Breaking Boxes: Countering Bias in AI and Human Thinking

Human thinking is wired for simplicity. Our brains naturally gravitate toward categorization, seeking patterns and reducing complexity into manageable boxes, which makes us vulnerable to bias. While advantageous for efficiency, this tendency fuels binary thinking: black and white, good and bad, us versus them. It’s a cognitive shortcut that drives the appeal of populist narratives and Hollywood blockbusters. While harmless in the realm of fiction, this oversimplification becomes problematic when applied to human relationships and societal dynamics.

The Danger of Labels: Missing Nuances

Stereotypes are pervasive across all cultures, communities, and contexts. No one is immune to the temptation of applying labels to others. Nor are we protected from being categorized ourselves. Labels form the scaffolding of social constructs, shaping perceptions and interactions. The framework through which we judge future scenarios based on past experiences is narrow and often fails to capture nuance. It’s a net with wide holes, allowing critical details to slip through unnoticed.

The brain’s reliance on labels and patterns aligns with Bayesian logic: We interpret new experiences through the lens of prior probabilities. While this mechanism aids in decision-making, it is fraught with limitations. Categorizing based on past labels often blinds us to differences in the present—details that may hold the key to deeper understanding. This binary logic, where anything not fitting neatly into one category is relegated to the "other" box, can perpetuate stereotypes and hinder meaningful connections. It also denies us the beauty of diversity—the spectrum of possibilities that exist beyond reductive boxes.

This limitation doesn’t just affect human relationships; it reverberates through systems, institutions, and technologies—particularly artificial intelligence.

Why Bias Matters in the Age of AI

Just like natural intelligence, artificial intelligence is not immune to cognitive biases. In fact, it amplifies them. The concept of data as a proxy plays a significant role in how AI systems are trained and operate. Proxy data is used when direct measurement of a variable is impractical or impossible, relying instead on closely correlated metrics. For instance, tree rings serve as proxies for historical climate conditions, and website traffic is often used as a proxy for consumer interest.

This reliance on proxies permeates every step when training large language models like ChatGPT. From the biases of the data scientists designing the model to the annotators labeling datasets and the users crafting prompts, inherent assumptions are encoded into the system. The result? Algorithms that reflect and potentially exacerbate the biases embedded in the data they are fed.

The Challenge of Bias in AI

Bias is not just a technological flaw; it is an echo of human cognition. Our tendency to stereotype and label is hardwired into the way we process information, and when algorithms are trained on human-generated data, these tendencies are replicated. AI systems, including generative models, inherit the same blind spots as their creators. Making this even more challenging and preoccupying is the use of training data that is not representative of the wider population, often containing historical inequalities or societal stereotypes.

Furthermore, even the design choices of apparently neutral algorithms can inadvertently perpetuate biases. When the developers or data collectors lack diversity, it leads to skewed datasets and biased outputs. The frontier models that we are presently using were built primarily by young white males.

Unfortunately, this is not an abstract or academic debate—algorithms perpetuate social inequalities. For example, facial recognition technology has been shown to have higher error rates for people with darker skin tones; also, when algorithms are trained on data from a company with a predominantly male workforce, the hiring tool might favor male candidates over equally qualified women; and AI systems used for credit scoring tend to unfairly penalize individuals from specific demographics.

A Way Out: From Awareness to Accountability

The path to mitigating bias in both human cognition and AI systems is not easy, but it's necessary. While the broader, systemic implications loom large, change toward online equality begins with acute personal awareness—acknowledging our predispositions and understanding how they shape our interactions with others and technology. Once it is established at the individual level, this self-awareness can extend to the development and use of generative AI.

The A-Frame offers a framework to navigate this challenge:

• Awareness: Recognize biases in yourself and the systems you navigate.
• Appreciation: Value diversity in perspectives, data, and contexts.
• Acceptance: Acknowledge limitations—both in human cognition and AI.
• Accountability: Take responsibility for your decisions and their outcomes, ensuring they align with ethical principles.

Bias Essential Reads

Would You Wear Hitler's Sweater?

As Simple as M&Ms: Teaching Kids About Racism Through Play

Beauty Beyond Boxes

Ultimately, the goal is to move beyond binary thinking—beyond biased boxes that constrain human potential and technological innovation. It is urgent to train our minds today, before their narrow version shapes the infrastructure of tomorrow’s algorithms.

When we blindly apply labels from the past, we lose out. Applying the black versus white, good versus bad principle—whereby everyone who does not fit neatly into one category automatically falls into the bad box—is harmful. We miss the rainbow that unfolds beyond boxes. Why should we deprive ourselves?