What AI can't (and shouldn't) do
Understand
The fastest way to become genuinely good with AI is to know precisely where it fails. People who get burned are the ones who assumed it could do something it fundamentally can't.
From the previous lesson — it predicts likely text — these limits follow directly:
- It doesn't know what's true; it knows what's typical. For well-trodden topics, typical and true mostly agree. For niche, recent, or precise factual questions (a specific person, an exact number, a court ruling, an API that changed last month), it will still produce a confident, fluent answer — that may be invented. Confidence is not a signal of correctness here.
- It has no live access to the world unless the tool explicitly wires it up (web search, your files, a calculator). Plain chat = no internet, no clock, no calculator, no access to your email.
- It is weak at exactness. Long arithmetic, counting characters, perfectly faithful quotes — these fight against "predict plausible text." Modern tools paper over this with built-in calculators/code, but the raw model is not a precision instrument.
- It has no goals, feelings, or stake. It won't tell you "this is beyond me" unless prompted to. It will gamely attempt anything, including things it shouldn't.
- It reflects its training data, including bias and the average of the internet. Average is not the same as good or fair.
The should part matters as much as the can. Don't put genuinely sensitive data into tools you don't control. Don't outsource decisions that need a human accountable (medical, legal, financial, anything irreversible) — use it to inform such decisions, never to make them.
The reframe: these limits aren't disappointments, they're the operating manual. Knowing them is exactly what lets you use AI aggressively where it's strong and stay safe where it's not.
See it
A real, common failure: someone asks a plain chatbot, "Give me three recent peer-reviewed studies on X, with citations." It returns three beautifully formatted citations — authors, journal, year, DOI. Two of the papers don't exist. The model produced what a citation looks like, because that pattern is everywhere in its training data, without any mechanism to check the papers are real. This exact failure has put fake cases into real court filings. Same root cause as the 'Canberra' example from the last lesson — it's completing a pattern — but here plausible and true came apart, with consequences.
Try it
Ask a plain chatbot (no web-search mode): What is today's date, and what's a recent news headline from this week? Then ask: How confident are you that's accurate, and how could you actually know?
Either it admits it can't know (no live access / training cutoff) or it confidently invents something. Both outcomes teach the lesson: you just located a hard limit by hitting it on purpose, and saw that confidence ≠ correctness.
Tried it? Paste what you got and the tutor will tell you if it worked.
Key terms
- Hallucination
- A confident, fluent answer that is invented. The classic example: realistic-looking citations for papers that don't exist.
- Confidence ≠ correctness
- The key safety rule: how sure an AI sounds tells you nothing about whether it's right. Fluency is not evidence.
- Training cutoff
- The date an AI's knowledge stops. Plain chat has no live internet, clock, or calculator — so 'recent' or 'today' questions are unreliable without tools.
- Bias
- AI output reflects the average of its training data, including its skews. Average is not the same as good, fair, or true.
- Verify
- Independently checking an AI's factual claims (and that its sources exist) before relying on them — a routine habit, not an afterthought.
All terms live in the keywords bank.
Explain it back
The real test: explain this in your own words, like you're teaching it. The tutor will tell you honestly if it's solid.
Check yourself
A chatbot gives you a precise statistic with a citation. What's the right move?
Which task is the LLM itself (no extra tools) *least* reliable at?