When AI helps and when it doesn't: realistic expectations

People quit AI for one of two opposite reasons: it disappointed them once, or they believed it about everything. Both mistakes come from the same missing piece — a realistic map of where AI is strong and where it trips. Here’s that map.

The one-sentence model

AI is a brilliant, well-read assistant with no hands, no calendar, and no shame about guessing. It has read an enormous amount of text, writes beautifully, never gets tired — and occasionally states nonsense with total confidence. Hold both facts at once and you’re set.

Where it genuinely helps

• First drafts of anything — emails, invitations, posts, speeches. You edit a good draft faster than you write a bad one.
• Understanding — explanations, documents, “what does this letter from the tax office mean”.
• Structure from chaos — turning your voice notes into a plan, your messy list into a table.
• Practice partner — interview prep, language practice, “poke holes in my argument”.
• Translation and tone — saying the right thing in the wrong language.

Where it trips

• Facts with stakes. Laws, medicine, taxes, official requirements. It can be right — and it can produce a beautifully formatted legal fantasy.
• Very recent events. Its knowledge has a cutoff date; yesterday’s news may simply not be in there.
• Exact numbers and counting. It’s brilliant at “about how much” and shaky at “exactly how many”.
• Your private reality. It doesn’t know your boss, your kitchen or your kid’s school. Feed it context or get generic answers.

The failure modes, named

“Sometimes it’s wrong” is too vague to work with. These are the specific ways it goes wrong — once you know their names, you spot them in seconds:

• Hallucination — a confidently stated fact that doesn’t exist. Not lying: the model writes the most plausible next words, and sometimes the most plausible thing is invented.
• Invented sources — a citation, a URL or a book title that looks perfect and leads nowhere. Always click one.
• Sycophancy — it agrees with you. If you ask “isn’t it true that…”, it tends to say yes, because agreement is the plausible continuation. Ask neutrally and you’ll get a more honest answer.
• Stale knowledge — training has a cutoff. Anything after it is guesswork unless the tool searches the web for you.
• Arithmetic and counting — summing a column, counting words in a file, exact dates. It estimates; it does not calculate. Give it a calculator or do it yourself.
• Lost context — in a long conversation it quietly loses the beginning. Details you stated early stop being honoured.

One caveat sits above all of them: confidence is not evidence. The tone is identical whether the answer is right or wrong — which is what makes every failure mode on this list easy to miss.

Hallucinations, in everyday terms

Everyday version: it’s the coworker who never says “I don’t know” and always sounds sure. Useful to have around. Not useful as your only source.

Press “send next” once and read the first answer on its own: case names, case numbers, a guideline, a confident short summary at the end. Now send the follow-up and watch it fold. Same assistant, same calm tone, opposite answer. That is the whole problem in two messages — the tone is identical whether the answer is right or wrong, so the tone tells you nothing. Asking “are you sure? show me the source” is the cheapest check there is, and it works even when the fabrication looks this professional.

The mechanic matters because it predicts when it happens: the model is filling in gaps. Ask about something well-documented and it’s usually fine. Ask about something obscure, brand new, or hyper-specific — a case number, a phone number, a clause in a local law — and you’ve created exactly the gap it will cheerfully fill.

How much to trust it, by task

Two expectations, one conversation

Notice the shape of a good answer: useful knowledge plus an honest warning. You should expect that shape — and distrust any answer that sounds final about high-stakes facts.

Verification habits (lightweight, not paranoid)

• Ask for sources, then actually open one. If it cites a page that doesn’t exist — that’s your hallucination alarm.
• The two-source rule for anything with stakes: money, health, law — confirm in a second place (an official site beats another AI).
• Ask “how sure are you, and what could make this wrong?” In my experience, models are often more honest when asked directly.
• Don’t lead the witness. “Is X true?” invites sycophancy. “What’s the evidence for and against X?” doesn’t.
• Reverse-check the important stuff: paste its summary of your document back and ask “what did I leave out?”

Steal this: the verification prompt

Answer my question: [your question]

Rules: 1) Mark clearly which parts are solid knowledge and which are your
guesses. 2) Tell me the date your knowledge might be outdated from. 3) List
the 2 most likely ways your answer could be wrong. 4) Say where I should
verify this myself. 5) Do not just agree with me — tell me if my question
assumes something false.

Rule 5 is the one people forget, and it’s the one that defeats sycophancy.

The takeaway

• AI is a brilliant generalist with no hands and occasional confident nonsense — use it for drafts, understanding, structure; verify facts with stakes.
• It fails in named, predictable ways: hallucination, invented sources, sycophancy, stale knowledge, arithmetic, lost context.
• The verification habit costs almost nothing. Not having it costs trust in the whole tool.

This closes the AI Basics track. You have an assistant, everyday skills, maybe even a first agent — and, most importantly, correct expectations. The deeper tracks start where this honesty ends. If you want to keep going: the mindset behind all of this is in The AI Mindset, and picking the right tool for the job is in Which AI assistant to pick.