The Expanding Reach of AI

You have now seen AI applied to medicine, finance, transport, science, education, and the arts. The tools differ from field to field, but the underlying lessons repeat. If you can name those patterns, you can reason about a field you have never studied — and see through a headline before you have any expertise in it.

The stakes are practical. Organisations spend heavily on systems that stumble in the real world, while quieter, well-scoped deployments deliver steady value. The difference is rarely the model's raw accuracy. It is whether the deployment respects what AI does well and what it does not.

Where AI genuinely helps

Three strengths recur. Scale: AI handles volumes of data or decisions no human team could review — every transaction, every scan, every log line. Pattern-finding: it surfaces regularities in high-dimensional data that people miss, from tumours to fraud rings. Augmenting experts: it drafts, flags, and triages so a skilled human spends attention where it counts. Notice these are all about handling volume and surfacing candidates, not making the final call.

Where AI struggles

The weaknesses recur too. Judgment: weighing competing values, or deciding what should happen, not just what usually does. Context: the messy, shifting conditions of the real world — a new camera, an unusual patient, a rule that changed last week. Accountability: a model cannot be responsible, explain itself under scrutiny, or be held to account when it is wrong. These are exactly the places a bold claim tends to fail quietly.

A framework for any "AI can do X" claim

Ask five questions. (1) What exactly was measured, and on what data? (2) Under what conditions — a lab benchmark or a messy deployment? (3) Compared to whom, doing what? (4) What happens on the hard, rare, or high-stakes cases? (5) Who is accountable when it is wrong? A claim that survives all five is worth taking seriously. Most headlines answer only the first.

A headline reads: "AI matches radiologists at detecting disease." Run the framework. Measured: accuracy on a curated image set. Conditions: a clean benchmark, not a working clinic. Compared to: radiologists reading in isolation, without patient history. Hard cases: often excluded from the test set. Accountable: unstated. The honest reading is narrower — the system may match average performance on typical images under ideal conditions. That is genuinely useful as a second reader, and far short of "replaces radiologists."

Augmentation is not always the answer. Spam filtering, optical character recognition, and sorting high-volume, low-stakes items are fully automated for good reason: the task is well-defined, errors are cheap and correctable, and no one wants a human reviewing every email. The framework is not "always keep a human." It is "match the deployment to the task" — full automation fits narrow, tolerant, well-scoped problems; augmentation fits open, high-stakes, context-heavy ones.

In 2020, researchers from Google Health published a human-centred study (Beede and colleagues, presented at the CHI conference) of a deep-learning system deployed to screen for diabetic retinopathy — an eye disease — across clinics in Thailand. In the lab the model was highly accurate. In the field it struggled: poor lighting and image quality caused it to reject many photos, slow internet stalled results, and nurses and patients grew frustrated. Where the workflow fit the nurses' real conditions, it sped up referrals; where it ignored them, it added friction. The model was strong; the deployment succeeded or failed on how well it respected the people and context around it.

• "Higher benchmark accuracy means it will work in the field." Deployment conditions, not benchmark scores, decide real-world value.
• "AI will simply replace the experts in this field." The durable pattern is augmentation; full replacement works only for narrow, tolerant tasks.
• "If it works in one setting, it works everywhere." Context shifts — new equipment, populations, or rules — quietly break models.
• "The model itself can be held accountable." Responsibility stays with the people and organisations deploying it.

Take any "AI can do X" claim you have seen this week and run the five framework questions on it. Write one sentence stating what the claim actually supports, and one naming the question it fails to answer.

Think Like a Maester: Before asking whether AI can do a task, ask what happens on the cases where it is wrong — and who answers for them.

1. Name the three recurring strengths of AI applications described in this lesson.
2. Name the three areas where AI applications recurrently struggle.
3. What is the first question to ask of any "AI can do X" claim?
4. Why is full automation appropriate for spam filtering but not for most medical decisions?
5. In the Thai diabetic retinopathy deployment, what caused the field difficulties despite strong lab accuracy?

Across every application in this module, the same patterns hold. AI earns its value through scale, pattern-finding, and augmenting skilled people; it struggles with judgment, context, and accountability. The most reliable deployments pair AI with humans and match the level of automation to the stakes and messiness of the task. Armed with five questions — what was measured, under what conditions, against whom, on the hard cases, and who is accountable — you can evaluate an "AI can do X" claim in any field, calmly, and without needing to be an expert first.

• Artificial Intelligence in Everyday Life
• Artificial Intelligence