What AI is actually good at
AI is not general cleverness. It is very good at a specific kind of task: finding patterns in large amounts of data, quickly, consistently and without getting tired. Every successful application below is that same capability pointed at a different problem.
It is correspondingly poor at anything requiring genuine understanding of an unfamiliar situation, at explaining itself, and at knowing when it is wrong — which is why the useful applications pair it with a human rather than replacing one.
| Field | What AI does | What it does not do |
|---|---|---|
| Healthcare | flags suspicious regions on a scan | decide treatment or tell the patient |
| Agriculture | identifies crop disease from a photograph | know that the field is waterlogged |
| Finance | flags unusual transactions as possible fraud | judge whether a person is trustworthy |
| Transport | recognises lanes, signs and obstacles | handle a situation never seen in training |
| Education | adapts practice questions to a student's level | notice that a child is upset |
| Language | translates and transcribes speech | grasp context, sarcasm or intent reliably |
Applications that matter in Pakistan
The syllabus asks for real uses rather than science fiction, and the strongest answers are local.
In agriculture, a farmer photographs a leaf and a model identifies the disease and suggests treatment — reaching farms an agricultural extension officer could never visit often enough. In healthcare, AI reading chest X-rays for tuberculosis lets a small number of radiologists cover a large population by flagging the scans that need attention. In banking, fraud detection watches transaction patterns continuously in a way no human team could. And in language, translation and speech recognition make services usable by people who do not read English or Urdu comfortably.
Why triage is the pattern that works
Notice that the medical example does not diagnose. It sorts — putting the scans most likely to show a problem at the top of the radiologist's list. The human still decides, but spends their time where it matters most. Applications built this way succeed because the AI's mistakes are caught, and because the value comes from speed rather than from replacing judgement.
The costs, honestly stated
Every exam question about AI applications wants the drawbacks as well, and vague worry scores poorly. The specific costs are these.
Employment. Automation removes some roles and creates others, but rarely for the same people in the same place — a data-entry clerk does not become a machine-learning engineer. Privacy. These systems need large amounts of data, often personal. Accountability. When an automated decision is wrong, it is frequently unclear who is responsible. Environmental cost. Training a large model consumes a great deal of electricity. Dependence. A skill that is always automated is eventually lost, which matters when the system fails.
A hospital proposes replacing its radiologists with an AI system that reads scans more accurately on average. Evaluate the proposal.
- Note the benefit: it is faster, works continuously, and does not tire or become less consistent late in a shift.Consistency is a genuine advantage — human accuracy falls over a long day.
- But "more accurate on average" hides the distribution of its errors.A system that is better overall may fail badly on an unusual case, which is exactly where expertise matters most.
- It cannot explain its reasoning, so a patient cannot be told why, and a wrong result cannot be argued with.Accountability for a medical decision cannot rest on an unexplainable output.
- It was trained on a particular population and may perform worse on patients unlike that group.A model trained abroad may not transfer to a different population.
- The sound proposal is AI as triage, with radiologists reviewing — the same accuracy gain, with the failures caught.Replacing the human removes the safeguard that makes the accuracy gain safe to use.
Use it to prioritise scans, not to replace radiologists — the accuracy gain is real, but its errors need a human to catch them.
Evaluating a proposal
Questions in this chapter usually describe a proposed use and ask whether it is a good idea. Four questions produce a full answer every time.
- What does it replace? A tedious task nobody wants, or a judgement someone is accountable for?
- What happens when it is wrong? An irritating recommendation, or a person denied medical care?
- Whose data trained it, and does that match who will be affected?
- Can the person affected find out why? If not, the decision should not be automated alone.
Before you leave this chapter
- AI finds patterns in large data quickly and consistently. That is the whole capability.
- The applications that work pair it with a human — triage rather than replacement.
- Name specific costs: employment, privacy, accountability, energy, deskilling.
- Ask what happens when it is wrong. The higher the stakes, the more human oversight is needed.
- A model trained on one population may not transfer to another.
Why an accurate model can still fail in the field
A model reported as 95% accurate has been measured on some test set. Whether that number survives contact with real use depends on whether the people it will be applied to resemble the people it learned from.
A disease-detection model trained on scans from one hospital may meet different equipment, different patient demographics and different image quality elsewhere, and quietly perform far worse. The accuracy figure was never wrong — it was answering a narrower question than the one that matters.
- Ask what the accuracy was measured on. A number with no stated test population means very little.
- Ask how the errors are distributed. A model 95% accurate overall may be 99% accurate for one group and 70% for another.
- Ask what a wrong answer costs. The same accuracy is fine for a recommendation and unacceptable for a diagnosis.
- Ask whether it can be monitored. A model whose performance is never re-checked after deployment will drift as the world changes.
Accuracy is misleading for rare events
If one transaction in a thousand is fraudulent, a model that simply answers "not fraud" every single time is 99.9% accurate and completely useless. That is why fraud and disease detection are judged on how many real cases they catch and how many false alarms they raise, not on raw accuracy. It is the single most useful thing to know about evaluating these systems.