What AI can and cannot do for you right now
An honest capability map, so you stop trying to make it do the things it is bad at.
Before you start
- Nothing — this is the first thing to read
What you will be able to do
- Judge in advance whether a task suits a language model
- Recognise the three failure modes before they cost you an afternoon
- Stop blaming the prompt for problems the prompt cannot fix
The gap between what AI is sold as and what it does is wide enough that most people form their opinion of it on the first bad answer. That is unfortunate, because the failures are predictable and the successes are repeatable.
This is a map, not a pitch. Read it once and you will waste much less time.
Where it is genuinely strong
Transformation, not creation. Anything where you supply the substance.
Models are reliably good at transforming text you already have: summarising, reformatting, translating, changing register, extracting structure from prose, drafting a first version from your bullet points.
The common thread is that you brought the facts. The model is rearranging them, which is a task where being confidently wrong is much harder.
Second-strongest is explanation: asking what an error message means, what a clause implies, how a concept works. Here it is drawing on genuinely common knowledge, and you can usually tell when the answer is off.
Where it is unreliable but still useful
Anything factual and specific. Useful as a first pass, never as a source.
Specific facts — dates, figures, citations, prices, who said what — are where models fabricate. Not occasionally: routinely, and in the same confident register as everything else.
That does not make the task worthless. Using a model to find candidates you then verify is fine and often fast. Using it as the source of record is not.
The rule that survives contact with reality: if being wrong about it would matter, you check it. If checking it takes longer than doing it yourself, do it yourself.
- Asking for sources and treating the ones it produces as verified. Fabricated citations look exactly like real ones.
Where it does not work
Anything needing current information it was never given, or real accountability.
Three categories fail outright. Anything after the training cutoff, unless the tool searches the web and shows you what it found. Anything about your private context — your codebase, your numbers, your customers — that you have not supplied in the conversation. And anything where someone needs to be accountable for the answer: medical, legal and financial decisions have a named human on them for reasons that a model does not change.
Tip The test to apply before you start
Could a smart stranger do this with only what you are about to paste?
Imagine handing the task to a capable person who knows nothing about your situation and has no internet access, giving them only the text you are about to send.
If they could produce something useful, the model probably can too. If they would have to ask you four questions first, send the answers to those four questions with the request. If they would have to look something up, so does the model — and it will not tell you that it could not.
Pick tasks from the first list for a month before you go anywhere near the third. Competence with the easy cases is what tells you whether a hard case is worth attempting.
Common questions
Was this guide useful?
90% of readers found this useful
Read next
Choose your first AI assistant without overthinking it
Every comparison table lists twenty differences and only three of them change your day. Here is how to pick in ten minutes and get…
Set up a daily AI habit that survives week two
The problem is almost never the tool. It is that nothing in your day reminds you to use it, so you default back to the way that al…
Write ad variants and test them properly
AI removes the cost of writing variants, which makes it very easy to run tests that cannot teach you anything. The discipline is i…
Connect two apps with an AI step in the middle
The genuinely useful automations are not the clever ones. They are a trigger, one AI step that makes a small judgement, and a writ…
