Context rot is the degradation of an AI model's output as the context it is working with grows.
The individual facts remain present. The relationships between them weaken, fragment and drift, so the
model keeps producing fluent, confident answers that are quietly wrong. It is the failure mode that most
often ruins otherwise good AI work, and it arrives far sooner than the advertised context window suggests.
Real business information is not a list of facts, it is a mesh of connected ones: sources, entities,
constraints, causes, timelines, evidence. As context grows, the connections weaken, fragment and drift.
What actually breaks. Not the facts, the connections between them.
How Fast
It happens sooner than you have been told.
MRCR-8 is the closest approximation of that mesh, and the results are sobering. At 16k tokens, roughly 20
pages of input, the leading models are already shedding accuracy fast. This despite 1M context
windows. Non-thinking modes fall off a cliff.
Bigger context windows do not fix this and are actively deceptive about it, because more
irrelevant context makes it worse rather than better. The number on the box is a capacity figure, not a
quality guarantee.
This is why "just paste the whole document in" stops working exactly when the work starts to matter, and
why the failure is so hard to catch: the output stays fluent long after it stops being reliable.
How fast it happens. Measured across the current frontier models.
The Diagnostic
Seven ways context breaks
Context rot is not one failure, it is seven. Each produces a distinct business symptom, and each has a
specific action that prevents it. Once you can name the failure mode, you stop blaming the model and
reaching for a different one.
Mode One
Task-frame drift
The model quietly stops answering the question you actually asked and starts answering a nearby one.
Mode Two
Source and version confusion
Two versions of a document are in context and the model blends them into one that never existed.
Mode Three
Entity and metric confusion
Numbers and names attach to the wrong entity. The figure is real, it just belongs to something else.
Mode Four
Aggregate and detail interference
Totals and line items contaminate each other, so the summary stops reconciling with the underlying detail.
Mode Five
Format degradation
The structure you asked for erodes over a long session until the output is no longer usable as given.
Mode Six
Reasoning-chain breakdown
Each step looks sound in isolation, but the chain connecting them has quietly come apart.
Mode Seven
False coherence
The most dangerous one. The answer reads beautifully and is wrong, and nothing in the tone signals it.
The diagnostic table. Seven failure modes, seven symptoms, seven fixes.
The Countermeasure
The fix is a filing habit, not a clever prompt
It sounds administrative. It is the difference between AI work you can hand to a colleague, a client or an
auditor, and AI work nobody, including you, can reconstruct a week later.
Step One
Convert everything to Markdown
A Word file can be a thousand times the size of the same text, and you pay for that out of the context you needed for thinking.
Step Two
One numbered file per stage
Input saved, output saved, and the output becomes the next input. It reads like admin until somebody asks how a conclusion was reached.
Step Three
Ask for a handoff when it drifts
A self-contained summary of every decision taken and option rejected, which you paste into a clean chat. This is the practical answer to context rot.
Structure your prompts, run expert methods, and reset often. Context rot is covered in full on day one of the
GenAI Executive Masterclass, with the diagnostic table above worked through live.
FAQ
Questions people ask about context rot
What is context rot?
Context rot is the degradation of an AI model's output as the context it is working with grows. Real
business information is a mesh of connected facts: sources, entities, constraints, causes, timelines and
evidence. As context grows, the connections between those facts weaken, fragment and drift. The facts are
all still in there. The relationships between them are not.
Do bigger context windows fix context rot?
No. Bigger context windows do not fix context rot and are actively deceptive about it, because more
irrelevant context makes the problem worse rather than better. On MRCR-8, leading models are already
shedding accuracy fast at 16k tokens, roughly 20 pages of input, despite advertising context windows of
1M tokens.
What are the signs of context rot?
Context breaks in seven recognisable ways: task-frame drift, source and version confusion, entity and
metric confusion, aggregate and detail interference, format degradation, reasoning-chain breakdown, and
false coherence. Each produces a distinct business symptom, and naming the failure mode is what stops you
blaming the model and reaching for a different one.
How do you prevent context rot?
The countermeasure is a filing habit, not a clever prompt. Convert source documents to Markdown,
because a Word file can be a thousand times the size of the same text and you pay for that out of the
context you needed. Work in one numbered file per stage, so output is saved and becomes the next input.
When a long chat drifts, ask for a handoff: a self-contained summary of every decision taken and option
rejected, which you paste into a clean chat.
Context rot is day one of the Masterclass.
Seven failure modes, the business symptom each one produces, and the action that prevents it. Worked through
live, on your own material, by the team that has trained 700+ South African executives.
