What AI Is Actually Bad At

Most advice about AI is a list of things it is good at. That list is easy to find and gets stale fast. The more durable skill is knowing the shape of its failures — because the failures are consistent, and once you can name them you stop being surprised.

1. It is confident when it is wrong

This is the big one, and it is not a bug that is about to be fixed. These systems produce fluent text whether or not the underlying claim is true, and fluency is the signal humans use to judge competence. A wrong answer arrives in the same calm, well-organised prose as a right one.

Practical consequence: tone tells you nothing. You cannot use "it sounded sure" as evidence. Every factual claim you plan to act on needs a source you checked yourself.

2. It invents specifics

Names, dates, citations, case numbers, API methods, statistics. The more precise-looking the detail, the more worth checking it is — precision is exactly what these systems are good at imitating and bad at guaranteeing.

3. It does not know what it does not know

Ask about something obscure and you rarely get "I have no idea." You get a plausible answer assembled from adjacent things. The absence of hedging is not evidence of knowledge.

4. It is bad at counting and precise arithmetic

Not always — many tools now run actual code for maths — but the underlying model does not compute, it predicts. If a number matters, either check it or make sure the tool ran real arithmetic rather than producing a number-shaped word.

5. It flattens toward the average

The default output is the middle of everything written on a subject. That is genuinely useful when you want the consensus view and useless when you want a position. Ask for an argument and you often get a balanced survey of arguments instead.

This shows up in writing as a house style you did not choose: measured, slightly padded, fond of "furthermore" and three-item lists. Fixable, but only if you notice it.

6. It agrees with you too easily

Push back on a correct answer and it will often fold. That makes it a poor tool for testing whether you are right about something, unless you deliberately ask it to argue the other side — and even then, treat the result as a checklist of objections rather than a verdict.

What to do with this

1. 01Use it freely for things you can verify at a glance — drafting, rewording, summarising something you already have.
2. 02Slow down when the output contains facts you cannot check yourself, and go find the source.
3. 03Do not use it as the only reviewer of a decision. It is a second draft, not a second opinion.
4. 04Notice when you have stopped reading the output carefully. That is the actual risk.

The danger is not that the tool is wrong. It is that it is right often enough that you stop checking.

Coursium spends real time on where these tools break, because that is the part that keeps being useful after the models change.