McCloskey.ai · ResourcesAll resources
← Resources Claude Academy Courses

AI Capabilities and Limitations

Written by Sean McCloskey, AI tutor. Taught one to one, written down here.

AI is not uniformly capable or uniformly unreliable. It is strong and weak along specific, predictable axes.

I teach the words before the tools: hallucination, sycophancy, automation bias. People stop being surprised once they have the names.

Four properties of generative AI
🔮Next token prediction where answers come from
📚Knowledge what it actually knows
🧠Working memory what it is attending to now
🎛️Steerability how much you are in control

Based on the AI capabilities and limitations course by the Anthropic team. Thirteen lessons and a quiz, listed at 3.5 hours.

AI Fluency teaches the human side, the four Ds. This page teaches the machine properties those four Ds are responding to: AI Fluency

Next token prediction

Where the answers actually come from

The model is not looking things up. It composes the answer word by word, based on what tends to follow what.
Closer to an extraordinarily sophisticated autocomplete than to a search engine.

  • Is this a well-worn path, or out near an edge?
  • How precise is the claim: a name, a date, a figure, a citation?
  • Would you notice if it were wrong?

A confident tone does not signal accuracy. Specificity is where fabrication concentrates, so check the names, dates, statistics, citations and URLs.

Knowledge

What it actually knows, and what it cannot

It knows what it read in training, and only that.
Training ends on a date, and everything after that moment simply is not there.

  • How well represented was this in what it read?
  • Is the answer time sensitive: current events, who holds a job, what something costs?
  • Can it tell you where that came from?

Verify anything time sensitive, and test before you trust in a new domain. Brilliance in one area does not transfer to the one next door. Search and retrieval tools are there specifically to patch these gaps.

Working memory

What it is paying attention to right now

Everything sits inside one fixed-size workspace called the context window.
When it overflows, the oldest material usually falls off, and usually in silence.

  • Has this conversation run long, and has the quality started to slip?
  • Is the part that matters buried in the middle of a long document?
  • Are you expecting it to remember a session it never saw?

This one has a cliff. Lead with what matters, chunk long work into passes, and start fresh when quality degrades. A new chat with a short summary can outperform pushing through.

Steerability

How much control your instructions give you

Specify a role, a tone, a format, a word limit, and the model applies them, often on the first try.
But steerability is not the same thing as understanding.

  • How much room is there between what you typed and what you actually want?
  • Is the instruction short, concrete and checkable?
  • Could a small error in step two carry quietly through steps three and four?

It honours what you said and misses what you meant. When that happens, restate the goal rather than the instruction. Short and checkable beats long and ambiguous.

02 · When they collide

Name the two properties before you touch the prompt

Most real surprises are not one property failing. They are two meeting at once, and naming which two makes the fix obvious.

  1. 1

    Say what was actually off. Not “that was not quite right”, but which part, and how.

  2. 2

    A citation that does not exist? Prediction meeting a knowledge gap. Verify the specifics, or use a tool that grounds them in real sources.

  3. 3

    Your constraints ignored twenty messages in? Working memory meeting steerability. Resupply the essentials, or start fresh with them up front.

  4. 4

    Only then rewrite the prompt. Jump straight there and you are guessing.

Ask it before you start too: which properties am I looking at here? More of the same triage sits in the FAQ

03 · Calibrated trust

Place the task, then decide

Safe handoff

a well-worn path
  • Summarising, reformatting, explaining.
  • Mainstream topics, before the cutoff.

Check it

thin ground
  • Niche, recent or local.
  • Any precise name, date or figure.

Bring your own

outside its reach
  • Your documents, your numbers.
  • Upload it, or give it a search tool.

The goal is not to distrust AI, and not to hand it everything. It is calibrated trust, neither granting it nor withholding it wholesale.

What you learned

This week: take one answer that annoyed you and name which two properties caused it.

Where next

The human half of the same picture: AI Fluency Close the gap between what you type and what you mean: How to Prompt

Do it with someone. Everything here is free and you can run it alone. If you would rather work through it with me, one to one, book a free 30-minute chat. No pitch.