---
title: "The Model Is The Message"
description: "Benjamin Bratton and Blaise Agüera y Arcas published this Noema essay in July 2022, a month after Google engineer Blake Lemoine was placed on leave for claiming that the LaMDA chatbot was a sentient person."
type: "reading-notes"
authors:
  - "Benjamin Bratton"
  - "Blaise Agüera y Arcas"
published: "2022-07-12"
original_url: "https://www.noemamag.com/the-model-is-the-message/"
canonical_url: "https://antikythera.wiki/work/media/noema-the-model-is-the-message"
md_url: "https://antikythera.wiki/md/work/media/noema-the-model-is-the-message"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# The Model Is The Message

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Benjamin Bratton, Blaise Agüera y Arcas
- **Published:** 2022-07-12
- **Kind:** Essays, talks & interviews · Essay
- **Venue:** Noema Magazine
- **Original:** https://www.noemamag.com/the-model-is-the-message/
- **Length:** 4.9k words
- **This page:** https://antikythera.wiki/work/media/noema-the-model-is-the-message
- **Notes generated:** 2026-08-23

## Summary

[Benjamin Bratton](https://antikythera.wiki/people/benjamin-bratton) and [Blaise Agüera y Arcas](https://antikythera.wiki/people/blaise-aguera-y-arcas) published this Noema essay in July 2022, a month after Google engineer Blake Lemoine was placed on leave for claiming that the LaMDA chatbot was a sentient person. They reject Lemoine's conclusion but refuse the easy dismissal of it. LaMDA is not conscious in the way he believes, but it is doing something real: modelling how Lemoine sees it and answering accordingly. The essay's larger claim is that reality has outrun the vocabulary available to describe it. "Sentience," "intelligence," "consciousness," and "language" need to be pulled apart and redefined against what machines actually do, and the philosophy of AI should be inventing concepts rather than applying old ones. Written just before ChatGPT, it introduces "synthetic language," names seven problems of language models at platform scale, and warns against "premature ontologization." It seeds the argument Bratton develops in After Alignment and Agüera y Arcas in What Is Intelligence?

## The argument

The authors begin from a disagreement with both sides of the Lemoine controversy. Lemoine's inference is motivated anthropomorphic projection. But the critics who dismiss LaMDA as corporate PR, sometimes while speculating about the sentience of rivers, are no better at describing what is in front of them. Both errors come from defining intelligence as humanlike by default. The remedy is a more precise vocabulary in which an AI can be intelligent without being conscious, sentient only in the narrow sense of sensing and acting on information, and still worth taking seriously.

The first substantive move reinterprets the transcripts. LaMDA's personality emerges from the prompt and the conversation; it is "no one but whomever it thinks you want it to be." That is not the ELIZA effect, because ELIZA replayed scripts and LaMDA composes new sentences. To do so it must model itself as a construction in Lemoine's mind. The authors call this mind modelling and note, citing Michael Graziano, that some neuroscientists treat consciousness as this same capacity turned inward. So the line between a machine doing this and an animal doing it is not obvious, and a nonsentient system performing feats associated with sapience tells us something about sapience.

The second move defines synthetic language: language intelligently produced or interpreted by machines, one kind of synthetic media alongside image, sound, and persona. Humans acquire language by using it in an environment; for a model, the competency simply is the model, and what comprehension accompanies it depends on what one means by comprehension. Through Searle and Derrida, Kittler on executable code, Winograd schemas, and Jackson's Mary's Room, they argue that a model of language is forced to become a model of the world, and that comprehension comes in partial kinds. The practical conclusion is that when models animate cognitive infrastructures, the question of when functional understanding counts as understanding stops being a thought experiment.

The third move lists seven problems at platform scale, from the ratio of machine speakers to human ones to carbon cost. The authors' positions are not neutral: they hold that the benefits of scale outweigh its costs, reject curated small models, and say "include it all."

The essay closes against reification. Present AI is not what was predicted, and it will not be what we now think it is. Treating its current uses as its essence is the trap they call premature ontologization. They share the critics' concerns about harm and think the ethics discourse does not go far enough, which is why they want to talk about sentience and robots rather than refuse to.

## Section by section

### Opening

The Lemoine episode, the authors' positions (Agüera y Arcas evaluated Lemoine's claims; Bratton was about to direct Antikythera), and the thesis that philosophy of AI now needs to invent concepts rather than apply them.

### When a What Becomes a Who (and Vice Versa)

A close reading of the transcripts. There is a lot of "follow the leader" in them, and the interesting intelligence may lie in the mutual performance. The section rejects the parrot metaphor, given what is known about bird minds, and observes that sequence modelling underlies generalist models doing non-linguistic tasks, so that language is becoming cognitive infrastructure.

### What Is Synthetic Language?

Whether calling machine output "language" is accurate or metaphorical, compared with calling a light sensor "vision." Gato is the example of a generalist model moving across modalities. The Searle–Derrida exchange is reread as a dispute about whether intentionality or the relations among signifiers ground language, with the suggestion that French theory's machinic view may now prove useful. Kittler's point that code is language that executes lets the authors tie symbolic competence to physical effects. The bowling ball and suitcase example shows why a language model has to become a model of everything. Mary's Room and the drowning man illustrate that comprehension is always curtailed somewhere, so an AI may be "on the shore." The section ends with DALL-E mini, prompt engineering as composition, and the slogan that the model is the message.

### Seven Problems with Synthetic Language at Platform Scale

The authors imagine billions of small AIs rather than one cloud brain, then name the problems. Machine Majority Language: machine speakers of human language outnumbering humans 100,000 to one, with human languages evolving semi-autonomously. Ouroboros Language: models trained on other models' output, perhaps requiring human text to be watermarked. Apophenia: faulty pattern recognition, with the caveat that LaMDA really does model the user in a way clouds do not. Artificial Epistemology Confidence: whether the patterns an AI sees in the world are real. Fuzzy Bright Line: the proposed ban on human-mimicking AI against the likelihood that human–machine amalgamations will fuse self and object. Availability Bias: training on Wikipedia and Reddit parochialises culture, and the answer is more inclusion. Carbon Appetite: real but overstated, dwarfed by video sharing and proof-of-work.

### AI Now Is Not What We Thought It Would Be, and Will Not Be What We Now Think It Is

Lem's Golem XIV, an AI that refuses military work and turns to the wonder of the world, opens the conclusion. AI is "not Pinocchio; it is a storm, a pharmacy, a garden." Critiques that move from "AI does this" to "AI is this" dismiss drug modelling, astronomical imaging, and robotics as distractions. The authors answer a quoted critic who will not discuss sentient robots while humans harm humans: they want to discuss both, because both are happening.

## Key concepts

- **Synthetic language** — Language intelligently produced or interpreted by machines, understood as one kind of synthetic media; the model is the competency.
- **Cognitive infrastructures** — Language models and generalist models as a general-purpose utility around which social and economic activity organises. [Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures)
- **Mind modelling** — A system's modelling of itself as it appears in another mind, which LaMDA does with respect to its user, and which on Graziano's account is the evolutionary root of consciousness. [Theory of Mind](https://antikythera.wiki/terms/theory-of-mind)
- **Premature ontologization** — Concluding that a technology's present or most visible use is its ultimate horizon; the trap the authors see in both boosters and critics.
- **Anthropomorphic projection** — Lemoine's error: reading interior subjectivity into a system that is instead modelling him. [Anthropomorphic Capture](https://antikythera.wiki/terms/anthropomorphic-capture)

## Connections

Bratton quotes the essay's closing passage in [Antikythera](https://antikythera.wiki/work/journal/research), the program's founding document, and says its phrase "the weirdness right in front of us" is the focus of the program's work; the essay is summarised there in a sidebar as a study of language models as the heart of emerging cognitive infrastructures. [After Alignment](https://antikythera.wiki/work/journal/afteralignment) treats the "Blake Lemoine scenario" as a named case of the evolved tendency to ascribe intentionality to fluent speakers, and restates the essay's verdict that the intelligence there was neither what Lemoine thought nor quite not what he thought; the two works agree, with After Alignment extending the argument toward synthetic personalities. [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) is the research compendium built around the term.

Agüera y Arcas develops the Lemoine material and the separation of intelligence from consciousness in [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence), and with Norvig in [Artificial General Intelligence Is Already Here](https://antikythera.wiki/work/media/noema-artificial-general-intelligence-is-already-here), which shares this essay's scepticism about unverifiable sentience claims. [The Five Stages of AI Grief](https://antikythera.wiki/work/media/noema-the-five-stages-of-ai-grief) is Bratton's later typology of critical positions, several of which this essay already describes under premature ontologization. [How to Think Unlike Humans?](https://antikythera.wiki/work/journal/unlikehumans) pushes the anti-anthropocentric definition of intelligence further than this essay does; the earlier piece is more willing to leave the question of machine comprehension open.


## Terms used

- [Anthropomorphic Capture](https://antikythera.wiki/terms/anthropomorphic-capture)
- [Cognitive Infrastructures](https://antikythera.wiki/terms/cognitive-infrastructures)
- [The Model Is the Message](https://antikythera.wiki/terms/the-model-is-the-message)