The intelligence move

The intelligence move is the choice of the word intelligence to describe what these systems are doing — a choice made by a small group of researchers in 1955, accepted by a field thereafter, and inherited by every contemporary public conversation about AI without being argued for again. The move is not the technology and not the research program. It is the act of naming, and the act of naming is doing political work that the technology and the research program would not be capable of doing on their own.

Three claims structure the rest of this thread. First, the original coinage was a positioning move in academic-political space, not a description of a phenomenon (§2). Second, the word intelligence arrived already loaded by the Galton/Spearman/Burt tradition that built it as a single rankable scalar so that humans could be sorted; calling computer systems intelligent inherits that ranking apparatus (§3). Third, the word is now functioning the way god functioned in earlier centuries and science functions now — as a master-word with clergy, an asymmetric claim on the public good, and the capacity to license harm in its name (§6).

The phrase artificial intelligence appears in the August 1955 proposal McCarthy, Minsky, Rochester, and Shannon wrote to the Rockefeller Foundation to fund the following summer's workshop at Dartmouth College. The text of the proposal is short, and worth reading in its own voice [to-expand]. The proposal is doing two pieces of work at once. It is making a technical bet — that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it — and it is making a naming choice that has to be read as distinct from the bet.

The naming choice was contested at the time. Wiener's cybernetics had a head start, a more developed theoretical apparatus, and institutional momentum at MIT. McCarthy's preferred term, by several accounts, was chosen partly to put institutional distance between the new field and Wiener's program, and partly because the word intelligence communicated ambition and legibility to non-technical funders in a way cybernetics and automata theory could not.

This is the first move to name as such: the choice was not a description of the system being built. It was a description of how the builders wanted the system to be received. Every subsequent argument about whether these systems are really intelligent has been organized by a word whose original work was political, not descriptive.

The word intelligence did not arrive in 1955 as a neutral term. It arrived loaded by roughly seven decades of psychometric work — Galton, Spearman, Burt, Terman, Yerkes — that had built it as a single, rankable, heritable scalar for the explicit purpose of sorting human populations. Stephen Jay Gould's The Mismeasure of Man (1981) is the definitive critique of that construct — both the statistical sleight of hand (the reification of g from a factor-analytic artifact into a real thing) and the political projects (immigration restriction, segregation, sterilization) it underwrote [to-expand].

When the 1956 cohort attached the word intelligence to their research program, this inheritance came with it. The inheritance is what makes AGI thinkable as a frame. AGI presupposes a single scalar called general intelligence, measurable on one axis, that systems and humans can be plotted on, and that more of is unambiguously the goal. Strip the Galtonian construct out and the word general in AGI becomes meaningless. So does most of the public-facing capability discourse — this model is smarter than the last one is a claim about a scalar that does not exist outside the measurement apparatus that produced it.

The Galtonian inheritance also explains why the framing slides so easily into ranking humans against machines. The scalar exists precisely so that things can be ranked on it. Once the systems are on the scalar, the question who is more intelligent, this person or this model is not a category error inside the frame — it is the frame's native question. The category error is the frame itself.

The naming move was contested inside the field essentially from the start. Four sources name four faces of the critique; none of them frame it as a single move, which is the contribution this thread is trying to add.

• Drew McDermott, Artificial Intelligence Meets Natural Stupidity (1976)

  The canonical insider naming. McDermott's paper is short, free, and angry: AI researchers routinely give their programs names that smuggle in the very competence the program is supposed to prove. A routine labeled UNDERSTAND does not understand. A symbol labeled knowledge is not knowledge. The vocabulary is doing work the code is not doing. Fifty years later the paper reads as a description of contemporary product marketing. [to-expand]

• Hubert Dreyfus, What Computers Can't Do (1972, rev. 1992)

  The phenomenological critique. Drawing on Heidegger and Merleau-Ponty, Dreyfus argued that the AI program of the 1960s and 70s presupposed a model of intelligence — explicit rules, context-free symbols, conscious deliberation — that bore no resemblance to how human competence actually works. The body, the situation, and tacit know-how were not edge cases; they were the phenomenon. Dreyfus was dismissed for a quarter century and then partially vindicated by the connectionist turn. The argument is now relevant again under different terms. [to-expand]

• Joseph Weizenbaum, Computer Power and Human Reason (1976)

  The ethical refusal. Weizenbaum built ELIZA, watched users — including psychiatrists — treat it as a real interlocutor, and wrote a book arguing that the field should not call what it was doing intelligence because the public would mistake symbol manipulation for judgment in ways that would be socially destructive. Weizenbaum was the conscience of the field. He was largely ignored. [to-expand]

• John Searle, the Chinese Room (1980) → I.3

  The analytic-philosophy critique. Searle argued that a system manipulating symbols by rules can produce intelligent-looking output without any understanding, because syntax does not constitute semantics. The argument has been contested since 1980. It also gave the framework its Cluster I.3 thread, where it is treated as the bridge to the institutional-facts argument rather than primarily as an AI critique.

Four critiques, four registers, half a century. None of the four authors framed the problem as a single naming move. This thread argues that they should have, because the naming move is the load-bearing structure underneath every specific objection — and because the move continues to operate now in contexts (regulation, capital allocation, education policy) where the four-author lineage is not in the room.

The cleanest contemporary version of the critique is Brian Cantwell Smith's in The Promise of Artificial Intelligence: Reckoning and Judgment (2019) [to-expand]. Smith distinguishes two capacities. Reckoning is calculation, prediction, pattern extension — the capacity to process and combine signals according to rules. Judgment is the capacity to be answerable to a world: to recognize that some answers are better than others not because they score higher on a metric but because they correspond to how things actually are and what actually matters.

Smith's argument is that contemporary AI systems have reckoning at extraordinary scale and have no judgment whatsoever. They cannot be held to account by a world because they do not have a relationship to one; they have a relationship to a distribution over tokens. The distinction is the strongest contemporary version of the older syntax/semantics line, made with full credit given to what the systems actually can do.

For this thread, Smith's contribution is that he gives the critique a defensible positive claim. The intelligence move is not just they are not intelligent; it is they have reckoning, they do not have judgment, and the word intelligence blurs the distinction in a way that systematically transfers judgment-authority to systems that have only reckoning. That transfer is what licenses the policy posture, the credentialing displacement, and the consumer reliance the rest of the investigation documents.

Castoriadis (Cluster I.1) called the words around which societies organize significations imaginaires sociales — master- words that do not refer to anything concrete but structure what counts as legitimate. Each historical era has its master-word. God organized the European medieval; science organized the modern; intelligence is organizing the present. The pattern is not metaphorical. Each master-word satisfies four structural conditions.

1. (1)

   Clergy. A class of interpreters with privileged access to the legitimate meaning, whose authority does not require lay verification. Theologians, then scientists, now AI researchers and the labs that employ them.

2. (2)

   Asymmetric paternalism. A claim to know better than the public what is good for the public — what the soul needs, what the evidence shows, what the model can be trusted to do.

3. (3)

   Denial of contestability. A refusal of the category itself as a thing the public is permitted to argue about — only within the category. You may dispute this doctrine, this paper, this model — you may not dispute the standing of theology, science, or AI to be the arbiter of the question.

4. (4)

   Licenced harm. Atrocities authorized in the master-word's name and resisting accountability in its name. The Inquisition, the colonial sciences, the deployed welfare algorithms Act IV documents.

The lineage matters because the available response strategies are inherited too. The Reformation was a response to the first master-word; the Enlightenment and the STS-and-critical-theory traditions are responses to the second; the response to the third is not yet named, but the same structural conditions generate the same kind of work.

• ·
  The naming-as-positioning observation.

  Every category-formation in the AI investigation should be read first as a positioning move and only second as a description. This applies to intelligence, but also to safety, alignment, frontier, open, and the category AI itself.

• ·
  The Galtonian critique of rankability.

  When a debate organizes itself around whether system A is more intelligent than system B, the framework's first move is to ask what scalar is being assumed and what work that assumption is doing.

• ·
  The reckoning/judgment distinction.

  Borrowed from Brian Cantwell Smith. The systems have reckoning; they do not have judgment; the word intelligence blurs the line in a way that systematically transfers judgment-authority to systems that have only reckoning.

• ·
  The master-word genealogy.

  Borrowed from Castoriadis. The framework reads intelligence as the third master-word in the god → science → intelligence sequence, and treats the structural similarities (clergy, paternalism, contestability denial, licenced harm) as a diagnostic rather than a coincidence.

A critique of intelligence lands harder when it sits next to the term we are committing to in its place. That commit is its own piece of work with its own argument structure and its own decision register; it has been promoted to a dedicated charter-level page rather than buried as the closing section of a critical thread.

What this thread still owes independently of the naming work: (a) a diagram for the move, (b) a close reading of the McCarthy / Minsky / Rochester / Shannon 1955 proposal in its own register, (c) a treatment of how the master-word operates differently in policy, marketing, and technical literature, and (d) an explicit articulation of what a Reformation of the intelligence master-word would look like — modeled on what worked and what did not in the earlier displacements.

1. McDermott 1976 — short, free, devastating; the right entry point.
2. Gould 1981, The Mismeasure of Man — the Galtonian critique; long, foundational.
3. Brian Cantwell Smith 2019, Reckoning and Judgment — the contemporary technical-philosophical critique.
4. Crawford 2021, Atlas of AI — the political-economy companion, queued in the AI investigation's to-expand list.
5. Bender, Gebru et al. 2021, Stochastic Parrots — the contemporary technical critique from inside the field.

Full URLs and a wider queue live on /investigations/ai-discourse/to-expand.