Key Idea: A machine that acts is not a program with wheels. When software is wrong you get a wrong answer and somebody reads it; when a machine that acts is wrong, the thing has already happened — so its safeguards must be preventive rather than corrective.
Paper 1
- Part a: identify robot types or characteristics of autonomy
- Part b: explain why degree of autonomy matters
- Part c: evaluate robots in care, or autonomous weapons
Paper 2
- Q2: analyse a source's claim about an autonomous system
- Q3: compare sources on safety and responsibility
- Q4: synthesise sources on automation and work
Everywhere else
- AI: the same responsibility chain, with physical consequences
- Health: care robots meet the ageing-population argument
- Conflict: target selection is the sharpest case in the subject
Autonomy is a range, not a category
| Level | Who approves an action | What a failure means |
|---|---|---|
| Assisted | A person, every time | A suggestion is ignored |
| Supervised | A person, who can intervene | Depends on whether they had time |
| Autonomous | Nobody, action by action | The event has already occurred |
'Autonomous' is applied to systems at every point on that range, so it does not answer the question that decides the consequences — whether a person approves each action.
Why testing is genuinely hard here
Three linked difficulties
- Learning from live errors means accepting them, because an error observed in operation has already produced its consequence.
- Simulation contains only imagined situations, and the situations causing the worst outcomes are the ones nobody anticipated.
- Safety moved from separation to perception. A cage could not fail quietly; a perception system can.
Important: Most digital governance is corrective — observe a harm, measure it, write a rule. Irreversibility removes that sequence, so approval, physical limits and restrictions on use have to come first, and being wrong in the cautious direction has to be acceptable.
Exam-style questions
Suggest two conditions under which a machine should be allowed to act without a person approving each action.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
Justify the claim that a robot used in care should not be introduced to reduce staffing.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
Quick check
Why is a machine that acts different? No person stands between the output and the world, so the error and its consequence are simultaneous.
Why does the label 'autonomous' tell you nothing? It is applied across the whole range. Ask whether a person approves each action.
What changed about safety? It moved from separation, which cannot fail quietly, to perception, which can.
What must governance do here? Act before deployment — approval, physical limits, restrictions on use — because the corrective sequence is unavailable.
Exam tips
- 'Tasks, not jobs' applies here too. Automation removes tasks and changes the job that remains.
- For weapons questions, propose a narrow definable ban rather than a general one.
- Physical safeguards that software cannot override follow directly from the irreversibility argument.