What is a context window, in human terms?

Every AI model has one. It turns out you do too — and you can measure it in two minutes.

The short version

A context window is how much text an AI model can hold in mind at once while it works: the conversation so far, the document you pasted, its own answer. Anything outside the window, the model simply cannot see. It's measured in tokens — chunks of text roughly three-quarters of a word long, so 1,000 tokens is about 750 English words.

The numbers have grown fast. GPT-2 (2019) held about 1,000 tokens — a page. GPT-4 launched with 8,000. Today's models advertise 128,000 to over a million, and one claims ten million: whole libraries in mind at once.

Do humans have one?

Not in the same shape, but yes. Psychologists have measured working memory for a century: you can hold roughly four to seven separate things at once (digit span, Cowan's "magical number four"), and when you read a passage once, you retain a limited number of its facts — the basis of the Logical Memory test in the Wechsler Memory Scale, still used in clinics.

The honest human analogue to a context window is therefore not "how many words can you recite" but how long a passage can you read once and still answer questions about. That's the same needle-in-a-haystack idea the AI labs use — run on a person.

How the test on this site measures yours

1. A short passage streams past, a couple of words at a time, at reading speed. No pausing, no re-reading — nobody can stare longer than anyone else.
2. It disappears and you answer a few multiple-choice questions about its facts: who, where, how many, what colour.
3. Get them right and the next passage is longer (40 → 100 → 200 → 350 → 500 words). Miss, and the test stops.
4. Your result is the longest passage you could still answer about, tokenized with the same tokenizer OpenAI's models use, and placed on a leaderboard of AI models by how much of the ladder you got right.

Harder modes don't make it harder; they make it more precise — more rungs, more questions, and a distraction step between reading and answering so you can't just repeat the words to yourself.

What the number means

Most people land between 40 and 200 words. That is not a verdict on intelligence — a model holding a million tokens can't tie its shoelaces. It's a snapshot of one specific skill, working memory for text, which is also one of the most trainable. The test is the first step; the second is training it.

Based on: needle-in-a-haystack and RULER benchmarks; WMS Logical Memory; complex-span working-memory tasks (Unsworth, Engle). See also: AI context windows compared.