PHIL7002 Chap.1 Foundations of AI Ethics and the System Lifecycle
Foundations of AI Ethics and the System Lifecycle
A definition that settles later arguments
The course opens by describing AI ethics as a multidisciplinary field concerned with the moral standards that ought to govern how these technologies are built, released and run.
One clause in that description does a great deal of work later: the question is not whether a system is itself ethical or unethical, but whether the people who create and operate it are held to standards. Keep that framing and questions about responsibility, cost and remedy all become answerable, because each of them needs someone to address.
Drop it and you are left asking what a piece of software deserves.
Why the urgency claim matters for how you argue
The stated reason the field is urgent is not novelty.
It is that these systems have left research settings and now influence decisions in healthcare, criminal justice, finance and employment, where the effect on a person is direct and significant, and that harms scale because one model can make millions of decisions with identical logic. The same passage insists that there is tremendous potential to do good and that opportunities are lost by not developing beneficial systems.
That gives you a two-sided cost structure.
Foregone benefit counts, so an answer in this course rarely concludes that a technology should simply not exist; it concludes that a particular use, at particular stakes, needs a particular control.
Six principle families, and the stage each one binds
A stable set of principle families has emerged across published frameworks, and the course names six: fairness and non-discrimination, transparency and explainability, accountability and responsibility, privacy and data protection, safety with security and reliability, and human oversight and control.
They are not interchangeable.
Each is at its most demanding at a particular point in a system's life, which is why a question that names a stage usually has one family as its best answer.
The lifecycle, and the return arrow at the end
The course runs a group activity across six stages: data collection, preprocessing, model training, deployment, monitoring and redress, asking at each what could go wrong and what would catch it.
Was consent meaningful. Whose data is missing. Does the user know a system is involved and can they opt out. Who is accountable when harm occurs and is there an appeal. Redress deserves special attention, because a working appeal route is the mechanism by which the other failures become visible from outside a project.
Losing it does not add a fourth problem to three; it removes the instrument that would have surfaced the other three.
Prediction, classification and the two standard cases
Machine learning is introduced through two applications. Prediction estimates a future event, which means its errors surface later and often only to the person they were wrong about.
Classification assigns an instance to a class now, so an error is checkable immediately. That distinction decides what evidence a complaint needs.
The chapter closes on the trolley problem, used to expose the tension between counting consequences and following rules, and on a large online study of driving dilemmas that gathered forty million decisions across two hundred and thirty-three countries and found three major cultural clusters. Aggregated preference is evidence about acceptance rather than about what is right, and saying so is the mark of a careful answer.
What this chapter covers
- 01
AI ethics as a field about people rather than artefacts
- 02
Harm at scale, and foregone benefit as a cost
- 03
The six principle families and the stage each binds
- 04
Six lifecycle stages, ending in redress
- 05
Prediction against classification, and the evidence each needs
- 06
Model cards as disclosure rather than audit
- 07
Aggregated moral preference and what it can support
Place four complaints at the stage where the control belongs
- 3Assign the disclosure and the stale-ranking complaints.
- 3Assign the unappealable rejection and the missing households.
- 2Say which failure would have surfaced the others, and why.
Key terms
- AI Ethics
- The field studying the standards that ought to govern how artificial intelligence is built, released and run, with the people who build and operate systems as the object of judgement.
- Lifecycle Stage
- One of the six points at which the course locates an ethical question: data collection, preprocessing, training, deployment, monitoring or redress.
- Redress
- The stage at which a person harmed by a decision can obtain review, correction or a remedy. It is also the mechanism by which failures at earlier stages become visible.
- Prediction
- Estimating a future event or behaviour from patterns in past data, so that an error surfaces later and often only to the person it concerned.
- Classification
- Assigning an instance to a class, so that an error is checkable immediately by anyone who can inspect the instance.
- Model Card
- A standardised description of a model written by its developer, covering intended use, out-of-scope uses, data, metrics, limitations and ethical considerations.
Foundations of AI Ethics and the System Lifecycle FAQ
Why does the course insist that AI systems are not themselves the object of ethical judgement?
Because judging the artefact leaves nobody to address a recommendation to. Once the object is the people who design, deploy and operate a system, questions about who should have foreseen a harm, who should bear its cost and who owes an explanation all have candidate answers.
The framing also blocks a common evasion in which responsibility is described as having dissolved into the technology, and it is the reason later chapters can treat a deployment decision as reviewable even when an output was not foreseeable.
Is it a mistake to conclude that a risky system should not be deployed at all?
Usually, yes, because the course states that opportunities are lost by failing to develop beneficial systems, which makes non-deployment a choice with costs of its own rather than a neutral default.
A refusal can still be the right answer at very high stakes, but it has to be argued for against the benefit it forgoes, and the stronger answer almost always narrows the question instead, naming the use, the stakes and the control rather than the technology.
What does a large survey of moral preferences actually establish?
It establishes what rules a population would accept, which is a constraint on legitimacy rather than a source of moral content. If aggregated preference settled the question, a majority preference for favouring one group over another would resolve a discrimination case, and nobody thinks it does.
So the honest use of such a study in an answer is to say what a proposed rule would cost in public acceptance, while keeping the question of whether the rule is defensible separate.
Exam move
Draw the six stages from memory and write one real complaint under each, taken from a deployment you know rather than from a case study. Then check the principle family you attached to each and ask what evidence would settle it. If two complaints end up under the same stage with the same family, you have probably described the same failure twice, which is the commonest way this chapter is misused.
Working through Foundations of AI Ethics and the System Lifecycle in PHIL7002? Sia is AskSia’s AI Arts and Humanities tutor — ask any PHIL7002 Foundations of AI Ethics and the System Lifecycle question and get a clear, step-by-step explanation grounded in how PHIL7002 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
