A ten-minute case for three small words

Ask If
Unclear

Your AI isn't reading your mind.
It's placing a bet.

The thing that keeps happening

YouCan you take care of Gary?
Reading 1

Made him a cup of tea.

Reading 2

Drove him to A&E.

Reading 3

Fed his cat while he's away.

Reading 4

Buried him under the patio.

It didn't misunderstand you. It understood you four ways — and picked one.

What is actually happening in there

It predicts the next
chunk of text.
Then it does that again.

That's the whole trick. There is no plan, no draft, no moment where it decides what the finished answer will be. Each piece is chosen with only the pieces before it for company.

Autocomplete that went to university.

Watch it think · one chunk at a time

YouCan you take care of Gary?
press Predict next…

Illustrative figures. A real model ranks its entire vocabulary — tens of thousands of options — at every step; these are the top four.

chunk 0 of 9

A small correction

They aren't words.
They're bites.

unclearunclear
clarificationclarification
Ask if unclearAsk if unclear

The pieces are called tokens — roughly three quarters of a word each. It reads and writes in bites, not words. Nothing else about the argument changes.

Remember 34 vs 31

A near-tie inside the machine is a coin flip on Gary.

…I've looked after · 34% → Gary gets a cup of tea disposed of · 31% → Gary is under the patio

It doesn't hedge. It commits — and then every chunk after that defends the choice.

Why it never just says "I'm not sure"

Ambiguity doesn't feel like anything to a predictor.

99 / 1 split

One confident sentence.

34 / 31 split

One confident sentence.

Same fluency, same tone, same certainty. The doubt exists for a fraction of a second inside the numbers and then it is gone — spent, not reported. A fluent answer to the wrong question looks exactly like a fluent answer to the right one.

Same prompt · one difference

YouTake care of Gary while I'm at my mother's.
currently: off

Cost of the guess: 15 to 20 years, and you find out on Tuesday.

Why three words move the needle

It's an instruction. And this thing is built to follow instructions.

It isn't detecting confusion and deciding to report it. It's doing as it's told — and two separate forces make that land hard.

Trained to comply

Following an explicit instruction is the behaviour it was trained hardest to produce. Those three words spend that.

Otherwise suppressed

Right after a request, producing the goods is what gets rewarded — asking is rare. The instruction lifts that brake.

The net effect · illustrative, not measured
your prompt
+ ask if unclear

It spots "unclear" by pattern, not by reading its own numbers. Useful — not reliable.

When to spend the twelve seconds

Worth it
  • Anything with a deliverable at the end
  • Anything expensive or slow to redo
  • Multi-step work that builds on step one
  • Whenever you used a vague adjective: better, cleaner, modern, tighter
Skip it
  • Quick lookups and conversions
  • Throwaway drafts you'll rewrite anyway
  • When any reasonable reading is fine by you

Rule of thumb: the more it costs to be wrong, the more the question is worth.

Steal this

Ask if unclear.

Paste it at the end of any request that matters. It costs you one line and it buys you the question the machine would otherwise have answered on your behalf.

click or to advance?notes
Speaker notes