AI FoundationsStarter2:07

Why AI Stops Following Your Instructions

Prefer to read?

Why does AI sometimes stop following your instructions?

You ask AI to draft an email, and then you revise it ten times. You change the tone and ask it to remove a point. It does, but a few drafts later, that same point is back.

One reason is how it uses its context window.

So, what is a context window?

It is all the information AI can see when it replies. Think of it as AI's desk. It can only work with what is on the desk right now. That includes your conversation, the files you share, and instructions from the app that you never see. The size of this desk is measured in tokens, which are small pieces of words.

Earlier models could hold a few thousand tokens, about a few pages of text. Some can now hold a million or more, which is like several books at once.

Does that mean AI uses everything on the desk correctly?

No. Mistakes can start long before the desk is full, and there is no fixed point where it happens. As the conversation grows, AI can get worse at using it. This is called context rot.

Why does that happen?

Your email chat now holds many drafts, rejected ideas, and instructions to be both formal and casual. AI has to work out which parts still matter, and old or conflicting information makes that harder.

Can't you just tell it to forget the old instructions?

Saying "forget that" only adds another note to the desk. It does not remove the old one. So both the old instruction and your request to ignore it are still there.

Then why don't companies just make the desk bigger?

They are. But a bigger desk also leaves room for more clutter, and AI still has to find what matters in all of it.

So what can you do about it?

If AI keeps mixing old requirements with new ones, start a fresh chat. Bring a short, checked summary of your goal, your decisions so far, and your latest draft.

What to remember

A longer chat does not make AI smarter. Old drafts and changed instructions stay on the desk, even after you say "forget that." When the answers start mixing up, a clean desk with a short summary works better than one more correction.

Why it matters: Every other skill rests on this: knowing how AI reaches an answer and when to check it.

Keep going