[Beginner] — the last note. Not about how the tool works, but about how to use it over years without it quietly making you worse at things.

The kid version first

A calculator made arithmetic cheap. That was mostly good — but the people who kept a feel for roughly what the answer should be are the ones who still notice when they’ve typed a wrong number.

Same deal here. The goal isn’t to use AI for as much as possible. It’s to use it for the right things, while keeping the judgement that lets you tell when it’s gone wrong.

Where it genuinely earns its place

Sorted by how reliably it pays off:

TaskWhy it works
Beating the blank pageA mediocre draft to react to beats staring at nothing. Reacting is easier than creating
Explaining something you don’t understandAsk five times, at five levels, with no embarrassment. Historically a tutor’s job
Changing the shape of text you haveSummarize, restructure, shorten, reformat, change register — grounded in your material, so low risk
Being a first reader”What’s unclear here? What would a sceptical reader push back on?”
Getting unstuckTwenty options in ten seconds; you only need one to be interesting
Translation, with the reasoning shownAsk why it chose a phrasing and you learn something
The tedious middleReformatting a list, extracting fields, tidying notes

Notice the pattern: it’s strongest when you supply the substance and it supplies the shape, and weakest when you ask it to supply substance you can’t check.

Where it quietly costs you

Skill atrophy is real, and it’s selective. Offloading things you’ve already mastered is fine — that’s leverage. Offloading things you’re still learning is where the cost lands, because the struggle you’re skipping is the mechanism by which the skill forms. A student who has an LLM write every essay gets essays and no writing ability. Someone who already writes well and uses it to draft faster loses nothing.

The rule: delegate what you’ve learned, not what you’re learning.

Homogenisation. It converges on the median — the average phrasing, the safe structure, the four-bullet answer with a summary. Fine for a status update, corrosive for anything meant to sound like you. If you use it on your own writing, use it to edit, not to voice.

Anchoring. The first draft it gives you shapes everything you think afterwards. Sometimes worth forming your own view first, then asking — especially on decisions.

Fluency mistaken for understanding. A clear explanation feels like comprehension. It isn’t. The test is whether you can reproduce it, apply it, or teach it without the chat open. Ask it to quiz you rather than explain again.

Be honest about using it

Norms are still forming, but a few are settling:

  • Editing your own work with it — no disclosure expected, same as a spellchecker.
  • Substantially generated content presented as your own — disclose. Academic work, journalism, and most professional contexts increasingly require it, and being caught not disclosing costs far more than disclosing.
  • Anything sent to another person as your judgement — a recommendation, a diagnosis, a review — you own it fully. Read every word before it goes.

The thing that consistently ends badly is passing on output you didn’t read or check. Fake legal citations, invented statistics in reports, code nobody understood. In every case the tool wasn’t the failure — sending unread output was.

Watch what happens to you

Two questions, occasionally:

Am I still able to do this without it? If a skill you rely on professionally has gone soft, do it manually sometimes. Deliberately.

Am I using it because it’s better, or because it’s easier? Both are valid answers — just know which one you’re giving.

Where to go next

You now have the working model: what it is, how it works, how to brief it, how to feed it, how to check it, and what not to give it. From here:

  • Just use it. Two weeks of real tasks teaches more than another course. Come back to note 06 when something surprises you.
  • Want it to run without you? Repeated, scheduled, unattended tasks are automation — ai-automation/.
  • Want to build things with it? The AI Engineer track picks up exactly where this leaves off: calling models from code, retrieval, tools, agents, evals. It assumes you can program.
  • Want to know how the models themselves are made? The AI/ML course — foundations, then the data scientist and ML engineer paths.

Key insight

The tool is very good at producing the shape of competent work. Whether it contains competent work is your job, and stays your job. Use it to go faster at things you could do slowly — and be deliberate about the skills you’re choosing not to build.

↑ Back to Using AI