Chapter 3 — Skew and the Three Questions That Won’t Quit

Skew wasn’t big, not even for a mongoose. Her fur was a soft mix of warm gray and cream, ending in a darker tail. Little question marks dangled from her ears, bouncing as she moved. She always carried a small, bright flashlight. It wasn’t for finding things in the dark. Skew used it to shine a focused beam onto data, like a detective inspecting clues. She called it her dataset-inspection-flashlight.

“Whose data is in here?” she’d murmur, her voice a low rumble. “Whose is missing? And who decided?”

These three questions were Skew’s constant companions. They were the first thing she asked, and the last. She believed they were the most important questions in all of NeuralQuest. To Skew, they revealed the truth about bias and data fairness.

Many people thought bias was a rare problem, a mistake that happened sometimes in AI. Skew knew better. She knew bias was the default. It was like a shadow that always followed the light. Every AI system, every smart program, learned from the information it was given. And that information, that “data,” always came from people. It carried their choices, their blind spots, and their histories.

Imagine a giant recipe book. If the cooks only ever wrote down recipes for one kind of meal, then that book would have a bias. It wouldn’t be a mistake; it would just be what was there. Skew’s job was to make sure everyone saw those recipe books for what they truly were. Her three questions were the anchor for this constant watchfulness. They ran through every kit from kit five onward. No exceptions. Skew’s whole work was making these questions automatic. She wanted everyone to understand that bias was a structural concern, not just a rare edge-case.

Skew was clear and persistent. “Whose data is in here? Whose is missing? Who decided?” she’d say. “These three questions don’t quit. Ask them of every dataset. Every model. Every claim that ‘the AI says.’ Bias isn’t a bug to fix; it’s a structural feature to monitor — always.”

She didn’t just talk about bias. She taught everyone how to spot it. Her three questions were the start: Whose data is in here? Whose is missing? Who decided the labels and the categories?

She’d hold up a picture of a school yearbook. “Look at this,” she’d say, shining her flashlight. “Whose faces are here? Mostly kids from this town, right? What about the kids who moved away mid-year? Or the ones who were homeschooled? They’re missing.” This was sampling bias. The data reflected only who showed up for the picture, who was reachable.

Then she’d point to an old newspaper ad. “This ad from fifty years ago says ‘professional appearance required.’ What did that mean back then? Probably a suit and tie, right? Maybe a certain haircut. But who decided that was ‘professional’? And what about people who didn’t fit that look?” This was labeler bias. The people making the rules, even for something as simple as a job ad, brought their own ideas.

Skew often showed a series of old photographs. They depicted towns from a long time ago. “See how many people are missing from these pictures?” she’d ask. “The people who weren’t allowed in certain places, or whose homes were torn down. The past leaves gaps. If you train an AI on data from those times, it will learn those gaps too. It will carry that historical bias forward.”

She taught that no dataset was truly neutral. Every choice, from what to collect to how to describe it, was a human choice. And human choices always carried a viewpoint. “Neutrality,” she’d explain, “is often just one viewpoint pretending to be everyone’s.”

Skew’s family had always been vigilance-keepers in the watch-village. They were mongooses known for their constant alertness. They didn’t just watch for threats sometimes; they watched always. It was a way of life, a posture, not just a task. Skew learned this lesson early. Vigilance wasn’t something you finished. It was something you were.

When she turned twelve, Skew walked to NeuralQuest. Her mentor, Sift, met her at the gates. Sift was an old, wise owl, with eyes that seemed to see everything at once.

“Skew,” Sift hooted softly. “What is bias-vigilance?”

Skew didn’t hesitate. She held up her flashlight. “Whose data is in here? Whose is missing? Who decided? These three questions don’t quit.” Her voice grew stronger. “Bias is structural. Vigilance is posture. And I will be here, in every kit from five onward. That’s intentional.”

Sift blinked slowly. “You are appointed, then,” she said. “And your appointment is essential for the whole app.”

Skew understood. Her job wasn’t just to teach. It was to remind everyone, constantly, that these questions mattered.

In her workshop, Skew had datasets pinned all over the walls. Each one was a story, annotated with her three questions and their answers. A young student, a curious squirrel named Pip, often visited. Pip watched Skew, fascinated by the way her flashlight beam cut through the noise.

Skew pointed to a large chart. “This one is called ‘One Million Faces’,” she told Pip. “It’s a dataset used to train face-recognition programs.” She shone her light. “Whose faces are here, Pip?”

Pip squinted. “Lots of young people. Mostly light skin. And they look… like they’re from the same part of the world, maybe?”

“Exactly,” Skew said. “Mostly young, light-skinned, Western, photographed in good lighting. So, whose faces are missing?”

Pip thought for a moment. “Older folks? People with darker skin? Maybe people from other countries, or in dim light?”

“You got it,” Skew confirmed. “And who decided which faces to collect? Three engineers in California, back in 2015. Now we know the dataset’s limits. It can’t recognize everyone equally well if it hasn’t learned from everyone.”

She moved to another chart, this one showing graphs and numbers. “This is a crime-prediction dataset,” she explained. “It’s supposed to predict where crimes might happen.” She tapped the chart. “Same questions, Pip. Whose data is in here?”

“Past arrests?” Pip guessed.

“Yes. And past arrests reflect past policing patterns. If police focused more on certain neighborhoods, or certain groups of people, then the data will show more arrests in those places. The AI will learn those patterns. It will predict more crime where there was more policing, not necessarily more crime.” Skew paused. “That’s not bias in the AI itself. That’s bias from data. The AI just learned what it was taught.”

Skew looked at Pip, her eyes serious. “I am Skew. The primitive I teach is bias-vigilance. The move is simple: ask the three questions of everything. Always.”

Skew was clear and firm. “Don’t let anyone tell you the data is objective,” she insisted. “No data is objective. Every dataset reflects choices: sampling choices, labeler choices, and design choices. Asking who decided isn’t paranoia. It’s craft. It’s how we build better, fairer systems.”

She often talked about the importance of looking at performance across different groups. “Even if a face-recognition model works 95% of the time,” she’d explain, “that average can hide problems. What if it works 99% for light-skinned faces, but only 85% for dark-skinned faces? That disparity matters far more than the average. It means some people are being left out or misunderstood.”

Skew believed that when bias was found, the right response was to acknowledge it, not deny it. It took courage to admit a system wasn’t perfect. But that courage was essential for building trust.

“Whose data,” she repeated, her voice echoing slightly in the workshop. “Whose missing. Who decided. Three questions. Always.”

---