What it can and cannot do

The hype and the dismissal are both wrong, and in the same way: they assume AI is one thing with one level of ability.

It is not. It is excellent at some things, poor at others, and the boundary is not where people expect it.

Genuinely good at

• Rewriting text you give it — shortening, changing tone, translating, making it clearer
• Summarising something it can see — a document, a transcript, a long email chain
• Extracting structure from mess — pulling names, dates and amounts out of unstructured text
• First drafts of anything conventional — cover letters, boilerplate code, polite emails
• Explaining things at a level you choose — "explain it as if I am twelve" works surprisingly well
• Transforming between formats — prose to table, notes to checklist, code to comments

Notice the pattern: almost everything it is good at involves text you provided.

Genuinely bad at

• Knowing facts it was not trained on — and it will not tell you when this happens
• Anything after its training cutoff — it may not know recent events at all
• Arithmetic beyond small numbers — it predicts plausible digits rather than calculating
• Counting — ask how many Rs are in a word and watch it struggle
• Knowing what it does not know, which is the root cause of everything above
• Being consistently right in the same way twice — ask twice and you can get different answers

The boundary, stated plainly

The model is a transformer of text, not a source of truth.

It is not a search engine, a database, or an expert. It is a machine for producing plausible text — and plausibility is a very high bar to clear, which is exactly why the failures are so easy to miss.

The three questions before delegating anything

1. Could I check the answer if I had to? If not, do not delegate it. 2. Does it need facts I did not give it? If yes, expect invention. 3. Does a plausible wrong answer cost me anything? If yes, treat the output as a draft.

Those three questions catch nearly every case where AI produces something that looks fine and is not.

The disappointment, explained

Most people who are let down by AI were let down in one of two ways.

They asked it to know something — and got a confident invention. Or they asked it to do something it is good at, did not check the result, and discovered the error later.

Both are avoidable by knowing which half of the list you are in.

Where to go next

• How to check what it tells you
• Try it today — three ways, none of which cost anything