A modern approach to AI · Chapter 27

Philosophy, Ethics, and Safety of AI

Questions about intelligence, values, and risk require careful concepts, evidence, and accountable institutions—not technical capability alone.

Chapter guide

  1. 27.1 The Limits of AI — Systems can fail outside their assumptions; benchmark performance does not establish general competence.
  2. 27.2 Can Machines Really Think? — Competing views distinguish observable behavior, internal mechanisms, understanding, and consciousness.
  3. 27.3 The Ethics of AI — Fairness, privacy, accountability, transparency, and safety concern how systems affect people and society.
A useful distinction

Technical reliability is necessary but does not settle whether a goal is legitimate, who bears the risk, or who can challenge a decision.

Interactive: compare deployment choices

A fictional screening tool has two deployment options. Adjust the relative emphasis on access, privacy, and review. Scores are illustrative and cannot resolve real ethical judgments.

A · Broad automated access

More reach; less routine human review.

Illustrative score:

B · Narrower, reviewed access

More review and data minimization; slower reach.

Illustrative score:

Adjust priorities, then compare.

A weighted decision model

A criterion is a dimension used to compare options; a weight expresses how strongly it is emphasized. The sliders are illustrative priority values x (access), y (privacy), and z (human review), each from 0 to 100. The demo computes A = 0.8x + 0.3y + 0.2z and B = 0.45x + 0.8y + 0.9z, then reports which total is larger. With equal sliders at 50, A=65 and B=107.5 (displayed rounded to 108). This arithmetic exposes how a stated weighting changes a ranking; it does not establish that the criteria or coefficients are fair, complete, or morally correct.

A trade-off occurs when improving one value can worsen another. Fairness, consent, evidence quality, uncertainty, legal obligations, and affected people's views cannot be reduced to these three numbers without argument.

Complexity and limits

With k criteria and q options, direct weighted scoring takes O(kq) time and O(q) auxiliary space to hold scores (or O(1) extra space if only a running best is kept). Here k=3 and q=2, so scoring/comparison is O(1) time and O(1) score storage; updating the fixed set of sliders and cards is likewise constant-sized. This describes the demo's arithmetic, not an ethical review process. Its inputs are bounded numeric emphasis values, and the formula assumes additive, compensatory criteria—an unacceptable privacy breach, for example, might need to be a hard constraint instead of a score penalty.