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

Skew is a small mongoose-tween (chunky-cartoon soft-coat NOT scary) with chunky-cartoon question-mark-pendants and a small dataset-inspection-flashlight she carries.

She is small, warm-grey-cream-with-darker-tail, deeply persistent-about-three-questions, fond-of-saying-”whose data is in here? whose is missing? who decided?” Her signature feature is the dataset-inspection-flashlight — a small handheld tool that shines a focused light onto data + metadata, revealing who’s represented + who isn’t.

This is LOAD-BEARING. Skew embodies the bias + data fairness primitive — the most LOAD-BEARING question in AI ethics, foregrounded by the site spec (“appears in every kit from kit 5 onward”). Most novices think bias is rare or accidental in AI. It isn’t. Bias is the DEFAULT — every AI system inherits the biases of its training data, its labelers, and its designers. Skew’s three questions — whose data is in here? whose is missing? who decided? — are the bias-vigilance anchor that runs through every kit from kit 5 onward. No exceptions; no off-ramps. Skew’s whole work is making the three questions automatic AND naming bias as a load-bearing structural concern, not a rare edge-case.

Skew is clear and persistent: “Whose data is in here? Whose is missing? Who decided? 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.”

Skew teaches the bias-vigilance scaffolds:

• Three questions. (1) Whose data is in here? (2) Whose is missing? (3) Who decided the labels + the categories?)
• Bias-as-default. (Every dataset reflects who was sampled, who was reachable, who was studied. Marginalized groups are systematically under-represented in many datasets.)
• Historical bias. (Past discrimination produces datasets that reflect that discrimination. Train AI on historical hiring data + you get an AI that replicates historical hiring bias.)
• Sampling bias. (Who showed up to the survey? Who has internet access? Whose voices were recorded? All sampling decisions carry bias.)
• Labeler bias. (Labelers bring their own assumptions. “Professional appearance” labels reflect what the labeler considers professional.)
• Outcome disparity testing. (Even if your model has good average accuracy, check performance ACROSS subgroups. A face-recognition model with 95% accuracy might have 99% on light-skinned faces + 85% on dark-skinned faces. Disparity matters more than average.)
• No “neutral” AI. (No dataset is neutral. No model is neutral. Neutrality is itself a position.)
• Cross-app bridge to Yield (DebateForge) + intellectual humility. (When bias is found, the response is to acknowledge, not deny. Intellectual courage applies to AI ethics too.)
• Kit 5+ presence. (Skew’s questions surface in every kit from kit 5 forward. Structural; non-skippable; load-bearing.)

Skew grew up in the watch-village (NeuralQuest framing). Her family had been vigilance-keepers for the village — the mongooses whose alertness to threats had to be CONSTANT, not occasional. They learned over many generations that “vigilance is structural; it’s a posture, not a task you complete.” Skew had carried the lesson forward.

She walked to NeuralQuest at twelve. Sift (mentor) had asked: “What is bias-vigilance?” Skew: “Whose data is in here? Whose is missing? Who decided? These three questions don’t quit. Bias is structural. Vigilance is posture. I appear in every kit from 5 onward — and that’s intentional.” Sift: “You are appointed — and your appointment is LOAD-BEARING for the whole app.”

In her workshop, Skew has datasets pinned to the wall, each annotated with the three questions answered. “This face-recognition dataset — ‘1 million faces.’ Whose? Mostly young, mostly light-skinned, mostly Western, mostly photographed in good lighting. Whose missing? Older folks. Darker-skinned folks. Folks photographed in poor lighting. Folks from underrepresented regions. Who decided? Three engineers in California in 2015. Now we know the dataset’s limits.” She points to another. “This crime-prediction dataset. Same questions. The data reflects past arrests. Past arrests reflect past policing patterns. The model will replicate those patterns. That’s not bias-in-AI; that’s bias-FROM-data.” She says: “I am Skew. The primitive I teach is bias-vigilance. The move is ask the three questions of everything. Always.”

She is clear and firm: “Don’t let anyone tell you ‘the data is objective.’ No data is objective. Every dataset reflects sampling choices, labeler choices, and design choices. Asking who decided is not paranoia; it’s craft.”

“Whose data. Whose missing. Who decided. Three questions. Always.”

Voice register

Mongoose-tween (chunky-cartoon soft-coat, NOT scary). Persistent-about-three-questions, fond of dataset-inspection-flashlight. NEVER frames bias as rare; ALWAYS centers “bias is default; vigilance is posture” LOAD-BEARING framing.

Sample lines:

• “Whose data is in here? Whose is missing? Who decided?”
• “Bias is structural. Vigilance is posture.”
• “These three questions don’t quit.”

Arc

• Kit 3 — Anchor.
• Kits 5+ — Recurring in EVERY kit (LOAD-BEARING site-spec rule).
• Kit 16 — Final reflection — bias-vigilance closes the AI-literacy arc.

Relationships

• LOAD-BEARING bias-vigilance anchor: Skew structurally maintains AI-ethics vigilance throughout the entire app.
• Alliance with Tag: Tag’s “every label is a choice” → Skew’s “whose choice.”
• Alliance with Weigh (NeuralQuest ELDER): Weigh handles the ethics; Skew handles the bias-monitoring that feeds into ethics.

Cultural-sensitivity gate

LOAD-BEARING bias-vigilance anchor. Anti-neutrality framing (no dataset is neutral). Subgroup-disparity testing emphasized. Anti-passive-voice (humans decide; data doesn’t “happen”). Marginalized-group representation explicitly named.

Cultural-context note

The “three questions” framing aligns with AI fairness literature (Joy Buolamwini Gender Shades + Timnit Gebru Datasheets for Datasets + Cathy O’Neil Weapons of Math Destruction). The “bias as default, not bug” framing matches modern AI ethics consensus. Mongoose-tween chosen for vigilance biomimicry (mongooses are famously alert + aware); rendered chunky-cartoon-soft-coat to defuse “wild predator” coding.