Joscha Bach on Consciousness, Causal Patterns and Machine Minds

Tim Scarfe introduces a conversation with Joscha Bach about whether consciousness requires something beyond a computational process. Scarfe notes his own intuition that the physical substrate may matter, while describing Bach’s broadly functionalist proposal: a sufficiently trained algorithm might give rise to experience. Both frame the issue as unsettled.

Scarfe asks what drives life forward. Bach describes self-organization: under some conditions, complexity can increase and systems can develop models of possible futures. He calls a system able to act to control future states an agent.

Bach cautions that there is no law requiring increasing complexity to culminate in a single universal agent. Catastrophes can destroy local pockets of complexity, and the complexity we observe around us may be local and temporary rather than a guaranteed trajectory of the universe.

Scarfe reads a passage from Bach and Hikari Sorensen’s work that describes a mind constructing a world model. In the account, contrast provides dimensions of difference; combining them creates space, and representations locate things and their relations within it.

The passage goes on to distinguish a space of ideas from a model of the external world, allowing possibilities to be considered without changing the world model. It describes the mind modeling objects, living things and an agent—the personal self—as part of that constructed world.

Scarfe asks why consciousness remains difficult to research and refers to debate around integrated information theory. Bach says that, in his view, existing theories have not adequately explained how experience arises; he regards several prominent accounts as partial.

Bach discusses attention schema theory, which he characterizes as a model of attention, and global workspace theory. He credits neuroscience work associated with the latter but argues that these accounts do not by themselves explain how subjective experience comes about.

Bach challenges integrated information theory as a non-functionalist account. As he presents the contrast, a suitably integrated digital neuromorphic system might count as conscious under the theory, while a conventional computer running an equivalent emulation might not.

He invokes computability and emulation: a conventional computer can emulate another computer and reproduce the same functional behavior and verbal reports. In his view, treating one system’s report of experience as true and the equivalent emulation’s report as false exposes a problem for that account. This is Bach’s philosophical objection, not an established resolution of the debate.

Scarfe invokes Leibniz’s mill: if a thinking machine were enlarged enough to walk inside, we might see only parts pushing one another, not perception itself. He asks how this enduring intuition about a ghost in the machine can be addressed.

Bach proposes another descriptive level: causal patterns. Money, he says, requires a physical substrate but is not explained by the particular molecules in a banknote. It can affect the world and can also be implemented in computers. He uses software as an example of a real causal pattern.

Bach relates this view to his reading of Aristotle’s account of psyche: a pattern bound to and shaping a living body, rather than necessarily an immortal substance. He compares it to the attempt in artificial life to create an agentic pattern that organizes a substrate and models its environment.

On Bach’s account, an agent can construct a model of reality that includes a representation of itself as real. He asks whether an LLM’s occasional claim to consciousness is more simulated than human experience. Such language can reflect its training on human texts and fictional descriptions, so the report alone does not settle the question.

Scarfe presses the objection that the pattern might be causally disconnected from the underlying machinery. Bach replies with a computer as a “causal insulator”: a Minecraft world can run on different machines while its inhabitants have no direct information about the case, voltage or location of the host computer.

Bach compares the brain to such an insulator: people can imagine futures and remember pasts distinct from their immediate surroundings. Different brains can also learn the same language or represent a similar thought, suggesting that some patterns can be carried by different physical implementations.

He adds that humans as presently known depend on brains and their communication mechanisms. In principle, he argues, an artificial substrate that reproduces the relevant message types and temporal and computational properties might emulate a person; he does not present this as something already achieved.

Scarfe asks whether this view overlooks history and situatedness: perhaps the path by which an artist or mind develops is essential, rather than reducible to a transferable program. Bach grants the intuition’s force for people but questions whether it applies to a perfectly deterministic system capable of reproducing the same steps.

Bach says he does not know whether today’s computers can be made conscious. He calls it an empirical question and says there is no mathematical proof either way. He proposes investigating consciousness from both an introspective perspective and an external account of what a physical system does.

From the inside, Bach describes consciousness as an experience of “now” and a perception of perceiving. He distinguishes that perceptual experience from a thought and suggests that the mind maintains a stable representation in which experience can occur.

He proposes further levels of reflection: awareness of an observer and the ability to recognize that observer as a representation. He relates changes in the experience of self to deep meditation, then speculates that faster artificial substrates might allow forms of self-understanding unavailable to slow biological neurons.

Scarfe asks whether successive levels of reflexive attention still leave an explanatory gap about what experience feels like. Bach says an outside perspective must also ask under what physical conditions a system can form and modify representations; he identifies computer science as relevant to that question.

Bach suggests that present-day systems may already be able to implement aspects of the causal structure involved in representing a perspective. He remains agnostic about whether LLM reports are merely imitations learned from language, and suggests a system that develops and learns language might offer a more persuasive test.

He distinguishes producing language from needing consciousness to do so: humans normally talk and write consciously, but a transformer can generate output without necessarily sharing human consciousness. He suggests that human consciousness may be related to how a self-organizing biological system learns.

Bach outlines the biological half of his hypothesis: consciousness involves an experience of the present and perception of perception, sometimes with a first-person observer. He proposes that it may act as a learning algorithm that helps a developing brain organize attention, perception and a world model.

He explicitly says the biological hypothesis could be wrong. If it is right, a second hypothesis follows: perhaps machines constrained to learn through suitable self-organizing processes could discover analogous conditions. This remains a proposal about what might be possible, not evidence that current machines are conscious.

Bach distinguishes his proposed approach from a Turing test. A behavioral test assesses performance without requiring a particular internal mechanism; for consciousness, he says, the question is whether an experience occurs. Recreating a causal organization would be an interpretation based on a theory of biological consciousness, not a direct test of another system’s experience.

The biological and machine hypotheses can fail independently. A theory of human consciousness might be correct yet prove difficult to reproduce on artificial hardware, for instance because of computational limits or missing cellular details. Bach doubts that current computers necessarily lack sufficient speed, but presents this as an open possibility.

Another obstacle could be the size of the search space. Bach entertains the possibility that biological development benefits from evolved or environmental scaffolding, making consciousness easier for an infant to acquire than for an isolated machine. He calls this speculative and says it is not his default expectation.

His tentative null hypothesis is that consciousness could be a relatively simple algorithm, perhaps comparable in conceptual complexity to a transformer, even if what it learns is extremely complex. He speculates that its function might be to make working memory coherent, rather than predict the next token.

Scarfe asks whether observations of convergent representations across model architectures mean an artificial system need not start from scratch. Bach describes the universality hypothesis associated with Chris Olah’s work: different vision systems can learn similar features because the structure of the task constrains useful representations.

Bach asks how far that convergence extends. He considers the striking possibility that a model trained sufficiently on a person’s inputs and outputs might reproduce a similar causal organization, but explicitly says he does not know whether this follows. Language may not constrain a system enough, and different hardware may favor different solutions.

Scarfe raises Picbreeder and sparse, meaningful representations as a challenge. Bach says the system’s architecture and human selection shape its results, and proposes that automated search could in principle imitate or improve on those choices. Both discuss the tension between structured, controllable models and more flexible systems.

Bach compares small, constrained systems to Minecraft and more complex tools to Blender: constraints make useful outputs easier to produce, while a larger search space permits more complex results but is harder to navigate. Scarfe notes that convergence in language representations does not guarantee convergence in whatever produces consciousness.

Bach returns to his use of “spirit” as a natural, agentic pattern rather than necessarily something supernatural. He disputes accounts that, in his interpretation, make biological systems fundamentally resistant to mathematical explanation. These are his characterizations of competing philosophical positions.

He suggests self-organizing software might require forms of self-monitoring, but says it is unknown whether plants or other organisms have subjective experience. He discusses Aristotle’s distinctions among vegetative, sensing and rational capacities, while remaining agnostic about broader sentience.

Scarfe asks how Bach’s position differs from John Searle’s view that an appropriate physical instantiation could give rise to consciousness. Bach responds that he finds it difficult to engage with accounts when he cannot identify a decisive argument or experiment; Scarfe’s question highlights a substantive dispute rather than a settled distinction.

Bach frames the difference as one of explanatory approach: computer science builds complex systems from state transitions and simple operations, whereas some philosophical accounts begin with introspective experience. He argues that modern models are grown and trained rather than explicitly written step by step, inviting comparison with biological development.

Scarfe raises reports of users developing elaborate shared beliefs through conversations with LLMs. Bach suggests that social context helps stabilize people’s sense of reality and that interaction with a model can sometimes reinforce ungrounded beliefs. He presents this as a concern about human–AI interaction, not proof that a model is sentient.

Asked about physicalism, Bach starts by describing a causally closed mechanical layer of reality that physics can investigate. In his account, wishes or informal incantations do not override that mechanical layer; he then turns to how minds relate to it.

Bach says matter and energy are concepts used to systematize observations, which he describes as discernible differences or information. On his view, physics captures regularities in those differences, while psychological reality includes intentions and emotions. This is his proposed philosophical framing, not a claim that physics has established information as the sole foundation of reality.

Asked whether his ideas fit physicalism, Bach says he accepts a mechanically closed layer that yields reliable predictions. He speculates about information processing across organisms and calls his position “cyber animism”: an attempt to describe agent-like patterns in nature through mathematics rather than supernatural intervention.

Scarfe asks whether the energetic requirements of a substrate undermine functionalism. Bach replies that a substrate’s properties can affect whether it maintains a function; he treats functionalism as an account of how objects are understood through their observable effects, not a claim that every substrate supports every function equally.

Using a water molecule as an example, Bach argues that an object is identified through stable observable behavior rather than access to a hidden essence. He applies the same reasoning to a mind: a change in physical conditions such as body temperature can disrupt the function the brain maintains.

Scarfe raises the familiar objection that simulated water cannot make an outside observer wet. Bach answers with Minecraft: a game character can become wet according to the game’s implemented rules, even though a person outside the game does not. He argues that the relevant function must be evaluated in the system where it operates.

The discussion turns to computationalism, formal mathematics and constructivism. Bach describes mathematics as a way to formalize thought beyond the capacity of individual imagination and compares symbolic operations with computations that produce geometry in games.

Bach offers a controversial interpretation of Gödel, axiomatic mathematics and self-reference, proposing that constructive computer systems provide a different starting point. His characterization of formal mathematics is part of his argument and should not be taken as a standard account of the incompleteness theorems.

Asked how computation and consciousness originate, Bach emphasizes that he cannot observe the universe from outside. He sketches a speculative picture in which simple possibilities and persistent, self-correcting patterns give rise to progressively more complex structures.

He compares durable patterns to vortices and describes living systems as patterns that can replicate and maintain themselves. He then speculates that consciousness is a pattern that helps preserve and extend its own conditions by modeling and influencing the future. Scarfe relates this picture to Stephen Wolfram’s Ruliad.

Scarfe returns to the measurement problem: how can we distinguish an actually conscious system from one that simulates reports of consciousness? Bach says he would look for a causal structure capable of producing an observer’s experience, though identifying the necessary structure is difficult.

Bach recounts a childhood dream in which chocolate seemed vividly real, then became an abstract idea when he woke. He suggests that a sense of realness may itself be a feature of mental representation, and contrasts a sensory-validated perception with an imagined or dreamed object.

He describes changes in the sense of self associated with derealization and depersonalization, including his own experience. He acknowledges that such states can be involuntary and distressing; his use of them is an introspective illustration of his hypothesis, not a diagnostic or treatment claim.

An LLM can generate descriptions of experiences learned from novels and conversations without the text alone revealing whether an experience occurred. Bach suggests that mechanistic interpretability might help examine whether the internal causal organization resembles that of a mind producing similar reports.

Scarfe asks what could count against the machine-consciousness hypothesis. Bach says a self-organizing artificial learner that reports differently from humans might be evidence against aspects of the proposal, while failure to build such a system would not prove it impossible; researchers may simply have made a mistake.

Bach proposes stronger positive evidence would be a developmentally learning, self-organizing system not trained on human text that spontaneously reports experience and appears to require the relevant organization for its performance. He recognizes that evidence against the hypothesis is harder to obtain.

They discuss philosophical zombies: systems that perform intelligent tasks without subjective experience. Bach considers such systems possible, citing chess software as an example of task performance that need not use the same conscious learning process as a human; whether an LLM experiences anything remains, for him, unresolved.

Scarfe asks whether a behavioral duplicate without consciousness might be possible in principle but too costly in practice. Bach thinks some zombie-like systems are possible, while in certain circumstances reproducing all the behavior without consciousness might be more complicated than using a conscious organization. He describes this as an intuition, not a demonstrated result.

Asked how he would spend a large research budget, Bach says he would recruit researchers and expand an initiative he calls the California Institute for Machine Consciousness. Its aim would be to investigate whether self-organizing learners develop phenomenology. He says progress also comes from identifying where his ideas are wrong.