Possible Minds, illustratedCaroline A. Jones

Possible Minds (2019)

Caroline A. Jones, “The Artistic Use of Cybernetic Beings”

Before “artificial intelligence” became the ruling metaphor, artists used cybernetics to build embodied, affect-driven beings that came alive only when coupled to people and environments; that path (left cybernetics) is the needed corrective to today's disembodied, task-fragmenting, profit-driven AI (right cybernetics).

Author
Caroline A. Jones (professor of art history, Department of Architecture, MIT)
Essay
“The Artistic Use of Cybernetic Beings,” chapter 24 of Brockman (ed.), Possible Minds (2019)
PDF pages
227–235 (p.227 is Brockman's headnote; the essay itself runs pp.228–235)
Print pages
254–65

Scroll, or use the arrow keys, to move through the essay one step at a time. The plate paints each step as you reach it.

1. The argument

Step 1. Artists wanted behaviour, not thinking machines. AI was not what artists first took from Wiener's cybernetics. They built sculptures that simulated responsive sentience out of switches and sensors, and that history has been erased: “Now obscured by an ideology of a free-floating ‘intelligence’ untethered by either hardware or flesh, AI has forgotten the early days of cybernetics' uptake by artists.” (PDF p.228) With sensor and effector organs sold as hobbyist kits, the computer was an adjunct organ rather than an electronic brain, so “artists were bricoleurs of electronic bodies, interested in actions rather than in calculation or cognition.” (PDF p.230)

Step 2. Cybernetics is old, bodily, and about control. The concepts Wiener named (feedback, damping, homeostasis, systems thinking) go back four centuries, and the word runs from the Greek steersman to Watt's governor, which Jones reads as a lineage of masculine control (PDF pp.228–229). Wiener's command, control and communication described information producing actions in an environment, “a muscular, fleshy agenda often minimized in genealogies of AI.” (PDF p.229) She closes on Wiener's warning that we must modify ourselves to live in the environment we have modified, and asks whether we have become the tools of our tools (PDF p.229).

Step 3. 1968: the viewer becomes a component. Jasia Reichardt's Cybernetic Serendipity at the ICA and Wen-Ying Tsai's Cybernetic Sculpture in New York mark the high-water mark (PDF p.229). Gordon Pask, whom Jones calls a second-order cybernetician, showed Colloquy of Mobiles (1968), in which a “male” apparatus of rods communicated with “female” fiberglass forms. Jones mocks its gender theatre but credits Pask's focus on interactivity and responsiveness (PDF p.230). The show's lasting idea was a paradigm: “the machinic ecosystem, in which the viewer was a biological part, tasked with figuring out just what the triggers for interaction might be.” (PDF p.230)

Step 4. The coupling was affective, and humans had to learn the protocol. Edward Ihnatowicz's Senster (1970), a fifteen-foot hydraulic claw run by a twelve-bit computer, was programmed to be shy, recoiling at loud noise and sudden movement: “Only those humans willing to speak softly and modulate their gestures would be rewarded by Senster's quiet, inquisitive approach” (PDF p.231). Ihnatowicz discovered this himself when the machine turned to him after he cleared his throat, and in Tsai's show an attendant probably had to tell visitors to clap (PDF p.231). The lesson: “we sense a growing necessity to train the public to experience itself as embedded in a technologized environment, modifying itself to communicate intuitively with machines.” (PDF p.231)

Step 5. Frank simulation, then feminist critique. The 1960s works “made no investment in ‘intelligence.’ Knowing machines were dumb and incapable of emotion, these creators were confident in staging frank simulations.” (PDF p.232) In the 1990s, women artists inspired by Radical Software and Haraway's A Cyborg Manifesto recast that innocent technosphere as a condition needing critique. Lynn Hershman Leeson's dolls made viewers aware of being both seeing subjects and objects looked at, and asked whose intelligence AI was simulating (PDF p.232). Judith Barry's Imagination, Dead Imagine (1991) stays locked into the body and refuses detached intelligence (PDF p.233).

Step 6. AI inflated a modest conjecture into a mission. Jones quotes the 1955 Dartmouth proposal by McCarthy, Minsky, Rochester and Shannon, which conjectured 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.” (PDF p.233) That goal has grown into Google DeepMind's ambition to solve intelligence, and what cracks is the social contract: lost driving jobs, privatised utilities, depersonalised care. She asks whether these are the whips Wiener feared we would learn to love (PDF p.233).

Step 7. A model is not life, and drills are not wisdom. Philippe Parreno's 2014 firefly piece animates drawings with Conway's Game of Life (PDF pp.233–234). Jones rejects the claim, voiced in a Hawking film, that such rules show how a mind might arise. Following Gregory Bateson, mind is not bounded by the skin (PDF p.234), and models are “parts of signaling systems constituting ‘intelligence’ only when their creaturely counterparts engage them in lively meaning making.” (PDF p.234) Then her sharpest charge: “Contemporary AI has talked itself into a corner by instrumentalizing and particularizing tasks and subroutines, confusing these drills with actual wisdom.” (PDF p.234)

Step 8. From right cybernetics back to left cybernetics. Jones names the path to automated weapons, Uber and Google right cybernetics and calls for a return to left cybernetics: theoretical biologists and anthropologists with a trans-species view of intelligent systems (PDF p.235). Bateson's diagnosis of the corporation is her model of failure: corporations “merely simulate ‘aggregates of parts of persons,’ with profit-maximizing decisions cut off from ‘wider and wiser parts of the mind’” (PDF p.235). The alternative is Bateson's immanent larger Mind, built from “inputs and actions that then become inputs for other actions in concert with other entities” (PDF p.235).

A note on voice. The headnote on p.227 is Brockman's; it reports the left/right distinction from an interview, so the steps above cite only the essay (PDF p.235). Nested quotations (the Dartmouth proposal, Gregory Bateson) are attributed to their authors as quoted by Jones.

2. Related work since 2019

Jones's own work

  1. “In Praise of Wetware” (Caroline A. Jones), 2019, MIT Ethics, Computing, and AI commentary series, summarised in MIT News. news.mit.edu/2019/ethics-computing-and-ai-perspectives-mit-0318 . Its stated action is to understand adaptive human wetware, and its dependence on a larger living ecosystem, before claiming machine intelligence: the essay's Bateson argument restated as advice to MIT's new college of computing.
  2. “Senster” (Caroline A. Jones), 2021, book chapter, Presses universitaires de Paris Nanterre. doi.org/10.4000/books.pupo.26717 . A chapter-length study of Ihnatowicz, the essay's central case. The fetched record confirms the author is the MIT art historian.
  3. “Unfolding Intelligence,” panel “Open Systems” (convened by Jones), April 2021, MIT Center for Art, Science & Technology. arts.mit.edu/unfolding-intelligence/ and unfoldingai.mit.edu/panelist/caroline-jones/ . Jones frames open systems as opening intelligence out to forms that are machinic, multi-species or “cosmically other”, extending the essay's claim that mind is not inside one skull to the definition of intelligence.
  4. Symbionts: Contemporary Artists and the Biosphere (Caroline A. Jones, Natalie Bell, Selby Nimrod, eds.), 2022, MIT Press and MIT List Visual Arts Center (exhibition October 2022 to February 2023). mitpress.mit.edu/9780262544481/symbionts/ and listart.mit.edu/exhibitions/symbionts-contemporary-artists-biosphere . The curatorial form of the essay's closing call for symbiotic webs: artists engaging living materials as collaborators rather than tools.
  5. “Symbiontics, Machine Learning, and the Internet of Earthlings” (Caroline A. Jones), lecture, 1 March 2024, Concordia Centre for Sensory Studies. www.concordia.ca/cuevents/artsci/cissc/2024/03/01/symbiontics-machine-learning-and-the-internet-of-earthlings.html (abstract: www.sensorystudies.org/wordpress/wp-content/uploads/2024/01/Symbiontics-Caroline-A-Jones.pdf). The abstract calls artificial intelligence a misnomer, since algorithms only iterate within the contexts and training sets human cognition has set. This is the essay's argument about models, applied to machine learning. (A 2023 SVA talk on GAN art and her “Latent Archive” project: sva.edu/events/the-algorithmic-state-what-are-machines-learning-and-what-are-we-learning-from-machines)

Work by others on the essay's cybernetic lineage

  1. “Gordon Pask's Conversation Theory and Interaction of Actors Theory: Research to Practice” (Tilak, Manning, Glassman, Pangaro), 2024, Enacting Cybernetics. doi.org/10.58695/ec.11 (repository copy: discovery.ucl.ac.uk/id/eprint/10196048/1/66c33b2ccd36d.pdf). Presents Pask's theories as tools for designing “human-to-human, human-to-machine, and machine-to-machine conversations”, treating users and tools as one homeostatic whole. It is the most direct bridge from Colloquy of Mobiles to agent-to-agent design.
  2. “Artificial Intelligence Seen Through the Lens of Bateson's Ecology of Mind” (D. Griffiths), 2021, International Journal of Interactive Multimedia and Artificial Intelligence 7(1). www.ijimai.org/index.php/ijimai/article/view/652 . Concludes that symbolic AI and neural networks fall short of Bateson's criteria for mind, but that AI applications “can form part of an ecology of mind”. That is Jones's position (models become intelligent only in coupling), argued from Bateson directly.
  3. “The Agent Use of Agent Beings: Agent Cybernetics Is the Missing Science of Foundation Agents” (Wang, Yang, Zhao, Lin, Hu), 2026, arXiv. arxiv.org/abs/2605.10754 . Maps six classical cybernetic laws (Wiener, Ashby, Shannon, von Foerster, Qian Xuesen) onto design principles for long-running LLM agents, with code generation as one application domain. It revives cybernetics for agents as a science of reliability, closer to Jones's right lineage than her left one, which is why it is worth reading against the essay.

Searched and not found. The logged searches (Exa, OpenAlex, arXiv; agent jones in resources/exa-cache/index.jsonl) found no post-2019 work by Jones on multi-agent or LLM systems. OpenAlex's author lookup for “Caroline A. Jones” returned only a different researcher (combustion and phosphor thermometry), so none of those items is used.

3. Bearing on multi-agent and multi-swarm orchestration

Jones never writes about orchestration, and her essay is partly a polemic against the culture that builds it. But the artworks she describes are specific coupling designs between people and machine collectives, and several translate into mechanisms for an orchestrator (Grok) dispatching to cross-checking workers (Claude Code, Codex).

3.1 The shy worker: gating on input quality. Senster recoiled from loud noise and sudden movement and approached only when people modulated their behaviour (PDF p.231). The gate is on the quality of the signal, not on the task. For a coding worker the analogue is a dispatch contract that recoils when the instruction is contradictory, underspecified or out of distribution: it returns a clarification request or a scoped-down plan, not a confident patch. The orchestrator then learns to speak softly (precise spec, minimal reproduction, named files). Today's workers are built to be maximally compliant, the opposite design. The nearest engineering precedent found is HULA (Takerngsaksiri et al., 2024, arxiv.org/abs/2411.12924), which adds a human agent at the planning and coding stages of a coding-agent pipeline deployed in Atlassian JIRA. It gates on human approval, not on the worker's own reading of its input.

3.2 The attendant who knows the protocol, and is inside the loop. In Tsai's show an attendant probably had to explain that clapping triggered the sculptures, and Ihnatowicz understood Senster only when it turned toward him (PDF p.231). The operator's highest-leverage role is often the attendant's: not steering every step, but knowing which inputs trigger which worker behaviours (prompt conventions, file ownership, escalation rules). That knowledge belongs in a versioned protocol file the orchestrator reads, not in the operator's head. And because the operator's prompts and review habits shape what workers do, evaluating the system without the operator measures a different system.

3.3 Cross-check as colloquy, not vote. Pask's Colloquy of Mobiles was machines conversing with machines, with humans able to join (PDF p.230). Conversation theory, as Tilak et al. (2024) present it, is about how agreement is produced between agents (doi.org/10.58695/ec.11). The usual cross-check, an independent review ending in approve or reject, discards the disagreement. A Paskian cross-check has the two workers converge on an explicit shared statement (what the change does, which invariants it keeps, what the tests show) and logs where they agree to disagree. JuryFlow (2026, arxiv.org/abs/2609.40103v1) makes a similar move for LLM judges: its abstract objects that majority voting discards inter-judge conflict and proposes a disagreement-guided, human-in-the-loop alternative.

3.4 Provenance confusion. Reichardt's catalogue warned that visitors would not know whether a work was made by an artist, engineer, mathematician or architect, and Jones calls the show prophetic for blurring that line (PDF p.230). A merged multi-agent change set has the same property, which raises a testable choice: reviewers can see authorship (to discount a model's known blind spots) or be blinded to it (to judge the artifact, not the author).

3.5 Multi-swarm: don't build a Batesonian corporation. Bateson's corporation, a set of partial persons cut off from wider and wiser parts of the mind (PDF p.235), and Jones's charge that AI mistakes task drills for wisdom (PDF p.234) are one diagnosis. A swarm of swarms in which each sub-orchestrator optimises its local metric (tests pass, ticket closed) is exactly such an aggregate of parts. The design response is a channel every swarm senses and none owns: an end-to-end acceptance suite, user-visible behaviour, or a reviewer that holds the whole task rather than a slice. This is Jones's machinic ecosystem (PDF p.230) read as architecture. The swarms share an environment, and coordination runs through it as well as through the meta-orchestrator.

4. Cybernetics

  • Second-order cybernetics, named and used. Jones describes Pask as a second-order cybernetician (PDF p.230) and builds her conclusion on Gregory Bateson, via Mary Catherine Bateson's paraphrase of his second-order cybernetics (PDF p.234). Von Foerster's distinction is between the cybernetics of observed and of observing systems (W6, Glanville 2004, www.emerald.com/insight/content/doi/10.1108/03684920410556007/full/html). Her artworks enact it: Hershman Leeson's dolls make the viewer both observer and observed (PDF p.232), and Senster made its designer a component (PDF p.231).
  • Feedback, extended from machine to ecology. Jones accepts feedback and homeostasis as cybernetics' core (PDF p.228) but relocates the loop across bodies, rooms and species, ending in Bateson's larger Mind of actions that become inputs to other actions (PDF p.235). Brockman's introduction gives the book's baseline, Wiener's world of interlocking feedback loops (W3, PDF p.24).
  • Purpose and control, contested. She reads the steersman etymology as a lineage of masculine command (PDF p.229) and asks whether automation's harms are the whips Wiener feared (PDF p.233). That links her to Wiener's warning about the purpose put into the machine (W2, www.cs.umd.edu/users/gasarch/BLOGPAPERS/moral.pdf). She goes further: the problem is not only mis-specified purpose but control itself as the organising metaphor.
  • Rejected: the capture of cybernetics by AI. Cybernetics, she writes, has collapsed into a discourse of AI that was not preordained (PDF p.228). The 2026 “Agent Cybernetics” paper (§2, item 8) shows the term being re-captured for agent reliability engineering, the right-cybernetic reading she warns against.
  • Not used: requisite variety. Ashby's law (W4) does not appear in the essay. Counting coupled humans and environments as part of a regulator's variety would be our extension, not her claim.

5. Agreement and clash with Society of Mind

The essay mentions Minsky only as co-author of the 1955 Dartmouth proposal (PDF p.233), which is her emblem of what went wrong. The comparison is therefore partly adversarial.

Agreement 1: behaviour from dumb parts. Jones's 1960s artists knew their machines were dumb and staged convincing behaviour anyway (PDF p.232). Minsky starts from the same premise.

> “To explain the mind, we have to show how minds are built from mindless stuff, from parts that are much smaller and simpler than anything we'd consider smart.” (SoM §1.1, www.aurellem.org/society-of-mind/som-1.1.html)

Both locate apparent sentience in organisation, not in a smart component.

Agreement 2: mind as interacting processes, not one principle. Jones's larger Mind is a web of actions that become inputs to other entities' actions (PDF p.235). Minsky also refuses a single seat of intelligence.

> “The power of intelligence stems from our vast diversity, not from any single, perfect principle.” (SoM §30.8, www.aurellem.org/society-of-mind/som-30.8.html)

Tension 1: decomposition as method versus decomposition as disease. Jones blames contemporary AI for instrumentalizing and particularizing tasks and subroutines (PDF p.234). For Minsky, splitting a job among subordinate agents is how competence arises.

> “building a tower is too complicated a job for any single, simple agent, so Builder has to ask for help from several other agents” (SoM §1.4, www.aurellem.org/society-of-mind/som-1.4.html)

An orchestrator dispatching subtasks is Minsky's Builder. Jones's critique says that the sum of drills is not wisdom unless something holds the wider context (§3.5).

Tension 2: where the mind's boundary lies. With Bateson, Jones places mind outside the cranium, in coupling with people and environment (PDF p.234). Minsky's society lives inside one head, and his supervisor is deliberately sealed off from the world.

> “don't connect the B-brain to the outer world at all; instead, connect it so that the A-brain is the B-brain's world!” (SoM §6.4, www.aurellem.org/society-of-mind/som-6.4.html)

Minsky's design argues for an orchestrator that monitors worker process and ignores the world. Jones argues that the decisive signals come from outside the agent society: users, environments, the operator.

Tension 3: simulation as honest artifice versus simulation as explanation. Jones praises frank simulations and denies that Conway's rules show how a mind arises (PDF pp.232, 234). The Dartmouth conjecture Minsky co-signed (PDF p.233) treats precise description and simulation as the route to intelligence, and Society of Mind (SoM §1.1, www.aurellem.org/society-of-mind/som-1.1.html) is an account of minds built from mindless parts. For Jones a simulation becomes intelligent only when a creature engages it (PDF p.234). For Minsky the right simulation would itself be a mind.

6. Seeds for open questions

  1. Shy workers. In a Claude Code/Codex coding pipeline, does a Senster-style input gate (the worker returns a clarification or scoped plan when the spec is contradictory or underspecified, instead of a patch) reduce confidently wrong merged changes and total operator time, compared with always-compliant workers, on a task set with deliberately injected spec defects?
  1. Colloquy versus vote. Does requiring two heterogeneous workers to converge on an explicit, logged shared statement of a change (intent, preserved invariants, test evidence) before acceptance catch more defects per token than independent review with an approve/reject verdict? Closest prior work found: JuryFlow (arxiv.org/abs/2609.40103v1), for LLM judges.
  1. Blind or labelled provenance. Does hiding which model authored a patch from the reviewing model change acceptance rates and defect detection, and in which direction? This tests whether cross-checks judge artifacts or authors.
  1. Who holds the wider mind? For tasks split into subtasks, what fraction of failures is detectable only with whole-task context, and does one whole-context reviewer (or a shared end-to-end acceptance suite across swarms) catch them more cheaply than tighter slice-level checks?

7. Sources

Book: John Brockman (ed.), Possible Minds: Twenty-Five Ways of Looking at AI (Penguin Press, 2019), PDF pp.227–235 (local extract in resources/book-text/, not committed). Brockman's introduction is cited once (PDF p.24) via the reference pack.

All URLs accessed 2026-10-03; each appears in resources/exa-cache/index.jsonl.

Jones's work

Other related work

Reference pack (resources/reference-pack/minsky-wiener.md)

Step 1 of 35The drawing shows the step of the essay currently in view. The text beside it is the full essay.