What AI can't do

Two jobs, one machine, opposite answers

Here are two jobs. Job one: sort fifty thousand incoming emails and flag the ones that look like junk. Job two: decide, alone and for good, whether a patient has a rare disease and should start a treatment that carries real risk. The same kind of pattern-prediction machine could be pointed at either. For the first, it’s a fine choice. For the second, handing it the decision would be reckless. The machine didn’t change. The task did. Learning to tell these two apart, before you hand anything over, is the most useful thing this course can leave you with.

Everything it can and can’t do comes from one fact

You’ve now seen the machinery. Pull it into one line: an AI predicts patterns from examples. Every strength and every danger flows straight out of that.

Because it works from patterns, it’s strong wherever a task looks like a mountain of past cases. Because it only predicts, it has no understanding of what it’s doing. It has no goals of its own — it does what it’s pointed at. It has no common sense beyond its data, so a situation that never appeared in the examples finds it flat. It can be fluent and confident and wrong at the same time. It mirrors whatever bias sits in the data it ate. And it can’t hand you a readable reason for its answer.

None of that is a bug to be patched. It’s the shape of a machine that predicts instead of knows. So the question is never “is AI good or bad?” It’s “does this task fit a pattern-predictor, or not?”

Four questions to ask before you hand it a task

Ask these four, in order, of any job before you give it to a machine.

One: is it pattern-rich? Is the task similar to lots of past examples? If yes, that’s exactly what the machine does best. Sorting, flagging, recommending, drafting, transcribing — all pattern-rich.

Two: can you tolerate wrong answers? Is every mistake cheap to catch and fix? If a wrong answer is a minor annoyance you can undo, the machine is safe to lean on. If a wrong answer causes real harm, it isn’t.

Three: is it novel or high-stakes? A one-off situation with no close match in the data falls outside the machine’s patterns — and that’s where its confident wrong answers live. The higher the stakes, the worse a confident wrong answer costs.

Four: does someone need to be accountable? If a person must own the decision and explain why, the machine can’t do that job. It has no readable reason to give.

Two yeses on the first pair and two noes on the second pair means the machine fits. Flip that, and it doesn’t — at least not alone.

Walk the test on the two jobs

Job A: flag which of fifty thousand emails are probably junk. Pattern-rich? Yes — junk looks like past junk. Error-tolerant? Yes — a missed one is a small annoyance, and you can fish it back out of the folder. Novel or high-stakes? No. Accountability? No one needs to explain why an email got flagged. Every arrow points the same way. This fits a machine well. Let it run mostly on its own, with a person glancing at the edge cases.

Job B: decide, finally and alone, whether a patient has a rare disease and starts a risky treatment. Pattern-rich? Rare means few examples — thin patterns, the exact blind spot from earlier. Error-tolerant? No — a confident wrong call harms a person. Novel or high-stakes? As high as it gets. Accountability? A doctor must own the call and explain it. Every arrow points the other way. Do not hand this to a machine alone.

Same technology, opposite verdict, because the two tasks have opposite shapes.

The design that fits: keep a human in the loop

Job B isn’t a job the machine has to sit out. It’s a job where the machine takes the right seat. Let it scan the images and flag the unusual ones for a doctor to look at closely. It’s fast, tireless, and good at spotting a pattern a busy human might skip. Then the doctor — who can weigh the whole person, handle the genuinely new, and answer “why” — makes the call and owns it. The machine narrows the pile. The human decides.

That’s the human-in-the-loop, and it’s not a compromise. It’s the design that matches how the machine actually works: brilliant at narrow the pile, poor at make the final, novel, accountable decision. Use it as an oracle you obey and its blind spots become yours. Use it as a fast assistant you check, and its strengths are yours while a person still catches the wrong ones.

So the honest map: lean on it for high-volume, pattern-rich, error-tolerant work — sorting spam, first-pass scans, recommendations, rough drafts, transcription. Keep a human firmly in charge of anything final, novel, high-stakes, or accountable — a diagnosis, a sentence, a hiring, a loan denied with no one to explain it.

On the whole

You now know the machine well enough to know where to trust it. Not from a slogan, but from the mechanism: it predicts, it doesn’t know. That single fact tells you it will be superb at the pattern-rich and dangerous at the genuinely new, sure-sounding when it’s wrong, and silent when you ask why. You live among these systems already — they sort your mail, suggest what you read, sift the images a doctor sees. Knowing what they are is what lets you take the good they offer and keep the final word, on anything that matters, in human hands.