The middle of the meeting

Think about a long meeting. You remember how it opened. You remember the last thing said before everyone stood up. The forty minutes in between go soft.

Models show a similar pattern with long inputs. A fact can sit inside the window, fully visible, and still get missed.

What the study found

In 2023, Liu and colleagues tested this. They gave models long inputs and moved the useful information to different places within them. The models did best when it was near the start or near the end. When it sat in the middle, they used it noticeably worse.

The paper’s title gave the effect its name: lost in the middle.

Hello, AI describes the same thing from the user’s side. In a very long chat, a model gets worse at using details buried in the middle, well before the window is full. A fact on page 2 of a 300-page conversation is technically visible and practically fuzzy.

This matters most when you paste in something long, like a contract or a report, and ask about one detail. The detail may be in there. The model may still answer as if it were missing.

So a big window tells you how much text fits. It tells you less about how well each part gets used.

Habits that help

Hello, AI’s advice follows from how the window works. You are the memory manager for the conversation.

• State anything you need used in plain words, and recently. A shellfish allergy mentioned once, forty messages back, is easy to lose.
• Paste in the text you want the model to work from, and ask it to answer based only on that text. Give it an honest way out, such as saying “not mentioned” when the text is silent.
• When a long chat keeps going wrong, open a fresh one and restate the task with only the good context. Every confused turn stays in the window and pulls on the next answer.

None of these need special words. They just keep the important parts where the model uses them best.

In one sentence: a fact can be inside the window and still get missed, so say what matters plainly and keep it close.