There's a line King wrote in two separate books that I keep pondering about how it relates to a common problem with AI chatbots.
Depending on which version you've encountered, in The Shining, it goes something like this:
*The arguments against insanity fall through with a soft shirring sound. Layer on layer.*
He was writing about a man losing his grip on reality inside a hotel that had decided to help him along.
A similar quote also appears in Lisey’s Story, in which Lisey also feels she is taking crazy pills:
“The arguments against insanity fall through with a soft shirring sound; these are the sounds of dead voices on dead records floating down the broken shaft of memory.”
I keep thinking about it every time I watch a long AI conversation slowly eat itself.
Context collapse in AI is quiet, a thing people often miss.
There's no jump scare, but more of a slow drift (a soft shirring…)
You start a long chat session with a clear set of instructions: write in this tone, remember this constraint, keep this persona, no em-dashes. The chatbot eagerly nods along, like a good little sycophant.
But, thirty exchanges later, it has stopped following the rules and you have found yourself cursing at an emotionless metal box. Earlier instructions have drifted outside its effective attention, and newer inputs have filled the space like blood rushing out of an elevator into a hallway.
An even more infuriating part of this is that the tool doesn't know it forgot.
Jack Torrance doesn't lose his mind all at once because that would be too clean, and it would be a boring story.
He loses it conversationally, incrementally.
The hotel feeds him inputs he can't verify, and he starts weighing the recent, vivid, emotionally charged ones over the older, quieter signal of reality.
His anchor (his family, his sobriety, his own prior self) falls outside his effective attention.
“The arguments against insanity fall through with a soft shirring. Layer on layer.”
An LLM (think GPT-5.4 or gemini-3.1) processing a 40,000-token conversation is doing something structurally analogous. (if you aren’t familiar with tokens, to give you an idea of length, the entire seven-book Harry Potter series is equal to about 1 million tokens)
The model's attention has to work across everything you've exchanged. The earlier material (the rules you set, the context you carefully established, the constraints you thought were load-bearing) gets progressively demoted as newer tokens crowd the window. The model can't attend to everything equally, and something has to give.
What gives is always the old stuff, which is typically the foundational stuff you thought would stay constant.
Bigger context window = solved problem, right?
The research says otherwise.
Models can become less reliable with significantly larger context windows — attending properly to all that data is genuinely hard, and the model's confidence stays high while its grip on the material quietly loosens (...soft shirring).
Every time you compress a long conversation into a new prompt, you lose nuance.
Do it enough times and you're rewriting a document until the original notes have vanished. A polished version of a thing that has drifted pretty far from where you started.
The goal is to orchestrate your sessions so drift hurts less when it happens. And if you use these tools enough, it will happen.
King understood something that applies here: the scariest part of the collapse is the overconfidence that things are going along just fine…not the collapse itself.
Jack doesn't spiral into insanity knowing he's spiraling. He arrives there feeling certain he is right.
A chatbot won’t flag its uncertainty in the midst of a mid-context drift. It will keep answering you with confidence as it slowly drifts away from your goal.
That's on us (the humans) to catch.
So, step outside the hotel. Start a new session. Re-anchor. Reread the original notes before they're gone.
You wrote the thing. You remember what it meant.