IB Digital Society 3.6E AI Dilemmas Notes

3.6E AI Dilemmas

What Makes AI a Dilemma

1. AI now shapes high-stakes decisions: who gets a loan, a job interview, a medical flag, or a longer prison sentence. The questions it raises are ethical as much as technical, and they rarely have a clean answer.
2. These are dilemmas, not simple problems, because benefit and risk come bundled. A faster decision can hide bias, cheaper service can cost jobs, and wider access can eat into privacy. Higher accuracy often arrives with no explanation attached.
3. The values and ethics concept means naming the competing goods, identifying who gains and who bears the cost, and accepting that stakeholders weigh trade-offs differently.

Fairness and Bias

1. The most discussed dilemma is fairness. Because a model learns from data, any bias in that data becomes algorithmic bias in outputs, reproducing or amplifying unfairness while appearing neutral and objective.
2. Bias creeps in through the data: if past hiring, lending, or policing was skewed, the model learns those patterns; if some groups are under-represented, it performs worse for exactly the people already disadvantaged.
   1. Documented case: audited facial-analysis systems had error rates under 1 percent for lighter-skinned men but over 30 percent for darker-skinned women, because training images skewed toward the former. The bias was unintended but real.
3. Contested: fairness itself is contested. Several mathematical definitions of fairness cannot all be satisfied at once, so making a model fairer by one measure makes it less fair by another. There is no purely technical fix.