Use these systems well and they save hours. Use them for the wrong task and they fail in a way that is unusually hard to spot, because the failure is written in the same confident prose as the successes.
Where they are genuinely strong
Transforming text you provide. Summarising, rewriting for a different audience, changing format, extracting structure from mess. The source is right there in the context; the work is judgement about form. This is the single most reliable category.
Producing a first draft. A blank page is expensive. A mediocre draft you improve is cheap. The model does not need to be as good as you — it needs to be faster than starting cold.
Explaining things at your level. "Explain this error message to someone who has never used Docker" is a task these systems are exceptionally good at.
Working with language mechanics. Translation, tone, grammar, naming things, generating variations.
Code you can immediately test. The feedback loop is what makes this safe: you run it, and it either works or it does not.
Where they are weak
Anything requiring current facts. A model's knowledge stops at its training data. If it has no live access to a source, recent information is guesswork.
Exact recall. Models reconstruct plausible text rather than retrieving stored records. This is why fabricated citations, invented page numbers and near-miss quotations are so common — the shape is right and the details are invented.
Arithmetic and precise counting. Improving, and better when the model can run code — but a system predicting text is not a calculator, and unaided it will produce confident wrong numbers.
Knowing what it does not know. A model has no reliable internal signal for "I am unsure here". Confidence in the writing does not track accuracy. This is the most dangerous property in the whole list.
Accountable decisions. Anything where a person deserves an explanation and a right of appeal — hiring, credit, medical, legal, disciplinary. The problem is not just accuracy; it is that "the model said so" is not a reason.
A test you can apply
Before using AI for a task, ask: how would I know if the output were wrong?
- Immediately obvious — code that fails, a summary of a document you have read, a rewritten paragraph. Low risk. Go ahead.
- Checkable with effort — a factual claim with a source, a calculation you can redo. Fine, if you actually do the check.
- You would not find out until it mattered — a legal detail, a medical dosage, a statistic in a report going to a client. Do not rely on it. Use it to draft, then verify every specific through a real source.
That question is more useful than any list of capabilities, and it keeps working as the systems improve.
The honest summary
These systems are a fast, tireless, occasionally wrong collaborator with an unusually convincing writing style. Treated that way — draft, then verify — they are one of the most useful tools available. Treated as an oracle, they will eventually embarrass you in front of someone who checks.