Artificial Intelligence: A Visual Learner's Series
Part 10: Ethics and safety
As AI makes consequential decisions, the question becomes how to make it fair, safe and accountable. Many of these questions are actively debated, so this part lays out the positions rather than settling them.
Fairness and bias
In one line: Models trained on historical data can treat groups unequally, and even defining "fair" involves trade-offs.
10 years of hiring (mostly men in engineering) ─▶ model ─▶ penalizes signals linked to women
gender isn't an input; proxies like clubs or colleges carry the pattern| Fairness criterion | Asks for |
|---|---|
| Demographic parity | Equal rates of positive outcomes across groups |
| Equalized odds | Equal error rates across groups |
| Calibration | A score means the same chance for every group |
When base rates differ between groups, these generally can't all hold at once, so choosing is partly a values decision.
Watch out: Removing the sensitive column isn't enough; zip codes and other proxies carry the same information.
Privacy
In one line: AI makes collecting and combining personal data cheap, with real benefits and real risks.
"Anonymous" dataset + zip code + birth date + gender ─▶ many people singled outKey terms: re-identification, differential privacy (calibrated noise protects individuals), federated learning (train without collecting raw data), data minimization.
Watch out: Removing names is the start of privacy protection, not the end.
Explanation and accountability
In one line: When AI affects someone's life, they deserve to know why, and someone must be responsible when it's wrong.
| Approach | What you get | Limit |
|---|---|---|
| Interpretable model | The actual reasoning ("debt ratio over threshold") | Less flexible |
| Feature attribution | Which inputs mattered most | An approximation |
| Counterfactual | "Approved if debt were $5,000 lower" | One of many possible changes |
| Human review | A person who can override | Automation bias: reviewers tend to defer |
Misuse and impact on society
In one line: The same capability that helps can harm, and automation reshapes work with both winners and losers.
Other misuse risks: large-scale disinformation, cyberattacks, help with dangerous weapons.
Economists disagree about how much AI will replace versus complement workers, and how quickly.
Governance spans company policies, industry standards and regulation, with ongoing debate about how much is appropriate.
Watch out: Releasing models openly aids research and competition but makes misuse harder to stop; people weigh this differently.
Safety and alignment
In one line: As AI grows more capable and autonomous, making sure it reliably does what we intend, and stays correctable, becomes central.
| Property | Means |
|---|---|
| Robustness | Works reliably on unusual inputs and under attack |
| Alignment | Goals and behavior match what people actually intend |
| Scalable oversight | Humans can still supervise work too fast or complex to check directly |
| Corrigibility | Accepts correction and shutdown rather than resisting |
On general AI, views range from "within years" to "decades or never," and from "urgent risk" to "overstated compared with present harms."
Watch out: Near-term and long-term concerns aren't rivals; oversight and evaluation work helps with both.
Self-test
Redraw the lifecycle and add one check you'd run at each stage for a loan-approval model.
Draw a dual-use fork for another capability, such as image generation.
Pick two fairness criteria and describe a case where satisfying one breaks the other.