Limits and Society

This is Unit Ten of Artificial Intelligence, Part Two: deployment. The first entry sets the terms, and it is worth stating plainly. Every ethical problem in deployed systems here is a mechanism this course already taught, arrived at the scale where its costs land on people. So nothing new is introduced. Learning from examples produced a system that reproduces its examples. Opacity at depth produced systems that cannot explain themselves. Optimising a written objective produced systems pursuing exactly what was written. Three error entries, and they correct three ways of getting the responsibility question wrong — that the arithmetic makes a system objective, that no one is responsible when a machine acts, and implicitly that the gap between the two is unavoidable. It is not: it is engineered. And the course closes on what it asks of the person taking it, rather than on a technical result. By the end you will be able to do five things. One: say why calling an algorithm objective misdescribes it. Two: name the three channels bias arrives by. Three: say why fairness cannot be settled technically. Four: say why the responsibility gap is a difficulty rather than an impossibility. Five: say what an optimiser optimises.

One entry framing everything after it, and the framing matters. This is not a new subject bolted onto a technical course — it is the same mechanisms, at the scale where they touch people.

The framing.

AI 0591 The Ledger, Presented

Every ethical problem in deployed systems is a mechanism this course covered, arrived at the scale where its costs land on people. The ethics is the technical content, met at size.

Five entries, and the second is the error entry that matters most. Arithmetic is neutral in exactly the way a photocopier is neutral — it reproduces what it is given, faithfully and quickly.

What a learner learns.

AI 0592 Fingerprints Become Verdicts

A system trained on human decisions learns those decisions as the correct answer. Historical patterns get codified and then applied at speed — which is Unit Eight's mechanism, with consequences.

And the error.

AI 0593 "The Algorithm Is Objective — It's Math"

The arithmetic is neutral the way a photocopier is — reproducing its inputs quickly and accurately. The inputs are human decisions, so neutrality of the machinery says nothing about neutrality of the output.

Three channels.

AI 0594 Three Kinds of Bias

Bias reaches deployed systems through what the data records, through the judgments of the people building it, and through the statistics of how examples were gathered. Three different channels needing three different remedies.

And the loop.

AI 0595 The Feedback Loop

A deployed system can shape the evidence it later learns from — one that denies opportunities to a group collects no data about how they would have performed. The error becomes self-confirming.

Why no technical fix settles it.

AI 0596 Fairness Is a Decision

Several reasonable definitions of fairness exist and they conflict — satisfying one can require violating another. So which to satisfy is a decision somebody makes rather than a property a system can be verified to have.

Four entries on explanation. What a person refused by a system is owed, a distinction between two kinds of opacity, and two ways of responding when the model itself cannot answer.

What is owed.

AI 0597 Owed a Reason

A person refused by a deep model is owed a reason the model cannot produce by being read. So reasons have to be manufactured by other means, or the obligation goes unmet.

Two kinds.

AI 0598 Standard and Deep Opacity

Some opacity is standard — removable in principle with effort and access. Some may be deep and ineliminable. Treating the first kind as the second excuses work that could actually be done.

The first response.

AI 0599 The Behavioral Audit

An unreadable model can still be examined from outside — matched cases submitted at scale, with outcomes measured. Behaviour is observable even where internals are not.

The second.

AI 0600 Auditable by Design

High-stakes systems can be built as pipelines of tested, inspectable components — trading some capability for accountability. That is a design decision available in advance rather than a constraint discovered later.

Four entries, and the third and fourth are what make this tractable. Responsibility for a learned system is genuinely distributed and genuinely hard to allocate — and the difficulty is a property somebody engineered in.

Where it travels.

AI 0601 The Responsibility Gap

When a learned, semi-autonomous system harms someone, responsibility runs through designers, trainers, deployers and operators — no single hand holds all of it.

And the error.

AI 0602 "The Machine Did It, So No One Is Responsible"

The gap is an allocation difficulty rather than an allocation impossibility. Every element of a deployed system was chosen by somebody — the question is hard to answer, not unanswerable.

One repair.

AI 0603 Human in the Loop

Putting a person in meaningful control re-attaches a responsible agent to the decision point, at a cost in speed and scale. The cost is real and so is what it buys.

And the key point.

AI 0604 The Gap Is Chosen

Opacity and distributed agency are engineering properties with engineering alternatives. So how much of a responsibility gap a system has is a decision made during its design.

Three entries on the specification problem, one worked instance at civilisational scale, and an entry pointing out that the standard fix moves the question rather than answering it.

The problem.

AI 0605 Written vs. Meant

An optimiser pursues what was written down, not what was intended. And the gap between the two matters more as optimisers get more capable — a weak one cannot exploit the difference, and a strong one can.

The instance.

AI 0606 The Engagement Machine

Platforms optimising for engagement produced polarisation, manipulation, and synthetic-media harms — alongside real benefits. Both came from the same objective, faithfully pursued.

And the standard fix.

AI 0607 Preference Tuning Relocates the Question

Steering systems by human judgments moves the specification problem rather than removing it — those judgments were made by particular people under particular conditions, and now they are the specification.

Two entries closing the course. An institutional answer that treats these as risks to be managed across a lifecycle, and then a final entry addressed to the person watching rather than to a builder.

The institutional answer.

AI 0608 Govern, Map, Measure, Manage

Governance frameworks answer with risk management — naming a system's risks, measuring them, and owning them across its whole lifecycle rather than at launch. It is an ordinary engineering discipline applied here.

The final entry.

AI 0609 The Citizen's Part

Deploying learned systems on people buys consistency and scale, and pays in fingerprinted decisions, manufactured explanations, and distributed responsibility. Knowing the mechanisms is what lets a citizen price that trade.

Five things you can now do, and the course is complete. You can say why calling an algorithm objective misdescribes it — the arithmetic is neutral the way a photocopier is, and the inputs are human decisions. You can name the three channels bias arrives by — what the data records, the judgments of the people building it, and how the examples were gathered — and say that they need different remedies. You can say why fairness is not settled technically: several reasonable definitions exist and conflict, so choosing between them is a decision. You can say why the responsibility gap is a difficulty rather than an impossibility, and that how large it is was decided during design. And you can say what an optimiser optimises — what was written, not what was meant — and that steering by human judgment relocates the problem.