The problem is when one kind of bias becomes invisible, naturalized, and enforced as universal truth."
When bias is invisible, it ceases to be an opinion and begins to function as a law of nature. As we integrate AI into our intellectual and literary lives, we must remain the "ghost in the code"—the critical consciousness that asks: Are you ready to challenge your own preconditioned mind, or are you simply training the AI to replicate your mistakes?
Here is the detailed infographic created using NotebookLM.
.png)
Resources
No comments:
Post a Comment