On this page
Some principles tell you what follows. Others tell you what to be careful about.
Imagine you’ve just taken over an old software system. It’s full of strange restrictions, redundant-looking checks, and a deployment process that appears designed by someone with a personal grudge against convenience.
Your first impulse is obvious: clean it up.
Then someone tells you that a particular validation rule exists because, seven years ago, a billing bug sent thousands of duplicate invoices. The rule is ugly, yes. So was the incident it prevents.
That doesn’t mean the rule should stay. It means you needed to know what problem it solved before deciding to remove it.
“Before removing a fence, find out why it was put there.” G. K. Chesterton made the point in The Thing (1929), using a literal fence across a road.
It’s one of my favourite bits of practical wisdom, partly because it doesn’t demand obedience to tradition. It demands understanding before demolition.
But there’s a catch. Chesterton’s fence is a heuristic. It isn’t a theorem. A rule might have survived because it helps, because nobody has bothered removing it, or because it benefits someone powerful enough to keep it there. The fence’s existence gives us a reason to investigate, not a proof that we should preserve it.
That difference—between an argument and the assumptions that make it work—is where much of the interesting thinking begins.
The simplest explanation is not always the right one
Suppose our old system occasionally fails at the same time every day. One explanation is straightforward: a scheduled job exhausts a resource. Another involves a complicated interaction among several services, their caches, and a race condition triggered by a deployment schedule.
Which would you test first?
“Of two equally good explanations, prefer the simpler one.” Occam’s razor is a good place to start, provided we pay attention to equally good.
There is a mathematical connection to Bayesian model comparison. A flexible model can accommodate many possible datasets. Its probability is spread across those possibilities; a more constrained model that assigns substantial probability to what we actually observe can compare favourably. This is often called the Occam factor. The details depend on the likelihoods and priors, and a complex model can absolutely win when the evidence supports it. Myung and Pitt (1997) discuss a Bayesian formalisation of the trade-off.
So the slogan is a model-selection heuristic with formal relatives, not a theorem saying nature always prefers fewer parameters. If the complex explanation predicts the failures and the simple one doesn’t, simplicity is no excuse to ignore the evidence.
The same caution applies when we try to infer a thing’s future from how long it’s already lasted.
Old enough to be trusted?
A book that has been read for two centuries feels as though it might still be read in fifty years. A social-media trend that began this morning does not inspire the same confidence.
This intuition is often called the Lindy effect: “The longer something has lasted, the longer it may be expected to last.”
J. Richard Gott presented a striking statistical argument in Nature in 1993. If the moment at which you observe a phenomenon is uniformly distributed across its total lifetime, you are as likely to catch it in the middle half of its life as outside it. Under that sampling model, the interval from one-third to three times the elapsed lifetime covers the true remaining lifetime in half of repeated observations. That is a statement about the procedure’s coverage, not a 50% posterior probability for a particular object from its age alone.
It’s a delightful piece of reasoning. It is also a conditional statement about a particular observation model, not a universal forecast of survival. Carlton Caves’ critique examines why the observation model matters. Knowing an object’s age can convey very different information depending on its failure process, selection mechanism, and the prior information available. The age of a centuries-old idea and the age of a worn mechanical part are not interchangeable clues.
The more modest practical lesson is still interesting: longevity can be evidence worth considering, particularly when surviving for a long time required repeated tests against the world. But sometimes things survive through inertia, monopoly, or sheer luck.
Now Chesterton’s fence looks slightly different. Perhaps it lasted because it solved something important. Perhaps it lasted because nobody wanted the meeting required to remove it. Both are live hypotheses.
Good advice should tell us where to look, not exempt us from looking.
A chart can’t tell you what to value
Let’s return to the old system. Suppose removing that complicated check saves the team ten hours per month, but raises the probability of a rare and costly billing failure.
A sufficiently good model might quantify the trade-off. It could estimate the engineering hours saved, the expected number of incidents, the distribution of potential losses, and the likely effect on customers.
But it cannot, from those facts alone, tell us which trade-off we ought to accept.
“You can’t get an ought from an is.” David Hume’s famous observation points to a gap between descriptions of the world and prescriptions for what we should do. The Stanford Encyclopedia of Philosophy notes that the exact interpretation of Hume’s argument remains contested, but its widely used logical form is clear: purely descriptive premises do not by themselves yield a normative conclusion without some value-laden bridge.
If we also assume that reducing expected financial loss is our only objective, we can optimise for it. If we place special value on preventing catastrophic harm to a small number of customers, we may choose differently. The mathematics can illuminate the consequences of those values, not select the values from nowhere.
This is easy to forget when a dashboard produces a beautiful single-number score. Choosing what to maximise is a decision of its own.
Sometimes, even after people agree on how choices should be represented, it turns out their wishes cannot all be reconciled.
You may have to give up a reasonable-sounding requirement
Suppose a team must rank three proposed projects. Every person has their own ranking. You want a rule that turns those rankings into one consistent team ranking.
Naturally, you’d like the rule to respect unanimous preferences. You’d like it not to change its comparison between projects A and B merely because opinions about project C changed. You don’t want one person to dictate the result. And you’d like the procedure to work for every possible combination of individual rankings.
All fairly reasonable.
Kenneth Arrow proved that, with three or more alternatives and the standard requirement that society’s rankings form a coherent ordering, no aggregation rule satisfies all those conditions simultaneously. That’s Arrow’s impossibility theorem, carefully explained in the Stanford Encyclopedia of Philosophy.
The aphorism “No voting system is fair to everyone” is catchy but too broad. Arrow did not prove that every imaginable decision procedure is unfair under every definition of fairness. He proved an incompatibility among a specific collection of axioms for aggregating ranked preferences. Change the requirements or the problem’s structure and different mechanisms become available.
The theorem is more useful with its assumptions visible. It tells us that we cannot get everything on our wish list at once, so we need to discuss which requirement to relax and why.
That’s a very different kind of help from a slogan that simply declares politics hopeless.
The real work begins after the aphorism
These principles don’t all have the same epistemic status. Arrow’s result is a formal impossibility theorem. Hume’s problem is a philosophical challenge about justification. Occam’s razor has mathematical interpretations but no unconditional guarantee that the simpler story is true. The Lindy effect depends heavily on how lifetimes and observations are modelled. Chesterton’s fence is a wise question to ask before making changes, not a law of nature.
Treating them all as proven wisdom would miss what makes each interesting.
The purpose of an aphorism isn’t to replace thought with a convenient sentence. At its best, it tells us which thought is worth having next.
What’s the simplest explanation that actually predicts the data? What did this old rule once protect? Why has this idea survived? Which values are shaping our recommendation? Which desirable properties are impossible to satisfy simultaneously?
A short saying can get you to the right question. The rest is still your responsibility.