IB Digital Society 3.2E Algorithmic Dilemmas Notes

3.2E Algorithmic Dilemmas

The Double Edge of Algorithms

1. Algorithms are prized for being fast, consistent, and able to run at huge scale. An algorithmic dilemma arises when those strengths turn into problems. The same efficiency that makes an algorithm useful also magnifies any mistake it carries.
2. Because an algorithm applies the same rule to everyone, a single flaw repeats identically across millions of cases. A biased hiring algorithm can shape the chances of every applicant a company ever sees.
3. These dilemmas are about responsible use, not a reason to abandon algorithms. Four are sharpest: bias, the meaning of fairness, transparency, and the erosion of human judgment.

Algorithmic Bias

1. Algorithmic bias is a systematic error in an algorithm that produces unfair outcomes, often favouring one group. Systematic means a consistent tilt, not random noise, disadvantaging the same people repeatedly.
2. Bias rarely means deliberate discrimination; it creeps in through the data. A model trained on historical records absorbs past unfairness as its target. It can also enter through feature choices and through teams too similar to spot problems affecting others.
3. The COMPAS US risk-scoring tool flagged Black defendants as future reoffenders at nearly twice the false-positive rate of white defendants, though race was never written into the rules. The bias was inherited from a justice system's history, not invented.
4. The dangerous part is that the output looks neutral and mathematical, giving old prejudice a new air of objectivity. Because the bias is spread evenly and dressed in numbers, it can run for years before anyone proves it, all the while shaping who gets a fair chance (power).

What Fairness Really Means

1. Fairness has more than one definition, and they can conflict. One asks for equal error rates (wrong equally often per group); another asks for equal accuracy (a score means the same whatever group you belong to).
2. Mathematicians have shown that when groups differ in the data you usually cannot satisfy every reasonable definition of fairness at once; improving one measure often worsens another.
3. In the COMPAS debate the maker showed equal accuracy while critics showed unequal error rates; both were true because the two kinds of fairness cannot fully hold at once when groups differ. There was no hidden setting to satisfy everyone.
4. This makes fairness a value judgment rather than a technical box to tick. Which fairness matters most for a court, hospital, or hiring tool is an ethical and political choice. Good teams run bias audits and say plainly which definition they chose and why.
   1. The disagreement is about values rather than the numbers: a defendant cares about not being wrongly flagged, a court about accuracy, a community about equal treatment. Choosing among them is a values and ethics decision that should be made openly.

Transparency and the Black Box

1. Many powerful algorithms are black boxes: you see inputs and output but not how one led to the other. Some are secret for commercial reasons; others (large machine-learning models) are so intricate even creators cannot fully explain a decision.
2. Opacity makes other dilemmas worse: bias is harder to detect, mistakes harder to challenge, and a person refused a loan cannot meaningfully appeal. Transparency is the antidote, through explainable models, external audits, and open code.
3. Regulation is starting to demand it: Europe's GDPR gives people a right to meaningful information about the logic behind automated decisions that significantly affect them.
4. Stakeholders feel opacity unequally: the company keeps its advantage while the person refused a loan, visa, or job cannot see or challenge the decision. Since you cannot challenge what you cannot see, transparency is a precondition for accountability.

Accountability and Human Judgment

1. Algorithmic accountability is the principle that those who build and deploy an algorithm are responsible for its effects, rather than hiding behind the machine.
2. Responsibility is easy to pass along a chain: data suppliers blame developers, developers blame the deploying company, the company points to the regulator, and the algorithm decided is used as if no human were involved. In the end, a decision many people shaped has no clear owner.
3. Example: when an automated benefits or fraud system wrongly cuts people off (as in several countries), each party points elsewhere while the claimant has a real loss and no door to knock on.
4. Closing the gap needs clear lines of responsibility set before deployment, meaningful human oversight of important decisions, and rules making an identifiable party answerable, so that whoever holds the power to decide is the named party held answerable for the outcome.
5. As algorithms take over hiring, medicine, lending, and driving, human judgment shrinks. Overreliance means trusting a plausible output even when wrong, switching off the scrutiny that would catch errors.
6. Over time comes deskilling: people lose the ability to do the task or notice failures. The aim is to keep a meaningful human in the loop and treat algorithms as advisers that can be questioned, not final authorities.