Through Time

Responsibility: who answers for what a machine did

Ethics in plain words, the well-known cases of algorithms gone wrong, who holds the power, and what an ordinary person can do.

A course for people who use AI, work beside it or live under its decisions, which is by now nearly everyone. You will learn to ask who made a decision and who will answer for it, to take apart disputed cases without picking a side in advance, and to tell what is known from what is contested. The question here is who should answer, and what the law says now is in the course on law.

So far there is only this table of contents. Lessons appear one at a time, and we promise no dates. Feed of new lessons (Atom)

17 lessons, 4 parts

Part 1 What ethics is

Not a list of prohibitions but a way of working out what to do when honest people disagree.

  1. Three questions of ethics

    What are the consequences, what is my duty, what kind of person does this make me: three old traditions, and why none replaces the other two.

    coming soon

  2. The moral machine

    A huge survey on whom a driverless car should save, and why it shows preferences, not what is right.

    coming soon

  3. No tool is neutral

    Who chose the data, the goal and what a model refuses to do: how other people’s values end up in a system, and how to spot them.

    coming soon

Part 2 Who answers

A machine cannot be held to account, so the question is always about the people who deployed it, bought it and believed it.

  1. A computer cannot be held accountable

    An old line credited to IBM: what being accountable means, and why it applies only to people and organizations.

    coming soon

  2. The responsibility gap

    If a system learns on its own, nobody seems to be at fault: why some see a new problem in this and others a convenient excuse.

    coming soon

  3. The human as a crumple zone

    Madeleine Elish’s image, with the pedestrian killed by a self-driving Uber: why the nearest person pays for a system’s failure.

    coming soon

  4. Horizon: believing the system over the people

    The British Post Office prosecuted its own subpostmasters for years on an accounting system’s data: how that was possible, and who answered.

    coming soon

  5. Robodebt: a debt worked out by averaging

    An Australian scheme sent people debts based on averaged income, and it was a simple formula, not AI: that is the whole lesson.

    coming soon

  6. Your name on the machine’s work

    What you take on when you send out text or a decision a model prepared, and what it means to check it well enough to answer for it yourself.

    coming soon

Part 3 Bias in systems

Almost everything here is disputed, and the dispute itself is worth learning.

  1. COMPAS: two truths about fairness

    The dispute between reporters and the maker over the algorithm’s errors by race, and why with different base rates both conditions of fairness cannot hold.

    coming soon

  2. Compared with whom

    How one algorithm looks beside a layperson, a group of people and a simple formula, and what to compare it with before you judge.

    coming soon

  3. Where bias comes from

    Data about the past, a proxy in place of the goal, who was left out of the sample: why one fix does not remove bias.

    coming soon

Part 4 Power and the rules

Who has the computing, the data and the decisions, who limits them, and what the worried and the unworried argue about.

  1. Who holds the power

    Computing and data in few companies, the unseen work of data labelers: what is fact here, and what is a dispute about whether that is bad.

    coming soon

  2. Who should write the rules

    Whom to trust with limiting powerful systems: governments, courts or the companies themselves, and why no answer is obvious.

    coming soon

  3. The safety debate

    Some warn of extinction and others call AI an ordinary technology: who claims what, on what grounds, and what stays disputed.

    coming soon

  4. What an ordinary person can do

    Ask who decided, insist on a human, complain, refuse to sign: not much, and more than nothing.

    coming soon

  5. A look ahead: agents that act on their own

    When an agent spends money and signs things on its own, responsibility becomes an everyday question: who pays for a mistake, and how to agree on it.

    coming soon