Bias at Each Step
Bias doesn't come from one place. It can slip into an AI at three different steps — the data it is given, the labels humans add, and the design choices the builders make. This lesson teaches you to trace a bias back to the step it came from.
Kavi Traces the Bias
Kavi the owl's face-app keeps failing for one group of users. "It must be the data," she guesses — but when she checks, the data is perfectly balanced! So she keeps tracing. The people who tagged the faces had marked the same smile "friendly" for one group and "suspicious" for another. "Found it!" she hoots. "The bias didn't start in the data — it crept in at the labels step." Kavi nods wisely. "Bias can sneak in at any step — data, labels, or design. To fix it, first you have to trace which step it came from."
What your child will learn
A peek inside
Try this at home: What are the three steps where bias can enter an AI?
Questions parents ask
+What are the three steps where bias can enter an AI?
Data — which examples get collected; Labels — how humans tag or judge those examples; and Design — the choices the builders make, such as who they test with, what clues the model uses, and what counts as 'success'. Bias can slip in at any one of them.
+Why isn't it enough to just fix the data?
Because bias can enter later too. You can have perfectly balanced data, but if the humans tag it unfairly (labels) or the builders test only on people like themselves (design), the model still comes out biased. You have to trace which step let the bias in and fix that step.
+What does it mean to 'trace' a bias?
Tracing means following the bias back to its source: ask 'was a group left out when the examples were collected (data)? Were the human tags unfair (labels)? Or was it a builder choice (design)?' Naming the step is the first move to fixing it.
+Can you give an example of label bias?
Human taggers marked the same smile 'friendly' for one group of faces and 'suspicious' for another. The data was balanced, but the unfair tags taught the model an unfair pattern — that bias entered at the labels step.
+Can you give an example of design bias?
The builders chose 'home postcode' as a clue for the model, and a postcode can quietly stand in for one group of people. Choosing which clues the model uses is a design decision, so that bias entered at the design step.
+What age is this 'Bias at Each Step' lesson for?
It is a Stage 6 lesson for roughly ages 11–12 in KAVI Academy's AI-literacy track — top of primary, ready for secondary. It uses a story, sorting and tracing games, with light reading.
+Is the lesson free?
Yes — this overview and the printable PDF handout are free. The full interactive lesson and the practice are part of KAVI Academy.
Ready to play and learn?
Start the free interactive lesson, or print the handout for screen-free play.
Part of AI Literacy · Stage 6 · S6 on KAVI Academy — a secular, whole-child learning platform for ages 5–12.