What AI Can and Cannot Do

You will regularly be asked to have an opinion on whether some AI idea is realistic, without the background to evaluate it technically. This lesson gives you a defensible way to do that.

The rule of thumb

A widely used heuristic, popularised by Andrew Ng, is this:

If a typical person can do a mental task with less than one second of thought, we can probably automate it with machine learning — now or in the near future.

Test it. Is there a face in this photo? Under a second. Is this email spam? About a second. Is that a stop sign? Instant. All solved to a useful standard.

Now the other side. Should we enter the Vietnamese market? Why did this customer really leave? Is this contract a good deal for us? None of these are one-second judgements, and none of them are close to solved.

The rule is not a law of nature — it is a fast filter that is right more often than a non-specialist's intuition. Its usefulness is that it moves the conversation from "is AI powerful enough" to "what kind of task is this actually".

A second condition sits alongside it: there must be a lot of labelled examples. A one-second task with no data is still not a machine learning project.

🔗 Match the Pairs

Reading a licence plate from a photoDrop here

Transcribing a recorded callDrop here

Deciding which of two acquisitions to pursueDrop here

Explaining why revenue fell last quarterDrop here

Flagging a transaction as probably fraudulentDrop here

Writing next year's company strategyDrop here

Where machine learning tends to fail

Four recurring conditions. Recognising any of them early saves a great deal of money.

1. Not enough labelled data. The most common blocker by far. The task is well-shaped, and there are two hundred examples where thousands are needed.

2. The answer is not in the input. Covered in lesson 1. If the cause of a late invoice lives in an unlogged phone call, no model can recover it from the invoice.

3. The world changes faster than the data. Google Flu Trends, from the previous lesson. Also any model trained on behaviour from before a pricing change, a regulation, or a market shock.

4. Errors are unacceptable and there is no safety net. A model is accurate, not correct. If being wrong once in twenty is catastrophic and nothing catches the failure, the honest conclusion is that this system needs a human in the loop — or should not be built.

Two capabilities people conflate

Making a decision is not the same as explaining it. Deep learning models are frequently excellent at the first and poor at the second. If your regulator, your customer or your board needs to know why, that requirement changes what you can build — sometimes forcing a simpler, less accurate but inspectable model. That is a legitimate engineering trade-off, and it should be raised at the start, not discovered at an audit.

Predicting is not the same as causing. A model can tell you which customers will churn. It cannot tell you what to do to stop them. Those are different questions, and the second usually needs an experiment rather than a model. Teams lose whole quarters to this confusion: they build an accurate churn predictor, then discover it offers no guidance on intervention.

"AI to automatically approve or reject expense claims."

One-second task? Largely yes — a finance clerk glancing at a receipt and a category makes most calls quickly. Labelled data? Yes, if past claims carry their approve/reject outcome. Answer in the input? Mostly. Verdict: plausible. Watch the tail: unusual claims are exactly where the model is least confident and the cost of error is highest, so route low-confidence cases to a person rather than automating end to end.

"AI to predict which employees will resign in the next six months."

Technically often feasible — the data exists and the label is unambiguous. But apply the other tests. Is the answer in the input? Partly at best: the strongest causes are usually a manager relationship or an outside offer, neither of which is in your HR system. Is being wrong acceptable? A false positive here is an employee quietly marked a flight risk on the basis of a model nobody can interrogate. Verdict: feasible and inadvisable. Feasibility is not the only test, and this one fails on grounds this course returns to in module 4.

"AI to write our monthly client reports."

This is generative, so the one-second rule does not apply cleanly. The right question is what the output is for. As a draft that an account manager edits and signs off, it is realistic today and genuinely saves time. As an unreviewed artefact sent to clients, it inherits every hallucination risk from the previous lesson, and the failure lands in front of a customer. Verdict: yes as a drafting tool, no as an automation. The distinction is not about the model's quality. It is about who takes responsibility for the output.

❓ Knowledge Check

A team has built a churn model that accurately identifies customers likely to leave, but leadership complains it "hasn't reduced churn". What has gone wrong?

📚 Flashcards1 / 5

Term

The one-second rule

Click to flip

Definition

If a typical person can do a mental task with under a second of thought, it is probably automatable with machine learning. A fast feasibility filter, popularised by Andrew Ng — not a guarantee.

Click to flip back

💡Key Takeaway

Judge feasibility with two questions a non-specialist can ask: could a person do this in about a second, and do thousands of labelled examples exist? Then check the four failure conditions — thin data, the answer not being in the input, a world that moves faster than the data, and unacceptable errors with no safety net. Finally, keep two pairs apart: deciding is not explaining, and predicting is not causing.