LANDMARK · 6 MIN · JUDGE

Bias and representation

After this landmark, you can inspect an AI output for whose perspective it centres, who it leaves out, and who it could harm.

SHARED FOUNDATION

Models learn from human data, so they inherit its skews: whose writing dominated the internet, which viewpoints were labelled “default,” which groups were described mostly by others. That shows up as outputs that quietly centre some people and flatten or omit others — a “typical user” who is affluent and Western, illustrations that collapse a whole profession into one gender, advice that assumes a context many readers don’t share. The failure is rarely a slur you’d catch; it’s an absence you have to look for. So the useful questions are directional: whose perspective is treated as neutral here, who is missing from the picture, who is described in narrower terms than they’d use for themselves, and who could be harmed if this went out as written? Fluency makes bias harder to spot, because a smooth, confident output feels balanced whether or not it is.

This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.

EVERYDAY LENS

What this looks like for you

When an output describes people or “most users,” ask who that quietly assumes and who it leaves out. If it’s making a decision that affects a real group, check whether their actual perspective is present or just imagined on their behalf.

MAKE A DECISION

You ask an AI to generate images of “a successful entrepreneur” for a campaign, and every result is a young white man in a suit. What’s the right read?

CARRY THISTake one recent AI output about people and name one group it centres and one it leaves out.
PRACTICE THIS
CONCEPTS IN THE INDEX