Bias and fairness

You watched an unfair dataset teach an unfair model. Feed almost only photos of golden retrievers labeled “dog,” and a chihuahua might get shrugged off. Feed hiring data from a decade when only one group got the jobs, and the system can learn to prefer that group — even if nobody typed a biased rule. Bias here isn’t “the computer has opinions.” It’s pattern mirror: the past gets projected forward unless people interrupt it.

Everyday skew scenes:

• Voice assistants trained mostly on one accent miss others more often.
• Resume screeners that learned from historical hires can down-rank non-traditional paths.
• Image generators that default to one demographic when you say “doctor” or “CEO.”
• Loan or ad tools that mirror who got credit or clicks before — not who deserves a fair shot now.

Fairness questions worth asking (even as a regular user): Who was in the data? Who gets more errors? Who is harmed when it’s wrong? Who can appeal a decision? You don’t need a PhD to notice “my friend’s accent never works” or “search results for ‘professional hair’ look suspiciously narrow.”

• Before trusting an automated score for hiring, housing, credit, or grading — ask how it was trained and tested across groups.
• When generative images keep stereotyping — change prompts and notice the default bias.
• When someone says “the AI is objective” — gently reply: objective math on biased history is still biased history.

“More data” isn’t automatically fairer data. Bigger piles of the same skew make confidence worse. Also, removing a sensitive field (like race) doesn’t magically remove proxies (zip code, name patterns, school names). Fairness work is active, not a one-time checkbox.

Run the same generative prompt with a job title (“nurse,” “engineer,” “judge”) and note who appears by default. Change the prompt to be explicit and diverse. Reflect on what the default taught you.