Battle-Tested

You may have noticed, if you glanced at the sources behind the last chapter, that a lot of them are old.

The study that named the behaviour is from 1963. The factory curtain is from 2012. The work on how attention gets allocated is from the late nineties. For a book whose central character is running a large-language model on her laptop in a year that starts with a two, that is a strange bibliography.

Hold that thought, because by the end of this chapter I want it to have flipped completely, from the thing you would use to attack the argument into the thing that makes the argument hard to attack. The age of this evidence is not a weakness I am hoping you won’t notice. In my opinion it is the whole point.

So let me show you the same behaviour surfacing, under a different name, in every decade since the Beatles released their first album, and you can watch what changes and what doesn’t.

1963: it gets a name

Donald Schön was studying how genuinely new inventions made it out of corporate research labs. Not incremental improvements, but the real departures, the products that didn’t fit the existing roadmap. He expected to find a process. What he found was people going around the process.

The interesting inventions, he reported, tended to have a champion: someone who believed in the idea and was willing to protect it, quietly, from the organisation that would otherwise have killed it in review. The champion often worked outside the official plan and sometimes in direct defiance of it. Schön gave the behaviour a name that survived, bootlegging, and made a claim that I think should have been far more disturbing to his readers than it apparently was — that for certain kinds of innovation the informal, hidden route was not a failure of the system but the only route that worked at all.

Note what is already present in 1963, before the personal computer, before the internet, before anything you would call technology in the modern sense. The hiding is there. So is the pro-organisational motive, because the champion is not stealing anything, he is protecting an idea from his own employer. And so is the structural cause, a formal process better at killing young ideas than at growing them.

Every element of Maria’s Tuesday is sitting in Schön’s 1963 paper. Only the tools have changed.

1970s–80s: it builds the airplane, and companies try to tame it

The next thing the pattern does is get famous.

Lockheed’s Advanced Development Programs, the Skunk Works, is where most people first met the idea. Kelly Johnson’s team produced the U-2 and the SR-71 Blackbird by operating deliberately outside the parent company’s normal process: small, autonomous, secretive, insulated from the reviews and sign-offs that governed everything else. The lesson corporations took from this was seductive and, I think, slightly wrong — that if you want radical output you build a walled garden and let a special team play in it.

I say slightly wrong because of what happened next, which is the more revealing move. Companies noticed that unofficial work produced value and they tried to domesticate it. If bootlegging works, the reasoning went, let us authorise it. 3M formalised its fifteen per cent time. Later, famously, Google offered twenty per cent time. The underground would be given a sanctioned room to happen in.

Hold on to that instinct, because we are going to watch it fail, repeatedly, for the rest of the book. For now just register that by the 1980s the pattern was strong enough that large, sophisticated companies were building official programmes to try to capture it. You do not build a programme to capture a fluke. You build one to capture something you have noticed keeps happening whether you like it or not.

1990s: it moves into software, and it gets a second name

When work moved onto computers the pattern moved with it, and picked up a label from a completely different discipline.

The information-systems world started writing about shadow IT, meaning systems, tools and workarounds that employees built or bought without the sanction of the IT department. The spreadsheet that quietly runs a business unit. The little database nobody approved. The tool the team depends on that appears on no official inventory. Silic and Back, studying it, described people bypassing sanctioned systems not out of malice but to close the gap between rigid corporate tooling and what the actual work demanded.

Same behaviour, same motive, new name, because a new profession had discovered it and assumed it was theirs.

This is the moment the pattern’s disguise turns into a real problem. Each field renamed it, each field studied it in isolation, and nobody stood on a chair and said that this is the same thing Schön described in 1963. The R&D scholars had bootlegging. The IS scholars had shadow IT. The strategy scholars, as we will see, had their own vocabulary again. One phenomenon wearing a different coat in every room, so that nobody quite noticed how old and how general it was.

2000s–2010s: it gets caught, and it gets punished

By now the pattern has a rap sheet, and this decade adds the fact that matters most for Maria, which is consequences.

Researchers began looking not just at whether shadow work happened but at what happened to the people who did it. One experimental study, by Allen and colleagues, is worth stating precisely, because it is the empirical spine under Maria’s fear. They looked at how employees who used shadow IT were evaluated, and found that those employees were less likely to receive a bonus in both the good-outcome and the bad-outcome conditions.

Read that twice. It did not matter whether the shadow tool worked. The penalty attached to the circumvention itself. You could break the rule and succeed, or break the rule and fail, and either way the evaluator marked you down for having broken the rule.

That is the exact calculation Maria ran in Chapter 1, and now you know it is not paranoia but a documented pattern. The organisation punishes the going-around rather than the outcome of the going-around, which means a rational employee learns to hide the going-around, which produces precisely the invisible fleet we started with.

2025: it goes mass

And then the tools got cheap, and the pattern stopped being a minority sport.

We saw the number in Chapter 1: roughly ninety per cent of employees using personal AI tools for work, against forty per cent of companies with official ones. What took a champion’s courage in 1963, a skunkworks’ budget in 1975 and a clever engineer’s evenings in 1995 now takes eighteen euros and a Sunday afternoon. The barrier to running your own submarine collapsed, so the fleet grew from a few daring captains to very nearly everyone.

That is the whole change. Not the behaviour, but the cost of the behaviour. We will spend Chapter 3 on what that collapse actually means, because it is the reason this book is being written now rather than in any of the six previous decades when it could have been.

The thing that didn’t change

Line those decades up and look at what stayed constant while the technology underneath was replaced four times over.

The hiding stayed. The pro-organisational motive stayed, and in every era, across every rename, the dominant finding is that these people are trying to help their employer rather than rob it. The structural cause stayed, which is a formal process systematically better at stopping young ideas than at growing them. And the disguise stayed too, with each generation greeting the pattern as a novel symptom of its own new technology, naming it afresh, and missing that it was looking at something with a sixty-year pedigree.

Four technology revolutions and one unchanged pattern. A behaviour that survives that does not depend on the tools. It depends on something a great deal more stable than the tools, which is the shape of the large organisation itself, and it will therefore not be fixed by a better tool, a new policy, or the current transformation programme. It is wiring.

This is what I mean by battle-tested, and the phrase needs discipline, because it is very easy to abuse.

What “battle-tested” is allowed to mean

There is a lazy version of this argument that says the sources are old, therefore they are proven. That is not true, and I want no part of it. Old and wrong is a common combination, and age on its own certifies nothing at all.

The claim is narrower than that, and stronger. A finding is battle-tested when it has been re-tested — examined again, by different people, in different places, in different decades, under different conditions — and keeps surviving. It is the survival across re-tests that carries the weight, not the birth date.

Take the sharpest example, one we will lean on hard in Part II. In 1987, research found that companies set their investment hurdle rates, meaning the return a project must clear to get funded, far above their actual cost of capital, and then wait until an opportunity clears that inflated bar before committing. An old finding, and easy to wave away as a relic of eighties finance.

Except that it kept getting re-tested. Through the two-thousands. Through the twenty-tens. And into the twenty-twenties, with interest rates having collapsed and then whipsawed in between, the finding held: hurdle rates stayed stubbornly, irrationally high, barely moving while the cost of money moved underneath them by enormous margins. A 2022 review of decades of data described them as very sticky. Work in 2024 and beyond found them high still. So the 1987 result is not a relic at all. It is the first data point in an unbroken line running to the present.

That is what I am willing to call battle-tested. Not old, so trust it. Re-tested across four decades and two interest-rate regimes, so trust it.

The same discipline has to cut the other way, and this is the part I care about most. Where a finding is old and has not been convincingly re-tested, or worse, where later work has failed to reproduce it, I say so and I don’t lean on it. There is a famous result about escalation of commitment, the tendency to throw good money after bad into a failing project, that fits this book’s argument almost too neatly. It is from 1976, it is canonical, and recent attempts to reproduce it in the lab have struggled, with some finding the opposite effect. So when it turns up later it turns up with that warning attached, doing lighter work than its fame would allow. A durable pattern earns confidence. A famous but shaky one does not, however much I might want it to.

The distinction matters because the reliability of the whole book rests on it. Battle-tested is a badge I am going to hang on specific findings that earned it by re-test and refuse to hang on ones that didn’t. Applied to everything old it would mean nothing. Applied only to what survived, I think it means quite a lot.

Why this makes the AI question sharper, not softer

Here is the payoff, and it sets up the hinge the whole book swings on.

If the pattern were new, a quirk of generative AI or a symptom of this particular technology, then the sensible response would be to wait it out. New things settle. Frontier tools mature, governance catches up, the exception gets absorbed into the rule, and in five years the whole shadow-AI phenomenon looks like an adolescent phase the industry grew out of.

But it isn’t new. It has survived every previous attempt by every previous technology to either enable it or stamp it out. It outlived the mainframe, the PC, client-server, the web and mobile. It shrugged off 3M’s programme and Google’s programme and every governance crackdown in between.

Which turns the question you actually care about into something much more pointed. Not “will this shadow-AI thing blow over,” because the historical record says it won’t, any more than the previous five waves blew over. The real question is the one this entire book is walking toward:

Is AI just the sixth wave of a pattern that has survived all five before it, or is it, for the first time, the thing that changes the pattern itself?

That question is only available to us because the research is old. A book built on 2025 citations couldn’t ask it, because it would have no way to establish that there was a sixty-year pattern for AI to either continue or break. The very thing that looked like this argument’s weakness, the dates, is what lets it ask its most important question with a straight face.

We will spend the book earning an answer. But we cannot even ask it honestly until we have done the next thing, which is to understand why the machine buries good small ideas in the first place, reasonably and defensibly and catastrophically, at every single step.

That is Part II. It starts with the gate.

Next: Chapter 3 — why a pattern this old is arriving now at a scale it has never had, which turns out to be two changes rather than one.