This piece builds on an earlier essay I wrote, “Why We Think in Fences, Guns and Cats,” which explores why we mix ideas like Chesterton’s Fence, Chekhov’s Gun and Schrödinger’s Cat into a single mental toolbox for thinking about risk and systems. In that post, I deliberately spliced metaphors and sayings together; here, I push that further and turn the cats themselves into a practical literacy layer for generative AI.
Generative AI is surrounded by technical language.
Transformers.
Attention weights.
Latent space.
Probability distributions.
Alignment.
And yet when people try to understand what is happening, they do not reach for equations. They reach for stories.
A cat in a box.
A razor.
A gun on a wall.
A law that says whatever can go wrong will.
That is not accidental.
Large language models are a type of generative AI trained on vast amounts of human text, learning statistical patterns in language rather than explicit rules. Human writing is built on metaphor, so when we reason about AI systems, risk, uncertainty, failure and relevance, we instinctively reach for physical props.
Strangely, those props are often more useful than many whitepapers.
Here are four “cats” that explain generative AI better than most technical summaries.
1. Chesterton’s Cat: Do Not Throw Away the Box
Chesterton’s Fence says:
Do not remove a fence until you understand why it was put up.
Translated into, “Do not throw away the box until you understand what it was protecting.”
This principle is deeply relevant to AI automation.
Across organisations, there is pressure to “remove the box”:
Replace human review with AI.
Automate customer support.
Remove approval layers.
Eliminate documentation because “the model can recall context for us.”
But every legacy process existed for a reason.
Maybe it slowed things down.
Maybe it felt bureaucratic.
Maybe it looked redundant.
Yet it might have been protecting:
Accountability
Risk containment
Legal defensibility
Brand trust
Human judgment in edge cases
When organisations rush to automate without understanding what a process was guarding against, they do not create efficiency. They create invisible risk.
This is where AI governance really begins.
Before replacing a human checkpoint with a model, ask:
What was this box protecting?
If you do not know, you are not modernising. You are gambling.
2. Occam’s Cat: Assume “Cat” Before “Portal”
Occam’s Razor advises us to prefer the simplest explanation that fits the facts.
A cleaner version: If you hear a noise, assume “cat” before “interdimensional portal.”
This is AI literacy in one sentence.
When a model produces a strange output, people often jump to dramatic explanations:
“It is becoming conscious.”
“It is manipulating us.”
“It is secretly intelligent.”
“It is biased on purpose.”
The simpler explanation is almost always structural:
Statistical pattern completion
Prompt ambiguity
Insufficient context
Training data artifacts
Overconfident probability distribution
Large language models operate by predicting the next token based on patterns learned from huge text corpora, not by reasoning from first principles. They can simulate reasoning, but under the hood they are still performing probabilistic pattern completion.
When something strange happens, assume cat.
Assume probability.
Assume training data.
Assume statistical mechanics.
This mindset does two things:
It prevents panic.
It encourages better system design.
AI governance is not about dramatizing risk. It is about diagnosing it calmly.
3. Murphy’s Cat: It Will Fail During the Demo
Murphy’s Law says:
If something can go wrong, it will.
A sharper version: If the box can fail, expect it to fail during the experiment.
Anyone who has run a live AI demo knows this is not theory.
The API times out.
The model hallucinates.
The output contradicts itself.
The agent loops.
The wrong file gets indexed.
The embeddings misfire.
Generative AI systems typically produce a distribution of plausible outputs rather than a single fixed answer for a given input. That variability is a feature, not a bug, but it introduces operational fragility.
The mistake organisations make is assuming AI behaves like traditional software.
Traditional software:
Input A → Output B.
Generative AI:
Input A → Distribution of plausible outputs.
Even when you configure models to minimise randomness, for example with low sampling temperature or greedy decoding, implementation details and infrastructure can still introduce subtle non deterministic behaviour between runs. That difference changes how systems should be designed, tested and governed.
Governance, in this context, is not about preventing failure entirely. It is about designing systems that:
Contain failure
Detect failure
Escalate failure
Log failure
Murphy’s Cat reminds us that resilience must be designed in from the beginning, not after the first public embarrassment.
4. Chekhov’s Cat: What You Emphasise Matters
Chekhov’s Gun states:
If a gun appears in Act I, it must go off by Act III.
Another variation: If the cat gets dramatic lighting, it is plot relevant.
This is essentially a lesson in prompt engineering and model attention.
In a transformer model, attention mechanisms assign different weights to tokens in context, so the model does not treat every word equally. It prioritises some patterns and signals over others, based on learned relationships in the data.
In practice, this means:
If you emphasise a constraint, the model will focus on it.
If you include a number, it will try to anchor to it.
If you mention a risk, it may amplify it.
If you insert a bias, it may propagate it.
Every token in a prompt has the potential to shift the model’s probability distribution, even if some tokens matter far more than others. Context is never truly neutral.
This is why careful prompt design matters. Not because the model “understands” in a human sense, but because it weights.
Chekhov’s Cat reminds us:
If you put something in the system, expect it to matter.
This applies to:
Prompt design
Training data
Fine tuning corpora
Retrieval Augmented Generation
System instructions
Relevance is not random. It is influenced by what you spotlight.
The Deeper Pattern
All four cats point to the same truth.
Generative AI is not magic. It is a probabilistic system trained on human language, learning patterns from enormous text datasets such as websites, books and other written sources. Human language, in turn, encodes centuries of metaphor about uncertainty, causality, failure and relevance.
When we try to understand AI systems, we instinctively reach for physical metaphors because that is how we think.
Boxes.
Cats.
Razors.
Guns.
The irony is this:
We trained AI on human stories.
Now we use human stories to understand AI.
That circularity matters.
Because it means governance is not purely technical.
It is linguistic.
It is conceptual.
It is cultural.
Before organisations implement frameworks, policies or ISO standards, they need mental models that make risk legible.
Chesterton’s Cat teaches restraint.
Occam’s Cat teaches calm diagnosis.
Murphy’s Cat teaches resilience.
Chekhov’s Cat teaches precision.
Together, they form a surprisingly practical literacy layer for generative AI.
Not a compliance manual.
Not a certification syllabus.
A thinking discipline.
And in a field moving as fast as generative AI, disciplined thinking may be more valuable than any single tool.
Because models will change.
APIs will evolve.
Vendors will pivot.
But the underlying principles, the cats, remain.
If you can see them clearly, you can navigate the rest.
