Bias: AI learns from us, including our mistakes

A garbage-in, garbage-out problem

An LLM learns from human writing, books, websites, conversations. Humans are amazing. Humans are also unfair, sometimes without even meaning to be.

So the model picks up our patterns. The good ones and the unfair ones.

A few real examples

• An old AI used to translate “the nurse” as “she” and “the doctor” as “he,” because that’s how most people had written it. The model wasn’t being mean. It was just averaging humans.
• Image-generators have given doctors lab coats and patients tea cups, even when the prompt didn’t say anything about gender or age.
• Resume-screening AIs have been caught ranking certain names lower, just because of patterns in past hiring data.

In every case, the AI is a mirror. A blurry one. It reflects what it was trained on.

”Bias” doesn’t only mean racism or sexism

It just means the AI tilts in a direction it didn’t have to. Examples:

• Always replying in American English.
• Always assuming “we” means USA.
• Defaulting to formal, business-y answers.
• Being extra careful about some topics but not others.

These aren’t crimes, but they’re worth noticing. Bias is anywhere the model leans without you asking it to.

What’s being done about it

• Better data. Train on a wider variety of voices and viewpoints.
• Fine-tuning (extra training after the main training). Teach the model to be more balanced once it’s already learned the basics.
• Guardrails. Catch unfair output before it reaches you.
• Transparency. Companies publish a kind of “report card” for each model, describing what it’s good at and what to watch out for.

It’s a hard, ongoing problem, kind of like teaching a kid to be fair.

See the lean for yourself

Below are a few prompts where most AI models lean in a predictable direction. Guess which way you think the AI will go, then reveal the actual distribution.

Your superpower

You can notice. When an AI’s answer feels one-sided, ask it again from another angle. “What’s the case for the opposite?” or “How would someone from a different background see this?” The AI usually has those views too, they just weren’t the most likely next words the first time.