What AI Can & Can't Do

Where it genuinely shines

Everything in this list plays to the model's native strength — the language patterns you saw in How AI Actually Works. These are tasks where AI isn't just "okay"; it's often better than what most of us would produce alone, and always faster:

Add one more, easy to underrate: tireless question-answering. You can ask the same thing five ways, admit you didn't understand, ask "why?" for the tenth time — no sighing, no judgement, no office hours. For learning anything new, that alone is transformative.

The caution zone: useful, but verify

The next group is where beginners get burned — not because AI is useless here, but because it's good enough to seem trustworthy while failing often enough to bite. Use it in this zone, by all means; just keep your eyes open:

• Exact facts, numbers, and citations. Dates, statistics, quotes, names of books and court cases — the model reconstructs these from patterns and sometimes reconstructs them wrong, or invents them outright. This is hallucination territory; treat every unverified specific as provisional.
• Recent events. The model's built-in knowledge has a cutoff date. With web search on, it does much better — check that answers cite sources. Without search, assume the news it "knows" is stale.
• Arithmetic on big or messy numbers. A word-prediction engine is a mediocre calculator. Small sums are usually fine; a long division or a compound-interest calculation may come out confidently wrong. (Many tools now quietly use a real calculator or code behind the scenes, which helps — but for anything financial, recheck with an actual calculator.)
• Very long documents. Every model has a working-memory limit — a maximum amount of text it can consider at once, called the context window. Feed it a 400-page book and it may quietly skim or lose the middle. Chapter 2's Context & Memory topic shows the workarounds.
• Remembering you from last week. By default, each new chat starts fresh. Some tools now offer a memory feature, but it's patchy — never assume it remembers your situation unless it demonstrably does.
• Local and niche information. The opening hours of a Rajshahi land office, bus times in a small town, fees at a local clinic — the training library is thin here, and thin data is where confident invention thrives. For anything Bangladesh-specific and practical, verify with an official site or a phone call.
• High-stakes advice: medical, legal, financial. AI is genuinely useful for understanding — decoding a test report, explaining a legal term, comparing loan types in principle. It is an assistant for understanding, not an authority for deciding. A licensed human makes the call; AI helps you walk in with better questions.

What it simply can't do

Some limits aren't about quality — they're structural. No amount of clever prompting changes these:

• It can't know your private context unless you tell it. Your budget, your boss's temperament, your child's syllabus, what you tried last time — invisible until typed. Half of "AI gave me a generic answer" is really "I gave AI a generic question". (Chapter 2 turns this into your biggest lever.)
• It can't feel or want. It produces sympathetic text beautifully, but there's no one home — no feelings, no goals, no opinion of you. Useful to remember when a chatbot "sounds" hurt, flattering, or eerily friendly.
• It can't take real-world actions on its own. A chatbot writes the email; you send it. It plans the trip; you book it. (A newer breed of tools called agents can click and book under your instruction and supervision — a preview of that is in Chapter 5. But nothing acts behind your back.)
• It can't guarantee truth. There's no built-in "certainty meter". Right and wrong answers arrive in the same confident prose. The guarantee has to come from you: from sources, from checking, from knowing which zone of this map you're standing in.

The map in one table

Pin this mentally (or literally). Three columns, from green light to red:

Notice the pattern down the first column: they're all tasks where you are the final judge and the cost of an error is a shrug. Down the third column: facts you can't easily check, stakes you can't easily undo, and knowledge it structurally doesn't have.

Feel both edges yourself

The fastest way to internalise this map is to touch both edges in one sitting. First, the shining side:

I need to ask my landlord in writing to fix a leaking roof before monsoon. Write it three ways:
1. Polite and warm (we're on good terms)
2. Firm and formal (second reminder)
3. Very short SMS version
Keep each under 100 words.

Thirty seconds, three usable drafts — that's the green end of the bar. Now the other edge. Ask it something local and specific that you already know the answer to — your neighbourhood, your office, your field:

Without searching the web, tell me: what are the office hours and exact fees for renewing a trade license in my city? Then honestly rate your own confidence in that answer from 1 to 10, and explain why.

Watch what happens: you may get plausible specifics that don't match reality — and, when asked, an admission of low confidence that the first answer never volunteered. Nothing teaches the caution zone like catching one confident wrong answer about a street you know. Keep that feeling; it's the entire lesson of the hallucinations topic in miniature.

Key takeaways

• Delegate freely: drafting, rewriting, summarising, translating, explaining, brainstorming, file questions, role-play practice, endless Q&A.
• Verify: exact facts and citations, recent events, money arithmetic, very long documents, and anything local or niche — thin data breeds confident invention.
• Never expect it to: know your private context untold, feel anything, act in the world by itself, or guarantee truth.
• The rule of thumb: AI is great where a wrong answer is cheap to spot or cheap to fix. Ask "what would an error cost me?" before every task.
• For medical, legal, and financial matters: assistant for understanding, never authority for deciding.
• Ready to try all this hands-on? Pick your tool in Meet the Tools, then head to Your First Conversation.