AI and the Bias Feedback Loop

Deepfake technology and AI-generated content are not just blurring the lines between reality and fabrication—they’re also exposing and amplifying our inherent biases. As these systems increasingly mirror and magnify societal prejudices, questions about their impact on truth and fairness have taken center stage.

December 2017 saw a surge on Reddit, where some users began sharing pornographic videos featuring celebrities’ faces. Although these videos looked real, they were digitally created using algorithms that swapped one person’s face onto another’s body—a phenomenon now known as “deepfakes.” Today, AI image and video generation has advanced to the point where users can simply describe a scene and receive realistic photos or minute-long videos, or modify existing media to depict virtually any person or event. This expanded definition of deepfakes now covers all AI-generated or altered content. The rapid development of these technologies has sparked concerns among both the public and experts, with warnings of an impending “infocalypse” and a potential “collapse of reality,” as noted by Immerwahr.

In his piece “Your Lying Eyes,” Immerwahr takes a closer look at deepfakes and their societal implications, challenging the common fear that these technologies could lead to widespread deception. He points out that despite the spread of deepfakes, there has yet to be a case where one misled people in any significant way. In his view, deepfakes tend to entertain rather than deceive, revealing a complexity beyond simple trickery.

Looking back, Immerwahr reminds us that society has long been skilled at spotting forgeries. Early photographers routinely edited images—adding clouds or rearranging elements—and even during the Civil War, photographer Alexander Gardner moved bodies to create dramatic effects. These historical examples show that while manipulated images are not new, society has generally managed to detect them without catastrophic fallout.

In an intriguing twist, Immerwahr uses Arthur Conan Doyle’s belief in staged fairy photographs as an example of how manipulated media often reinforces what we already believe rather than tricking us. Doyle, best known for creating Sherlock Holmes, accepted these photos as genuine thanks to his longstanding fascination with the supernatural—a fascination partly fueled by personal loss during World War I. Even simple fakes can gain traction when they confirm our preexisting ideas.

According to Immerwahr, the real issue with fakes isn’t what they hide, but what they reveal about our own biases. The concern isn’t so much the increasing sophistication of fake content, but our tendency to embrace information that confirms our worldview—a habit that can eventually lead to harmful biases. In extreme cases, this can reinforce damaging stereotypes, such as outdated and sexist rumors about historical figures, or even facilitate the harassment of women.

In her article “Why Facts Don’t Change Our Minds,” Kolbert digs into why we so readily accept information that confirms our beliefs. Citing psychological studies, she explains that once impressions form, they tend to stick. For example, Stanford researchers found that students continued to believe in the accuracy of performance feedback—even after it was revealed to be entirely fabricated—highlighting a deep-rooted flaw in human cognition.

Kolbert further notes that this stubborn resistance to changing our views is rooted in our evolution. As cognitive scientists Hugo Mercier and Dan Sperber argue, reason evolved not to solve abstract problems but to help us navigate social life. This helps explain why people tend to accept fake content that aligns with their beliefs: our brains favor social cohesion over objective truth. This confirmation bias can lock us into unfair judgments and stereotypes. The question then arises: what happens when AI-generated content actively reinforces these biases?

AI isn’t just reflecting our biases—it can also amplify them. An article by Tiku, Schaul, and Chen in The Washington Post reveals that AI image generators like Stable Diffusion and DALL-E often produce images echoing racist, sexist, and Western-centric stereotypes. For instance, prompts for “attractive people” frequently return images of young, light-skinned individuals with European features, while “productive people” are typically depicted as white men in corporate attire. These outcomes largely stem from the datasets used to train these models—datasets that carry forward existing societal prejudices.

This raises a concerning link between human psychology and technological bias. As Kolbert observed, our deep-seated biases can become embedded in technology, leading to the systematic reinforcement of stereotypes. In line with Immerwahr’s view that the problem with fakes lies in what they reveal about us, AI-generated biases are particularly worrisome because they not only mirror our prejudices but also amplify them into an “average stereotype” that reflects dominant cultural views.

This process could lead to a self-reinforcing “bias feedback loop,” where society’s prejudices are encoded into AI systems that, in turn, intensify these biases. When people encounter AI-generated content reflecting such stereotypes, our natural confirmation bias only strengthens the cycle. The big question is: what real-world impacts might this bias feedback loop have on communities?

In her poem “AI, Ain’t I A Woman?”, Buolamwini shows that AI biases don’t just mirror prejudice—they can actively erase and misinterpret the identities of Black women. Testing commercial facial recognition systems, she discovered that prominent Black women like Shirley Chisholm, known as “unbought and unbossed, the first black congresswoman,” were misclassified. Even figures like Michelle Obama face similar confusion, with AI often misidentifying their features. This misrecognition underlines how AI systems, trained on datasets that favor white features, frequently fail to accurately represent Black women.

So how can we address this complex problem? Firestein offers a fresh perspective in his chapter “A Short View of Ignorance” from Ignorance: How It Drives Science. Rather than viewing science as merely an accumulation of facts, he argues that it is our gaps in knowledge—our ignorance—that drive innovation. By treating facts as a starting point for deeper questions rather than as finished products, we can challenge ingrained assumptions. For example, when AI image generators consistently produce light-skinned figures in response to prompts for “attractive people,” we should ask: why do these biases exist? What data informs these choices? Recognizing these questions can help break the cycle of narrow perspectives.

Firestein’s approach—to see facts as raw material for further inquiry—provides a potential way to counter our natural cognitive biases. Instead of letting first impressions cement our beliefs, we can choose to ask: What perspectives are missing? What assumptions underlie these outputs? Such a questioning mindset can help disrupt the bias feedback loop by turning our inherent predispositions into opportunities for deeper understanding.

That said, putting this approach into practice isn’t easy. Our evolutionary wiring, as Kolbert explains, is designed for social cohesion rather than rigorous truth-seeking. In today’s fast-paced world, constantly questioning our beliefs demands time and mental energy—resources that are often in short supply. This challenge is even greater with the rapid pace of AI-generated content, which can quickly overwhelm our capacity for critical analysis.

There’s also a social cost to questioning our beliefs. As Kolbert points out, our reasoning evolved to help us navigate social relationships. Challenging widely accepted views can sometimes put us at odds with our communities—whether risking professional relationships, cultural norms, or even our online reputation. Many feel that the social cost of challenging established beliefs is simply too high.

Yet, there is room for optimism. Tech companies—the creators of these AI systems—can take steps to rethink their approaches. They can ensure that diverse voices are included in the design process and critically examine the societal impact of their technologies. Educational institutions and policymakers also have crucial roles to play. Schools should teach students not only to use AI responsibly but also to question its outputs, while regulators work to balance innovation with protection for marginalized communities.

In summary, the interplay between AI and human bias forms a reinforcing cycle. Immerwahr has shown that people tend to accept content that confirms their views, Kolbert explains our stubborn belief systems, Tiku, Schaul, and Chen reveal how AI can amplify prejudices, Buolamwini highlights the harm to marginalized communities, and Firestein offers a strategy of questioning to break the cycle. As AI continues to influence decisions from hiring practices to law enforcement, addressing these biases will require a concerted effort by tech companies, governments, and individuals alike. Only by continuously questioning and challenging what we see can we hope to harness AI to reduce—rather than reinforce—human biases.