What AI can and cannot do — an honest map

You now hold four ideas: AI learns patterns from examples (L3), stored as tuned dials (L4), used to generate plausible continuations (L5–6), with no built-in truth check (L7). Those four ideas are a prediction machine of your own — point them at any task and they’ll tell you whether AI will shine or stumble. Let’s draw the map.

Strong zone — pattern work you can check

Tasks that are language-shaped, pattern-rich in training data, and easy for you to verify:

• Drafting and rewriting: emails, summaries, outlines, translations, tone changes (“make this friendlier”)
• Explaining: “explain this contract clause / error message / concept simply”
• Brainstorming and structuring: options, plans, counter-arguments, names
• Transforming what you provide: notes → minutes, table → prose, long → short

Why strong? Oceans of training examples, and the output sits in front of you where wrongness is visible.

Shaky zone — usable with verification

• Factual questions: often right, invisibly wrong (L7) — verify specifics
• Maths and step-by-step logic: improving, still slips; check the steps
• Anything recent: pattern memory has a horizon; needs an attached search tool
• Long tasks with many constraints: constraints quietly drop; re-check against your list

Weak zone — keep humans firmly in charge

• Unverifiable specifics: citations, statistics, quotes produced from pattern memory
• Accountable judgement: hiring, diagnosis, legal strategy, grading people — wherever an error lands on a human being and someone must answer for it
• Knowing its own limits: the model can’t reliably tell you when it’s out of its depth — that assessment is your job, permanently

The golden rule

Use AI where you can verify; keep a human in the loop where errors cost. Every workflow in the rest of this course is built on that sentence.

Try it now (3 minutes)

Take three real tasks from your week. Place each on the map. For anything outside the strong zone, name the verification step you’d add. You’ve just done your first piece of AI workflow design — the skill the whole Explorer track builds.

Check your understanding

1. Which task sits most safely in AI's strong zone?

Producing exact market-size statistics from memory Rewriting your rough notes into a clear email you'll read before sending Deciding which job applicant to reject

2. Why does 'keep a human in the loop where errors cost' follow from this module?

Because regulations require it everywhere Because the model generates plausibility without a truth check, so stakes determine how much human verification is needed Because AI refuses high-stakes tasks

Check answers

Recap

Strong: verifiable language and pattern work. Shaky: facts, logic, recency — verify. Weak: unverifiable specifics and accountable judgement — human in charge. Module E1 complete. Next module: meeting the assistants themselves, and putting this judgement to work.