The Bias in the Bones – How AI Made Me Confront My Own Prejudices

1. The Face That Wasn't There

I was testing a facial recognition system for a class project.

The model had been trained on a large, publicly available dataset. I fed it a photo of myself – a fairly standard headshot, good lighting, neutral expression. It identified me correctly. I fed it photos of my friends. Most were identified correctly, though accuracy varied.

Then I fed it a photo of my neighbor, an elderly Black woman named Mrs. Patterson who had lived next door for fifteen years.

The model returned: “No face detected.”

I tried again. Different angle. Still nothing. I tried a photo of her grandson, a teenager with dark skin and a bright smile. The model detected a face but labeled it with low confidence, suggesting it might be an animal.

I sat back, stunned.

These were clear, high‑quality photos. The kind of photos any human would immediately recognize as faces. But the AI couldn't see them. Not because there was anything wrong with the photos – but because the training data had been overwhelmingly white.

2. The Data That Built the Mirror

AI models learn from data. If the data is biased, the model is biased. This is not a bug; it's a feature of how learning works.

The facial recognition dataset I had used was estimated to be over 75% male and over 80% light‑skinned. The model had never seen enough dark‑skinned faces to learn what they looked like. To the model, Mrs. Patterson's face was an anomaly, an outlier – something to be dismissed rather than recognized.

I felt a wave of guilt. I had built this system. I had chosen the dataset without thinking about its composition. I had assumed that because the data was large, it was representative. I was wrong.

But the guilt quickly spread beyond the project. I started thinking about my own “training data” – the experiences, media, and people that had shaped my perceptions.

I grew up in a predominantly white suburb. The TV shows I watched featured mostly white casts. The news I consumed covered crime in a way that subtly associated Blackness with danger. My “dataset” was biased. And that bias had shaped my own “model” of the world – my assumptions, my instincts, my split‑second judgments.

I had never chosen to be biased. But I had never actively sought to correct my training data either.

3. The Stereotype That Whispers

AI bias is often explicit: a hiring algorithm that downgrades women's resumes, a loan approval system that charges higher interest rates to minority applicants. But human bias is more insidious. It lives in the milliseconds between seeing a face and making a judgment.

I remember crossing the street once when I saw a young Black man walking toward me late at night. I told myself it was just caution, just being smart. But was it? Would I have crossed the street if he were white?

I don't know. And that not‑knowing is the problem.

The AI model didn't intend to be racist. It had no intentions at all. It simply optimized on the data it was given. My brain is similar: it optimizes on the data it has absorbed, building heuristics and shortcuts to navigate the world quickly. Some of those shortcuts are useful. Some are poisonous.

The difference is that I can choose to audit my own biases. I can seek out counter‑evidence. I can deliberately expose myself to data that challenges my assumptions.

The AI couldn't fix itself. I had to fix the AI. And I have to fix myself.

4. The Correction Loop

After my facial recognition failure, I retrained the model on a more balanced dataset. I found a collection called “FairFace” that explicitly balanced for race and gender. The new model performed dramatically better on Mrs. Patterson's photos. It wasn't perfect, but it saw her face.

That experience became a metaphor for my own growth.

I started reading books by authors from backgrounds different than mine. I subscribed to news outlets that covered stories I wouldn't otherwise see. I made a conscious effort to notice my own snap judgments and question them.

It was uncomfortable. There were moments of defensiveness, of wanting to argue that I was “not a racist” and therefore didn't need to do this work. But I remembered the AI. The AI wasn't malicious – it was just undertrained. And I was undertrained too.

We are all undertrained. No one is born with a perfect understanding of justice, of empathy, of the full humanity of people different from themselves. We learn it – or we don't. The choice is whether we will seek out better data or remain in our comfortable, biased training sets.

5. The False Positive of Fear

One of the most troubling aspects of AI bias is the false positive – the model that sees a threat where none exists.

In some law enforcement facial recognition systems, false positive rates are significantly higher for people of color. The model doesn't just fail to recognize them; it actively misidentifies them as someone else, often someone on a watchlist.

Human fear works the same way. We see a stranger and our brain runs a threat assessment. If our training data has overrepresented certain groups as dangerous, we will see danger where there is only a person walking home from work.

I have felt that false positive. I have felt my body tense when a group of teenagers laughed loudly on the subway, and I have had to ask myself: would I feel the same if they were wearing business suits? If they were white? If they were women?

The answers are uncomfortable. But discomfort is not harm. Discomfort is the signal that my model needs updating.

6. The Responsibility of the Engineer

AI systems are not neutral. They carry the values, oversights, and blind spots of their creators. Building ethical AI requires more than technical skill; it requires humility, curiosity, and a willingness to be wrong.

I am not an AI researcher at a major lab. I am just a person who learned to code. But I believe that everyone who uses AI – which is almost everyone now – has a responsibility to understand its limitations. To question its outputs. To remember that behind every algorithm is a dataset, and behind every dataset are human choices.

When I use a translation app, I remember that it might not handle my friend's dialect well. When I use a recommendation engine, I remember that it's optimizing for engagement, not for my wellbeing. When I see an AI‑generated image, I remember that it was trained on a slice of the internet that is not the whole world.

This skepticism is not cynicism. It is a form of care. It is the recognition that tools are not magic – they are extensions of us, with all our flaws.

7. The Face I Finally Saw

A year after my project, I showed Mrs. Patterson the new model.

I had taken a photo of her in her garden, smiling, wearing a bright yellow hat. The model detected her face instantly and drew a box around it with high confidence. It didn't know her name, but it knew she was there.

She laughed. “So the computer can see me now?”

“Yes,” I said. “It learned.”

She looked at the screen, then at me. “And what about you? Did you learn anything?”

I told her about the biases, about the retraining, about the books I had read and the questions I had asked myself. She listened quietly. When I finished, she nodded.

“Good,” she said. “Keep learning.”

I will.

The AI model will never be perfect. Neither will I. But both of us can improve. Both of us can seek better data, adjust our weights, and try again. That is the work of a lifetime – not to eliminate bias, but to see it clearly and correct it, one epoch at a time.

Mrs. Patterson's face is not an anomaly. It is a face, full of years and stories and a dignity that no algorithm can fully capture. But the algorithm can at least learn to see it. And so can I.