Build the mental model: thermostat schedule

There is no single technical test accepted for every use of the term “artificial intelligence.” A practical definition is: a designed computational system that uses a model to produce an output for a task that people associate with prediction, perception, language, planning, or decision support.

The task boundary matters more than the label. A system may classify book-cover images well and fail on handwritten shelf labels. Capability depends on the task, data, people, and conditions.

Harbourview’s old thermostat follows a schedule written by a facilities officer: 20°C at opening, 16°C after closing. Its behavior comes from explicit rules.

A proposed controller uses past temperatures, occupancy readings, and energy use to predict heating demand. Its behavior depends partly on patterns fitted from examples. That is a common form of machine learning, a major approach within AI.

Both systems automate heating. Only the second is described as learning a predictive model from data. The distinction does not make it automatically better.

Mira compares both options:

• Task: keep occupied rooms comfortable while limiting energy use.
• Rule system: fixed timetable; easy to inspect; slow to adapt to unusual events.
• Learned system: predicted demand; may adapt; harder to diagnose when readings or occupancy patterns change.
• Evidence needed: comfort readings across rooms, energy use under comparable weather, and failure behavior when sensors are missing.
• Human owner: facilities staff, who can override either system.

Her decision is not “AI or no AI?” It is “Which method meets this task’s requirements with acceptable cost and risk?”

• “AI learns by itself.” People choose data, objectives, feedback, and deployment conditions.
• “AI is the same as a robot.” A robot is a physical system; it may use AI, fixed rules, or both.
• “Machine learning is all AI.” It is a central approach, not the only historical or conceptual approach.
• “AI replaces human responsibility.” A model has no organisational authority. People remain responsible for adoption and consequences.

For a photo classifier, fill in: task, learned pattern, test condition, and owner. A strong answer might say: “Classify covers as fiction or nonfiction; patterns fitted from labelled cover images; test on local languages and damaged covers; catalogue lead owns correction.”

Compare a fixed spam rule (“block messages containing this exact phrase”) with a learned spam filter. Name one advantage, one failure mode, and one test for each. Do not choose a winner until you state the operating conditions.