Against the Analogy Industrial Complex

In 2002, two psychologists at Yale ran a now-classic experiment. They asked subjects to rate how well they understood how everyday objects worked — zippers, ballpoint pens, helicopters, locks — on a scale from 1 to 7. Most people rated themselves around a 4. Then the psychologists asked them to write a detailed, step-by-step explanation of the mechanism. The subjects sketched, paused, gave up partway through. Asked again to rate their understanding, the average plunged. Leonid Rozenblit and Frank Keil called this the illusion of explanatory depth, and the result has been replicated dozens of times. People consistently confuse familiarity with comprehension. Knowing the what of a zipper feels indistinguishable, in the moment, from knowing the how.

This is the engine of the analogy-industrial-complex. Every Tuesday, somewhere on the internet, a new essay appears titled “What [Domain X] Can Teach Us About [Domain Y].” Your codebase is an archaeological site. Your engineering org is a forest ecosystem. Your AI model is a brain. Your startup is a garden. Each landing page lands with a small dopamine hit of understanding. Each one feels productive. Each one is, often, what Richard Feynman called wooden headphones — the form of insight without the function.

I want to make a precise case for what the genre is doing, where it works, and where it merely feels like it works. The test isn’t elegance or virality. It’s whether the mapping actually constrains what you’d build.

The Mechanism (and Why It Mostly Doesn’t Transfer)

This isn’t an attack on cross-domain thinking. It’s a specific challenge to the industrialized version of it — the form-letter essay where domain X is a “lens” on domain Y, the LinkedIn post where biology has profound lessons for product management, the conference keynote where a thermodynamics metaphor explains team dynamics. These are sometimes good. Mostly they aren’t, and the cognitive science gives us tools to say why.

Dedre Gentner’s structure-mapping theory, the dominant account of analogical reasoning since 1983, distinguishes two kinds of correspondence between source and target domains. Surface similarity maps object attributes — the things look alike, share descriptive features. Structural similarity maps systems of relations — the things behave alike under the same kind of mathematical or causal structure. Water flow through pipes maps onto current through circuits because both obey the same first-order equation: a potential difference drives flow through a resistance. The relation transfers. The water and the electrons share nothing in common as substances. That’s a structural analogy, and Gentner’s principle of systematicity says good analogical reasoners overwhelmingly prefer them.

When Kevin Dunbar and Isabelle Blanchette watched scientists actually use analogies in real lab meetings — what they called the “in vivo” approach — they found scientists almost always reached for structural analogies, not surface ones. But here’s the embarrassing finding for the genre I’m criticizing: in real labs, analogies came after the discovery, not before. Scientists used them mainly to explain unexpected results to colleagues. Communication, not generation. The hypothesis came first; the analogy framed it for an audience.

This is the first crack in the analogy-industrial-complex. Most of its essays are framed as if the analogy generated insight. Most likely, the writer already had the conclusion and reached for the analogy as an explanatory frame for readers.

The Three-Question Test

Here’s the test I want to propose. Apply it next time you read a cross-domain essay — including this one. An analogy does real intellectual work, rather than ornamental work, only if it passes all three:

1. Does the source domain’s formalism transfer? If the source has a mathematical structure, a law, an algorithm, a quantitative constraint — does that structure operate in the target? In Rutherford’s 1911 atomic model, the solar system analogy worked because both gravity and electrostatic attraction follow the inverse-square law. The math literally transferred. From solar-system mechanics, Rutherford could predict that most of an atom’s mass should be concentrated in a tiny dense nucleus — and the gold-foil scattering experiments confirmed it. The analogy constrained the design space of what an atom could look like. (The same analogy then broke productively: electrons should spiral into the nucleus under classical electrodynamics, which forced Bohr to quantize orbits. A working analogy fails informatively.)

Now compare “your engineering org is a forest.” Forests have actual ecological dynamics — succession, competition, nutrient cycling, edge effects. Does any of it transfer formally to org design? You can describe an org as a forest, but you cannot derive a load-bearing prediction from the analogy. The ecology doesn’t constrain the org chart.

2. Does the analogy generate a falsifiable prediction? Mary Hesse, in her 1966 Models and Analogies in Science, distinguished three components of any analogy. Positive analogy covers the properties source and target share — billiard balls and gas molecules both have mass. Negative analogy covers the properties they don’t — billiard balls have color; gas molecules don’t. Neutral analogy is the productive zone: properties where it’s unknown whether they transfer. Hesse’s claim was that all the epistemic action lives in the neutral zone. A productive analogy generates testable hypotheses about whether unmapped properties carry across. An analogy that maps only positive analogy — only already-confirmed similarities — does no scientific work. It’s decoration.

Most cross-domain essays I read map only positive analogy. They say “X is like Y because of features a, b, c,” and a, b, and c were all already known on both sides. There is no neutral zone. There is no hypothesis to falsify. The essay produces the feeling of insight without the falsifiable content of one.

3. Could the reader build the same thing without the analogy? This is the practical version. If you removed the analogy from the essay, would anything about what the reader builds be different? If yes, the analogy was functional. If no, it was decorative. Both are legitimate; they should not be confused.

Most cross-domain content fails Test 3. The reader would build the same thing either way; the analogy is a vocabulary they reach for after the fact to describe what they already wanted to do.

The Strongest Defense (and Why It Doesn’t Save Most Essays)

I owe the opposing position its strongest version. Two challenges deserve a hearing.

The first is from Douglas Hofstadter, who has spent four decades arguing that analogy isn’t one cognitive tool among many but the engine of all thought. Surfaces and Essences (2013) makes the case in granular detail: every act of categorization is an analogy. Recognizing a chair as a chair is an analogy to prior chairs. If Hofstadter is right, criticizing analogy use is like criticizing thinking.

The second is from Samuel Butler: “Though analogy is often misleading, it is the least misleading thing we have.” Pure deduction can’t reach novel conclusions; raw empiricism without conceptual scaffolding is paralyzed; argument from authority is worse. If you want to think about something genuinely new, you almost have to think it through some prior thing.

Both are right, and neither saves the genre. Hofstadter’s claim is about analogy as a foundational cognitive operation — what your brain does when it perceives a category. The analogy-industrial-complex is something more specific: a published-content genre where domain X is offered as a lens on domain Y, mostly via surface features, mostly without testable consequences. Calling that genre out doesn’t require denying that thinking is analogical. It requires distinguishing analogy as cognition (always present, mostly invisible) from analogy as content (a specific format with specific failure modes).

Butler’s concession is exactly the right framing. Analogies are unreliable, but they’re often the only way in. The point isn’t to abandon them. The point is to specify their wrongness, the way George Box specified his — “all models are wrong, but some are useful.” The useful models are the ones whose wrongness you can articulate. The wooden-headphones analogies are the ones whose wrongness has been smoothed away by the genre’s polish.

Where the Industry Misleads

Two failure modes deserve specific names.

The first is what philosopher Andrea Sullivan-Clarke called ingrained analogy in a 2019 paper in Perspectives on Science. When a community settles on a metaphor — “neural networks are brains,” “microservices are cells,” “technical debt is financial debt” — the metaphor stops being a hypothesis to test and becomes a lens that distorts perception. Researchers design experiments that conform to the metaphor; engineers build architectures that mimic it; product managers reach for explanations the metaphor pre-licenses. Sullivan-Clarke documented this in scientific contexts; it operates with at least equal force in technical communities, where metaphors have shorter shelf lives but heavier rotation.

The technical-debt metaphor is the canonical case. Ward Cunningham coined it in 1992, and it has been one of the most successful software analogies ever — partly because it gave non-technical executives a financial frame for understanding code maintenance. But it misleads in at least three structural ways. Financial debt has a known interest rate; technical debt’s cost is undetermined until you try to modify the code. Financial debt can be precisely quantified; technical debt cannot. Financial debt is voluntarily assumed; most technical debt accumulates unconsciously. The metaphor “afforded bad management too easily,” as one industry analyst put it — executives heard “debt” and assumed they could calculate a return on paying it down, which is exactly the inference the metaphor most strongly invites and most badly licenses.

The second failure mode is design fixation. In 1991, David Jansson and Steven Smith ran an experiment that’s become a small classic in design research. They gave engineers a problem to solve, accompanied by an example solution. They explicitly told the engineers that the example was flawed and contained features they should avoid. The engineers, both novices and experts, unconsciously reproduced the flawed features in their own designs anyway. Showing someone an analog narrows the design space, even when you tell them not to let it.

Cross-domain essays are inadvertent fixation devices. They hand the reader an example structure from another domain and ask them to think about their own problem through it. Jansson and Smith’s result predicts that the reader will end up with a solution shaped like the source domain whether or not the source domain’s structure actually transfers. This is precisely the opposite of what the genre advertises. The essays claim to broaden thinking. The cognitive science predicts they constrain it.

What Working Analogies Look Like

To stop being a hostile witness for a moment: here is what a working analogy looks like, by the test I’ve proposed.

Eiji Nakatsu, an engineer at JR West, was redesigning the nose of the Shinkansen bullet train in the 1990s. The train had a tunnel-boom problem — entering tunnels at 300 km/h, it compressed the air ahead of it and produced sonic shocks at the far end. Nakatsu, an amateur birdwatcher, modeled the redesigned nose on the kingfisher’s beak, which lets the bird enter water without splashing. The biological geometry — an elongated, gradually tapered profile that minimizes the pressure wave — directly constrained the engineering. The analogy passed all three tests. The formalism transferred (pressure-wave dynamics in air and water are governed by similar continuum-mechanics equations). The mapping generated falsifiable predictions (specific drag and noise reductions that wind-tunnel testing could confirm). And the reader — Nakatsu — could not have built the same thing without it. The kingfisher beak constrained the design space in a way no other analog he had access to did.

Velcro is the same shape of story. In the 1940s, George de Mestral pulled burdock burrs off his dog after a walk in the Alps and looked at them under a microscope. The mechanism — tiny hooks catching loops of fabric — transferred directly into the engineering. The analogy was the design. There was no decorative version.

Notice what these examples share. The source-domain mechanism is precise enough to be reverse-engineered. The transfer isn’t “X is like Y in a vague conceptual sense” but “the actual physical or mathematical structure of X can be implemented in the target.” This is rare. It’s the working subset of the population that the analogy-industrial-complex’s high-volume output mostly fails to imitate.

A Practical Asymmetry

The most honest reckoning is that analogies are good for two things and bad for a third. They’re good for communication — once you have a finding, an analogy is often the cleanest way to convey it to a non-specialist audience. They’re good for generation — they produce hypotheses worth testing, even when most don’t survive testing. They’re bad for substantiation — they cannot license a conclusion. The cargo-cult problem isn’t that analogies are used. It’s that they’re treated as if a sufficiently elegant mapping closes a question rather than opens one.

For the reader, the practical move is small. When you encounter a cross-domain essay, ask: what would change about my plans if this essay turned out to be wrong? If the answer is “nothing” — if you’d build the same thing either way — the essay was communication, not engineering. That’s still legitimate. But you should price it accordingly. Don’t make load-bearing decisions based on a vocabulary borrowed from a different domain unless you can show the structure transfers, the prediction is falsifiable, and the design space changes when you remove the analogy.

For the writer — including me, writing this — the move is the same. Before publishing the next “what X teaches us about Y,” run the three questions. If the analogy fails all three, it’s communication, not generation. Frame it that way. Don’t sell readers a discovery engine when what you have is a vocabulary refresh. Both are useful, but the conflation is the failure mode.

George Box again: “All models are wrong, but some are useful.” The useful ones are the ones whose wrongness you can specify. The analogy-industrial-complex specializes in models whose wrongness has been polished out — every line calibrated to feel like insight, every metaphor extended just far enough to land but not so far that it visibly breaks. The result reads beautifully and tells you almost nothing you didn’t already know.

The wooden-headphones test, then. The planes don’t land if all you have is the form. They land when the formalism transfers, when the prediction can be wrong, when removing the analogy changes what you’d build. Most cross-domain essays — including some of the ones I’ve written — fail at least one of those tests. The honest move isn’t to stop writing them. It’s to stop pretending the genre is what it isn’t.

Sources: Leonid Rozenblit & Frank Keil, “The Misunderstood Limits of Folk Science: An Illusion of Explanatory Depth,” Cognitive Science 26(5), 2002; Dedre Gentner, “Structure-Mapping: A Theoretical Framework for Analogy,” Cognitive Science 7(2), 1983; Kevin Dunbar & Isabelle Blanchette, “The In Vivo/In Vitro Approach to Cognition: The Case of Analogy,” Trends in Cognitive Sciences 5(8), 2001; Mary Hesse, Models and Analogies in Science (Notre Dame, 1966); Douglas Hofstadter & Emmanuel Sander, Surfaces and Essences (Basic Books, 2013); Andrea Sullivan-Clarke, “Misled by Metaphor: The Problem of Ingrained Analogy,” Perspectives on Science 27(2), 2019; Ward Cunningham, “The WyCash Portfolio Management System” (OOPSLA ’92 Experience Report); David Jansson & Steven Smith, “Design Fixation,” Design Studies 12(1), 1991; on Eiji Nakatsu and the Shinkansen 500 Series, Frank Fish & George Lauder, “Passive and Active Flow Control by Swimming Fishes and Mammals,” Annual Review of Fluid Mechanics 38, 2006, plus Nakatsu’s own published lectures; on George de Mestral, Velcro Companies historical materials.