Language That Makes Power Hard to See

There is a peculiar magic trick powerful institutions perform with words.

The trick is not necessarily lying.

Lying is crude. It has the decency to leave a fact-checking target. It can be disproven. The more durable technique is to describe something accurately in a vocabulary that changes what the listener notices.

Consider a few invented formulations. Ten thousand people are not fired. The enterprise executes a workforce optimization initiative. Same parking lot. Fewer badges. Much calmer noun.

Jobs are not removed and handed to automation. The organization captures productivity through an AI-enabled operating-model transformation. The work disappears into a passive verb; the savings get a strategy deck.

A customer is not kept on an expensive legacy plan because the company expects inertia. The customer remains on a standard post-promotional rate.

A package does not contain less product for the same sticker price. The brand executes a pack-size architecture adjustment.

A financial intervention is described in terms of liquidity support or market functioning. Those descriptions can be accurate while leaving its effects on private portfolios and asset prices unexplained. The shared words do not make the operations identical. Their balance-sheet mechanics matter.

None of those phrases has to be false.

That is why they work.

The strongest research basis for this argument is not cultural cynicism. It is economics. Gabaix and Laibson developed a model of “shrouded attributes” in which firms deliberately obscure costly add-ons from consumers who are not fully attentive. Their crucial result is that informational shrouding can survive even in highly competitive markets and even when advertising is cheap. Competition does not automatically save the consumer because a firm that educates everyone may lose the profitable inattentive customer while sophisticated consumers exploit the subsidized headline offer. [S101]S101 · Shrouded Attributes, Consumer Myopia, and Information Suppression in Competitive MarketsXavier Gabaix; David Laibson · 2005peer-reviewed economics / NBER · mechanism evidenceClose

That model is more than twenty years old. The digital economy did not retire it. It gave it CSS.

The Federal Trade Commission’s dark-pattern work documents companies hiding or obscuring material information, burying mandatory fees, disguising advertising as independent content, creating intentionally difficult cancellation paths, and manipulating users toward surrendering data. [S102]S102 · Bringing Dark Patterns to LightFederal Trade Commission · 2022official regulator report · documented practiceClose The relevant information often exists. It may even be technically disclosed. The design problem is to make the information arrive too late, require too much effort, appear visually subordinate, or disappear into a forest of terms the user is unlikely to parse.

So I want a hard distinction throughout this essay:

Disclosure is not transparency.

Disclosure is the presence of information in a formally available location.

Transparency is the practical ability of a relevant human being to perceive, understand, compare, and act on it before the decision becomes costly to reverse.

Those are different engineering objectives.

The same distinction applies to expertise. Specialized fields require specialized vocabulary. I do not expect a cardiologist to replace “ventricular fibrillation” with “the heart is doing a weird little dance” for the sake of accessibility. Precision matters. Technical distinctions matter. The problem begins when the specialist controls both the underlying complexity and the way outsiders are permitted to encounter it, then treats the outsider’s inability to contest the explanation as evidence that the explanation has been accepted.

The Federal Reserve preserved a useful historical account. In a November 2009 oral-history session, Alan Greenspan recalled a tradition of deliberately maintaining uncertainty about policy to support two-sided markets. He also described that rationale weakening as markets grew. The published transcript notes that he did not review it. The evidence is narrow and useful: a retrospective account of one deliberate communication strategy, and a concrete example of ambiguity serving an institutional purpose. [S103]S103 · Alan Greenspan Oral History InterviewFederal Reserve Board · 2009official institutional oral history · historical primary evidenceClose

And this is where vocabulary becomes power rather than mere style.

Sociologists Wendy Espeland and Mitchell Stevens describe commensuration as the process of transforming different qualities into a common metric. [S104]S104 · Commensuration as a Social ProcessWendy Nelson Espeland; Mitchell L. Stevens · 1998peer-reviewed sociology · conceptual mechanismClose That sounds abstract until you notice how much of modern institutional life depends on it. Credit scores. University rankings. Employee performance bands. ESG scores. Vendor quadrants. Risk ratings. Customer lifetime value. Engagement. Net promoter scores. Productivity metrics. Sovereign credit ratings. “Talent density.” Every one of these instruments compresses messy reality into a form that can be compared.

Compression is useful. It is also political.

Every metric decides what counts, what does not count, which dimensions are made visible, and which disappear. Once the metric becomes institutionally authoritative, organizations begin adapting themselves to it. Espeland and Michael Sauder documented this reactivity in rankings: the measurement does not merely observe the world; it changes the behavior of the people being measured. [S105]S105 · Rankings and Reactivity: How Public Measures Recreate Social WorldsWendy Nelson Espeland; Michael Sauder · 2007peer-reviewed sociology · mechanism evidenceClose

That mechanism is much more important than a dispute over who first coined a technology label. The deeper power is the ability of a highly trusted advisory institution to define categories, construct taxonomies, establish evaluation dimensions, populate quadrants, create maturity models, forecast adoption curves, brief buyers, brief vendors, brief the press, convene conferences around the resulting vocabulary, and sell access to analysts who interpret the map. Draw the map long enough and people eventually start confusing cartography with geography.

I have a sore point about Gartner’s Magic Quadrant for exactly this reason. I spent years inside Gartner. I know why a compressed market map is useful. I also know what happens when the map graduates into an answer key. My crude analogy is buying a television from a matrix that compares size and price:

I’ve made you this matrix that compares size and price. Which size and price do you want? That’s the best one.

The dimensions may be measured perfectly. Congratulations: we have performed rigorous measurement on two variables I did not ask to optimize. They still do not tell me which television has the best picture, which fits my room, which has the input lag I can tolerate, which lasts, or which works with the equipment I already own. A real technology purchase is worse, because the missing dimensions can include regulatory requirements, migration cost, contractual terms, architecture, staffing, data residency, accessibility, reliability, security, and the simple fact that a specialist vendor may do six things magnificently while an incumbent does seven hundred.

The problem is not that the quadrant contains no information. The problem begins when the abstraction inherits the authority of the institution that drew it. A dot in the upper-right corner can become a procurement argument all by itself. The dot enters the meeting before the software does. A buyer’s actual utility function quietly gets replaced by the evaluator’s categories.

Language becomes infrastructure.

This is why corporate jargon deserves more serious treatment than jokes about consultants. Sometimes jargon is merely ugly. Sometimes it is compact precision among specialists. Sometimes it is semantic anesthesia. Anesthesia is useful too. The question is what we’re numbing.

Consider the moral emphasis in these two invented recommendations:

We recommend eliminating 10,000 jobs and automating as much of the remaining work as possible.

and saying:

We recommend rightsizing the workforce while accelerating digitally enabled productivity capture across the target operating model.

The second sentence does not necessarily contain a factual lie. It changes the object of attention. People disappear into “workforce.” Firing becomes “rightsizing.” The human consequences become transition costs. The moral decision is converted into an administrative transformation.

Organizational research gives us reasons to expect this abstraction to matter. Moral-disengagement research finds that people use cognitive mechanisms that allow harmful behavior to be reconciled with a positive self-conception. [S109]S109 · Why Employees Do Bad Things: Moral Disengagement and Unethical Organizational BehaviorCelia Moore et al. · 2012peer-reviewed organizational psychology · behavioral evidenceClose Diffusion of responsibility, euphemistic labeling, advantageous comparison, and displacement of agency are not exotic pathologies. They are ordinary ways humans reduce moral friction.

The language of modern institutions is astonishingly good at producing exactly that friction reduction. Once the sentence is wearing enough nouns, nobody appears to have done anything to anybody.

This need not involve a coordinated plan. Consider a hypothetical chain of local choices. A consultant favors a less confrontational phrase. A lawyer favors a formulation that limits exposure. A product manager favors a design that increases conversion. Each choice can be intelligible on its own. Their combined language may still make it harder to connect an institutional decision with its human consequences.

Eventually people outside the institution experience the consequence in plain English while being told the cause in a language they do not speak.

Their bill went up.

Their package got smaller.

Their job disappeared.

Their rent increased.

Their mortgage became unaffordable.

Their neighborhood was bought by investors.

Their pension depends on the market going up.

Their government spends a trillion dollars a year on interest.

Their news feed makes everybody look insane.

And the systems producing those outcomes answer in acronyms. The person says, “My bill went up.” The system replies with a three-letter initialism and a PDF.

Information can be present and still be unavailable

A disclosure can exist without becoming usable knowledge. That is the difference between placing a fact somewhere and giving a person a reasonable chance to understand it before the fact matters.

Imagine that a service discloses its introductory price on one screen, its later price in a linked document, its equipment charge in a separate table, and its cancellation procedure in an account menu available only after purchase. Every piece might be accurate. The person still has to assemble a working model of the transaction. The seller has one integrated model. The buyer receives fragments. One side gets a database. The other gets a scavenger hunt.

The problem is not simply that the text is long. Sometimes a long explanation is necessary. The problem is that integration has a cost, and the parties do not bear that cost equally. A firm can spread the work of designing a contract across millions of transactions. A household must understand its own contract while doing everything else required to live. The firm learns from repeated encounters. The customer may confront the issue once. The asymmetry is not verbosity. It is rehearsal.

Gabaix and Laibson’s model of shrouded attributes is important here because it does not assume competition automatically makes every hidden term clear. Under their model’s conditions, firms can profit from consumers who overlook add-on costs, and revealing those costs need not be the winning strategy for a rival. It gives us one rigorous mechanism by which complexity can remain profitable even in a competitive market. [S101]S101 · Shrouded Attributes, Consumer Myopia, and Information Suppression in Competitive MarketsXavier Gabaix; David Laibson · 2005peer-reviewed economics / NBER · mechanism evidenceClose

The Federal Trade Commission’s dark-pattern report documents a related practical problem: interfaces that hide material information, make cancellation difficult, disguise advertising, or steer privacy choices. Its categories show how a formally available choice can be shaped by presentation and friction. [S102]S102 · Bringing Dark Patterns to LightFederal Trade Commission · 2022official regulator report · documented practiceClose

What I draw from these sources is a distinction between information scarcity and usable-understanding scarcity. More words do not necessarily solve the latter. An AI system that generates an excellent explanation could help a customer. An AI system that generates an excellent explanation of why the customer’s objection does not qualify could help an institution maintain the same barrier more efficiently.

The difference lies partly in who gets to define the question. “Explain the charges” is not the same task as “help this person establish whether the charges match the offer.” One produces a description. The other compares a promise, a bill, and a remedy. A system can perform the first beautifully while never being asked to perform the second.

The cost of making a claim

The administrative-burden literature distinguishes learning costs, compliance costs, and psychological costs. A person has to find out how a process works, do what it requires, and endure the experience of navigating it. Moynihan, Herd, and Harvey developed that framework in research on citizen-state interactions, including a historical Medicaid case. The framework identifies costs that a simple count of completed transactions can miss; the motive behind any particular process is a separate question. [S205]S205 · Administrative Burden: Learning, Psychological, and Compliance Costs in Citizen-State InteractionsDonald Moynihan; Pamela Herd; Hope Harvey · 2015peer-reviewed public administration / JPART · administrative-friction frameworkClose

The distinction travels usefully into a hypothetical customer dispute. Learning the procedure is different from finding a document. Finding the document is different from having time to upload it. Completing the form is different from tolerating another rejection written in language that gives no useful next step. We should not treat these stages as one undifferentiated failure of effort by the person trying to proceed.

A system can therefore filter people without ever saying that they are excluded. It can require a sequence of small tasks whose combined cost exceeds what some people can bear. A person with spare time, stable internet access, organized records, and confidence in dealing with institutions may complete the sequence. Another may abandon it. The resulting data will show different rates of completion. It will not, by itself, explain why.

This is where automation creates two distinct possibilities. It can reduce the burden by carrying information across steps, explaining requirements, and helping repair a mistake. Or it can increase the rate at which institutions generate demands for information. A cheap, personalized request is still a request someone must answer. Removing work from one side of an exchange does not prove that total work has fallen.

Suppose a machine can review a claim instantly but requires the claimant to provide ten new documents. The firm’s processing time falls. The person’s burden rises. A report that measures only internal processing time will describe an efficiency gain. A report that follows the whole transaction may describe a transfer of work.

This is an accounting problem with an ethical consequence. The edge of an organization is not the edge of the world. A cost does not vanish when it is pushed beyond the boundary of the budget being measured. Excel is not a portal to another dimension.

Precision and obscurity are not the same thing

Technical language is not the enemy. I need it when distinctions matter. A Treasury buyback is not quantitative easing. Revenue is not profit. Statistical association is not causation. An observed model behavior is not an authenticated report of the model’s internal state. Removing those distinctions would make this essay easier to read in the worst possible way: by making it less true.

The useful test is whether a term preserves a distinction that helps the reader understand the matter at hand. “Duration” can be explained as a measure of how sensitive a bond’s price is to a change in yield. Once explained, it saves repeated paragraphs. “Value realization” often needs a second question: value for whom, realized in what form, over which period?

The same phrase can be precise in one setting and evasive in another. “Resource allocation” may be an exact description in a computer scheduler. In a meeting about people losing their jobs, it can conceal the human referent unless the speaker restores it. The fault is not in syllable count. It is in what the abstraction lets the conversation forget. If the term preserves a distinction, keep it. If it buries a person, dig.

This distinction matters for anyone trying to use AI to make complex systems legible. Simplification must not remove the conditions that determine whether a statement is true. A plain-language explanation of a loan must retain the rate, the timing, the fees, and the consequences of missed payments. Replacing a dense agreement with cheerful prose is not enough when the cheerful prose leaves out the costly part.

There is a second danger: fluent explanation can create confidence without understanding. An answer may feel clear because every sentence is smooth, even when the underlying relationship remains unclear. A person can understand every word in “your request does not meet the relevant criteria” and still have no idea which criterion failed or what evidence could change the decision.

That is why an explanation should sometimes be judged by what it allows a person to do next. Can they check the claim? Can they correct an input? Can they identify the responsible party? Can they distinguish a binding rule from a discretionary choice? Can they see when an appeal would be pointless rather than spend hours in a process with no available remedy?

The standard I want is simpler: distinguish language that describes a wall from language that identifies a door. A polished description of the wall is still a wall. A human-facing system needs to know which of those tasks it has actually completed.

When the description becomes the world

Commensuration makes unlike things comparable. It also creates a temptation to forget that comparison required a choice. A numerical score can look discovered when it was partly designed. Give it three decimal places and the design starts dressing like geology. The categories may be useful, well studied, and transparent. They still do not arrive from outside human judgment. [S104]S104 · Commensuration as a Social ProcessWendy Nelson Espeland; Mitchell L. Stevens · 1998peer-reviewed sociology · conceptual mechanismClose

Imagine two hypothetical schools evaluated only by their graduates’ immediate earnings. One may serve students who already have strong advantages. Another may help students overcome severe barriers. The ranking can be calculated correctly while answering a narrower question than a reader assumes. Adding more data will not fix that mismatch unless the definition of the question changes.

The same problem appears when a company reports average customer satisfaction after excluding customers who never reached the survey, or reports average response time without measuring whether the answer helped. These are illustrative measurement designs. Their lesson is logical: a metric can be accurate about the selected records and incomplete about the experience it appears to represent.

Once organizations respond to a metric, the measurement joins the process it was meant to observe. Espeland and Sauder’s work on rankings describes this reactivity. People adjust behavior in response to being measured, and the adjustment can reshape the institution. [S105]S105 · Rankings and Reactivity: How Public Measures Recreate Social WorldsWendy Nelson Espeland; Michael Sauder · 2007peer-reviewed sociology · mechanism evidenceClose

AI adds another layer. A model may learn from records already shaped by a ranking. It may then recommend behavior that performs well under that ranking. The next records reflect those recommendations. The system can become increasingly competent at reproducing the measurement regime without independently establishing that the regime measures what people value. It can become a world-class optimizer of a question nobody bothered to reopen.

The mechanism follows from the sequence: define a target, reward behavior that matches it, record the rewarded behavior, and train on the record. The feedback loop can get smarter without getting wiser. The learning is still not free from the original choice.

That is why I care about the institutions that sell categories and evaluations. Their product is not merely a set of words. It can become part of the environment in which other organizations decide what success looks like. The next question is who gets to draw that map, who pays to understand it, and what happens when the map becomes a condition of entry.