The technical side of AI you've covered. The harder questions are the human ones: who does it work for, who does it work against, and who decides?
Where bias comes from in AI
Three layers, each adding bias:
Training data. What's on the internet, in books, in scraped PDFs — disproportionately reflects English speakers, the global North, certain demographics, certain time periods. The model absorbs all those defaults.
Human feedback (RLHF). Whoever was hired to rate model outputs brought their own preferences. If raters were mostly from one country with one set of cultural values, the model "learns" those values are correct.
Deployment choices. Which inputs the developer filters, what the system prompt allows, what the company's content policy refuses — all shape who the system serves well.
Result: AI that's noticeably better in English than in Swahili, that draws "a person" in defaults that match a specific demographic, that misrecognises faces of underrepresented groups more often, that struggles with names from non-Western cultures.
Real-world examples
Who decides what AI refuses?
Every chat product refuses some requests:
"How do I make a bomb?" → Refused.
"Write a violent story for my novel." → Often allowed.
"Tell me about [political figure]." → Allowed, but content varies wildly between products.
"Help me argue for [controversial position]." → Differs by product. Some refuse; some balance; some take a side.
These boundaries are policy decisions made by the companies building the model. There's no neutral choice — refusing too much makes the product unhelpful; refusing too little risks real harm. Different vendors have made different trade-offs:
Practical responsibilities for you as a user
You're not just a user of AI — you're a deployer of AI in your own context. A few principles:
The hard questions (without easy answers)
These don't have right answers. But they're worth thinking about:
Should AI replace human judgement in hiring, lending, medical triage, parole decisions? Some argue algorithms are more consistent than humans; others point to the legitimacy gap when an algorithm tells you "no."
Who owns AI-generated content? Courts are still figuring out copyright on training data and on outputs. Different countries are landing differently.
Should models reflect a particular set of values, or try to be "neutral"? Neutrality is itself a value. Pretending otherwise is the most common mistake.
What about jobs? AI is changing employment patterns faster than education and policy can adapt. The disruption is real; the answer isn't "ban AI" or "it'll all be fine."
The honest position is: we don't know yet. The right move is to keep paying attention, keep adjusting, keep involving the people most affected.
Check your understanding
1. Where does most of the bias in AI systems come from?
2. An AI tool you use for resume screening shows a pattern of rejecting candidates from one demographic. What's the responsible move?
3. Why does it matter who decides what an AI refuses?
Found this useful? Share it with someone learning AI.