The ideas on this page (34), each with its own short page
Worked out here: Classification torque · Conceptual metaphor · Conway’s law · Cost asymmetry · Defaults · Einstellung effect · Framework · Framework feedback · Framework pricing · Framework rotation · Framing effect · Goodhart’s law · Inductive bias · Influence, not imprisonment · Law of the instrument · Legibility · Linguistic relativity · Looping effects · Notation as a tool of thought · Option space · Path dependence and lock-in · Performativity · The map is not the territory · Theory-ladenness and paradigms
Examples: Black-Scholes and the option market · Chess experts and the Einstellung effect · Leibniz’s dy/dx and the British lag · Pirahã and exact number · Prompt formatting and the 76-point swing · Russian blues · Scientific forestry and the Normalbaum · The Asian disease problem · The Columbia slide
Open questions: Does a medium forbid thoughts, or only price them?
You have filled in a form that had no box for your situation. You picked the closest one, and from then on, as far as the office was concerned, that was your situation. Nobody lied to anyone. The form did it.
So far the damage to an idea has happened between people: in the retelling, and at the gates that decide what gets retold. There is damage one step earlier, inside the tool a single person is thinking with, before anything is sent. Chapter 5b says a variant the option space cannot hold “never forms.” A research survey of the neighbouring literature (filed as a source) says the same machinery runs inside one head, and that “never” is too strong.
A slide that could not say it
In 2003, while the shuttle Columbia was still in orbit, Boeing engineers briefed NASA managers on whether a foam strike could have done real damage. They had the worrying fact: the test data behind their reassuring estimate came from debris far smaller than the piece that hit. The fact made it into the briefing. It sat buried in a nested bullet list, under the headline “Review Of Test Data Indicates Conservatism for Tile Penetration.”
The accident board later wrote that “it is easy to understand how a senior manager might read this PowerPoint slide and not realize that it addresses a life-threatening situation.” Nothing was removed from that slide. The engineers were reporting inside a tool that makes a ranked list of fragments cheap and a sustained argument expensive, and the caveat went where the tool puts caveats. The board documented many other causes, and this is one slide; I use it as a picture of the mechanism (the case, with its limits).
A framework is a medium you think in
By a framework I mean any structure someone built that other people then think inside: a language, a notation, a metaphor, a classification, a model, a metric, a tool, an org chart. A slide deck is one. So is a tax form, a diagnostic manual, a spreadsheet, a unit of measurement.
A framework is a medium for thought: selection performed once, in advance, on what its users will find easy to notice, record and say. That is the medium argument moved one step upstream. The pipeline starts where the world becomes a measurement, and Chapter 1 already says “raw data is misnamed. The ‘raw’ is the first compression, pretending to be the source.” What chooses that first compression is the framework: the instrument, the list of categories, the form with its boxes. The first gate has a medium attached, like every later one.
A dozen fields found this separately and named it a dozen ways. Linguists call it linguistic relativity. Philosophers of science call it theory-ladenness. Psychologists found framing effects and the Einstellung effect. Students of bureaucracy found legibility and torque. Economists found Goodhart’s law and performativity. Engineers found Conway’s law. Each is the same move in a different room: the framework sets the default, and defaults drive outcomes.
What a framework does to a thought
Four things, and three of them are the three things a medium does to a gate, seen from inside.
It decides what can be recorded at all. Eighteenth-century German foresters kept tables of saleable timber. The tables had no column for underbrush, soil life or the mix of species, so those dropped out of forestry’s reasoning. James C. Scott calls this making a thing legible, and what a framework cannot hold does not get weighed.
It makes some operations cheap and others laborious. Leibniz’s dy/dx carries the chain rule in its symbols; Newton’s dots do not. Place-value numerals make long division a routine; Roman numerals push it onto an abacus. That is cost asymmetry, and over time people mostly do what is cheap.
It supplies the answer you get when you do not choose. Tversky and Kahneman described one set of outcomes two ways. Framed as lives saved, 72 percent of people took the certain option; framed as lives lost, 78 percent took the gamble. The arithmetic was identical. The default moved, and the choice moved with it.
And it feeds back on the world, which gets its own section below. Put together, a framework is a menu with prices on it: some dishes are not listed, some are cheap, and one arrives if you say nothing. Most people, most days, order accordingly. I call the mechanism framework pricing.
The price list, not the wall
“Never forms” is the strong version of this claim. It has been tested for decades, mostly on language, and it mostly fails.
The Pirahã language has no stable word even for “one.” Its speakers still match large exact quantities perfectly while the objects stay in view, and lose accuracy only when the task makes them hold a number in memory. The researchers concluded that number words are “a cognitive technology for keeping track of the cardinality of large sets across time, space, and changes in modality”: a tool that extends a capacity, whose absence did not remove the thought. Russian has two obligatory words for blue, and Russian speakers tell two blues apart faster when the shades fall on opposite sides of that boundary, an edge English speakers do not show; give them a verbal task at the same time and the advantage disappears. Chess experts who have found a familiar solution keep looking at it while saying they are searching for something better, but when the familiar move is an outright blunder, every one of them avoids it.
A framework rarely forbids a thought. It prices it. The careful, expensive variant can form; it costs more, so across a population it mostly does not. That is influence, not imprisonment, and it changes the remedy. You tear down a wall. You can pay a price, with effort, with training, or with someone else’s subsidy. Paying it spends the reader’s finite budget, and want decides whether it gets paid. A plotting library that makes the standard chart one line and the unusual chart a fight (seaborn against matplotlib) is a price list in this sense, and so is every artifact on Chapter 4’s curve.
When the map starts moving the territory
The bias gets dangerous when people act on it.
The foresters did not stop at measuring. They defined a standard tree, cleared the underbrush, cut the species down toward one, and planted in rows, until the forest matched the table. The first generation of trees grew well. After the second, yields fell (Scott quotes a source putting the loss at 20 to 30 percent) and German acquired a word for forest death, Waldsterben (scientific-forestry).
The same loop runs in markets: an option-pricing formula fit real prices poorly when it was published in 1973, fit them well once traders priced with it, and stopped fitting after the 1987 crash (black-scholes). It runs in metrics, where a measure that becomes a target stops measuring. It runs in categories applied to people, who change in response to being sorted (looping effects).
The bias matters most where a framework is acted on, because the world is then rebuilt to fit it and the evidence of misfit disappears. I call this framework feedback. Chapter 8b finds the pipeline wired into its own cargo for truths made of agreement, like money (constitutive transmission). Feedback is that wiring met from the other side, and it reaches further than agreement does. A forest is not made of agreement.
The clearest case is a machine
Machine-learning researchers call a model’s built-in assumptions its inductive bias. Theirs is the one field that treats framework bias as unavoidable and chooses it on purpose. It is also where the effect has been measured most starkly. Change nothing about a prompt except its formatting (separators, spacing, capitalization) and one open model’s accuracy moves by up to 76 points (prompt-format-sensitivity). That is a reader whose decoding hangs on the wrapper. And a handful of benchmarks has come to stand for progress across the field, so labs build toward the score (ai-benchmarks): Goodhart’s law operating on a research community. Both bear on the capture surfaces of Chapter 11.
Where I land
A framework is frozen selection running inside the thinker. It works by price and default, so it bends populations and leaves individuals free: any one person can think the expensive thought, and most people, most of the time, will not. The damage compounds where the framework is acted on. The defense is plural. A single framework fails silently, and it is the disagreement between two of them that shows the shape of each. The practice is framework rotation: use several lenses, report results across framings, keep measures apart from targets, and question the categories as hard as the conclusions. Versatile expertise is rotation practiced by one person, and the curse of expertise is the Einstellung effect at the size of a career.
How far to trust all this: the dependable finding across fields is a weak but measurable bias toward whatever the framework makes easy, visible or rewarded. Framing effects replicate (31 of 36 samples in one large rerun). Much of the famous language work does not. Several of the best stories (the slide, the calculus notation, the formula blamed for 2008) have one cause among many, the keyboard story is disputed outright, and each example’s page says how solid it is.
Where I’m still uncertain
- Is a price nobody pays still a price? If the training that lets people think against a framework hollows out for a generation, the expensive thought may be unaffordable for nearly everyone. This is the time-scale caveat the medium note already carries: a nudge over a year, possibly a wall over a generation (tracked as an open question).
- The forest breaks one of my own carvings. I have said the pipeline can garble a claim about objective reality but cannot touch the fact. The foresters touched the fact. A framework with power behind it can rebuild a managed piece of the world to match its map, at least until biology objects. The line between truths made by reality and truths made by agreement may need a third region for realities that are administered.
- My own frameworks are in play. The pipeline diagram makes a straight, stage-by-stage flow cheap to think and loops expensive. The sorting of ideas into mechanisms and concepts fits some of them badly. Both deserve the questioning recommended above.
- I have these sources secondhand. Everything here comes through one survey. Scott, MacKenzie, and Bowker and Star are the first three to read directly.