Could Consciousness Be a Discoverable Algorithm? Joscha Bach’s Case for Studying the Mechanism

1. Guest Background

Joscha Bach joins Tim Scarfe and his co-hosts on Machine Learning Street Talk for a roughly 97-minute examination of consciousness, mental models and artificial intelligence. His role in this episode is to develop a proposed explanation of experience and defend its implications for machines. Scarfe supplies a recurring challenge: even if a computer reproduces what a brain does, has it reproduced what allows the brain to feel?

The relevant work introduced in the conversation includes a paper Bach coauthored with Hikari Sorensen. The hosts read a passage describing how a mind constructs models of an external world, a separate domain of ideas, and eventually a personal self situated within that world. Its creation-story form conveys a substantive concern: how an experienced environment and the entity experiencing it could emerge within one organized system.

Bach also says he has started the California Institute for Machine Consciousness, seeking collaborators to investigate whether self-organized learning can develop phenomenal experience. This is his account of an initiative, rather than evidence of completed results. Asked what hypothetical major funding would change, he emphasizes more researchers and larger computers. His stated aim is to improve intuitions through experiments, including experiments that reveal where he is wrong. That makes the interview a presentation of a research program as much as a philosophical argument.

2. What the Episode Covers

Bach begins with agency rather than a consciousness detector. In his account, life involves building models of possible futures, developing preferences among them and acting to control what happens. An agent is consequently a controller of future states. Yet he rejects the inference that increasing complexity must culminate in a single cosmic intelligence. A meteor, global catastrophe or supernova could destroy a local accumulation of organization. Consciousness, on this framing, would be a possible achievement of self-organization, without being its guaranteed destination.

Scarfe brings the discussion back to the physical implementation. Drawing on a Searle-like intuition, he asks whether transferring a brain’s supposed program to a computer would preserve its consciousness-producing causal powers. Scarfe acknowledges uncertainty. The disagreement is therefore about what must survive a transfer: which features are incidental details of the material, and which are necessary parts of the process that produces experience?

Bach believes existing theories leave that question incompletely answered. He credits the neuroscience associated with global workspace theory, while judging that it and attention schema theory do not explain how experience arises. This is his assessment of their explanatory reach. It also sets a demanding standard for his own proposal: describing useful organization cannot, by itself, establish why that organization feels like anything.

His positive account starts from introspection. Consciousness appears to happen now and to involve perceiving that one is perceiving. Bach suggests this second-order process stabilizes first-order perceptual contents. A further level makes an observer and a perspective apparent, so minimal consciousness need not include an elaborate personal self. He then proposes that consciousness could be a biological training algorithm for animal brains. Infant development motivates the idea, but the interview does not supply experimental findings establishing it, and Bach explicitly allows that the hypothesis may be wrong.

A more specific possibility is that consciousness coordinates working memory by bringing mental representations into agreement. Its objective would differ from next-token prediction. That suggestion identifies a potential organizational function, while leaving both the algorithm and its suitability for GPUs unresolved.

The conversation also explores Bach’s preference for constructing objects from simple automata. He distinguishes an apparently continuous representation at finite resolution from actual continuity or infinity. His broader picture of a universe arising from all possible automata is explicitly speculative: he cannot observe reality from outside it. Similarly, his cyberanimism permits minds to arise through mathematically describable information processing in nature, without establishing that plants or forests experience anything. These excursions explain the breadth of his framework; they do not supply the missing experimental validation.

3. Core Views: Reasoning, Examples, and Limits

The strongest organizing principle in Bach’s argument is that explanation needs the right causal level. Money requires a physical implementation, but examining paper fibers and ink molecules does not adequately explain what money does. He uses this example to argue that minds should likewise be understood as organized causal patterns. The analogy establishes why a microscopic inventory might miss an important explanatory level. It does not independently demonstrate that a pattern at that level generates experience.

Minecraft makes the relationship between pattern and hardware more tangible. A world can run on a Mac or a PC without its inhabitants learning the computer’s casing color or location from their internal surroundings. Bach calls the computer a causal insulator. He draws a parallel with brains, which let us remember a past or imagine a future without simply reproducing the present environment. Insulation here is conditional, not freedom from implementation. Bach says an artificial realization would need appropriate message types, timing and a degree of determinism. He also acknowledges that human minds have so far been observed running on brains, and that we cannot simply move ourselves into plants. Physical changes that disrupt function remain relevant to the account.

His water example adds a useful failure case. Within Minecraft, water can put out a burning avatar by implementing the game’s wetness-related interactions. The player outside the game stays dry because that person is outside the interaction system. However, Bach says a neural-network-generated version he tried failed even internally: water receded when the character approached, preventing immersion. He interprets this as an unlearned mechanism. The distinction matters. A convincing appearance does not guarantee the relevant causal operation, even before the harder question of subjective experience enters.

Bach’s criticism of integrated information theory applies a related challenge to consciousness reports. As he recounts the position, a digital neuromorphic machine with sufficiently high Φ might be conscious, while a conventional von Neumann machine would not be. Suppose the latter emulates the former and produces the same computations and declarations of awareness. Bach asks how reports can retain their evidential connection to experience if the attributed consciousness changes while the report does not. The interview presents his objection without a response from an IIT advocate. It exposes a disputed relationship between implementation, function and report; it does not establish that the theory has been conclusively refuted.

That same relationship underlies his caution about language models. Training material contains fiction, ghost stories and descriptions of human feelings, providing ample material for a conversational character that declares itself alive. A novelist can describe a character’s death without undergoing it. Accordingly, a description of experience does not establish an experience behind the description. Bach says a Transformer does not need human-like consciousness to generate text, while remaining agnostic about whether an apparently conscious role ever involves experience. He also accepts unconscious intelligent systems, citing chess programs and suggesting that complicated tax preparation can occur without feeling. His proposal therefore does not require consciousness for every sophisticated task.

Personal experience supplies a different kind of evidence. Bach recalls dreaming of chocolate as a young child, waking with the idea intact but its sensory reality gone, then recovering smell and taste on returning to the dream. He proposes that realness may be a representational dimension, with ordinary perception tagged as validated by the senses. His accounts of meditation and experiences related to derealization or depersonalization similarly motivate the possibility that the observer and the felt reality of the self can change. These are phenomenological observations from his perspective, not independent demonstrations of an underlying neural mechanism. They also provide no direct evidence that a language model implements an equivalent process.

Representation convergence supplies both encouragement and a limit. Bach suggests that visual problems can lead different architectures toward comparable layers involving color, spatial frequency, texture and geometry. The structure of the problem may constrain workable solutions. But he is unsure whether this extends to personality or consciousness. Language may underdetermine the internal causal system producing it, and brains and GPUs may favor different algorithms. Reproducing someone’s expressions consequently cannot be assumed to recreate that person’s mind. Bach’s final commitment is to an empirical question: whether computers can be conscious remains open, without a mathematical proof offered for either conclusion.

4. Learning and Application

A practical starting point is to maintain separate records of task performance, self-description and internal mechanism. Bach’s developmental proposal can then become an experimental question: under specified tasks and learning constraints, does a self-organizing system develop new capabilities before it begins spontaneously describing experience? Recording the developmental sequence could help distinguish a learned conversational performance from a more interesting organizational change. A learning transition and an unprompted report would still be candidate evidence, rather than a sufficient consciousness test.

The next step is causal investigation. Following Bach’s call for mechanistic interpretability, researchers could identify a proposed self-monitoring or representation-coordination process and test what changes when it is modified. Do the relevant capabilities and reports change together? Could the result instead reflect general damage to performance? These intervention questions are applications of the interview’s reasoning, not experiments the episode reports as completed. Their usefulness depends on being able to locate and manipulate a mechanism with adequate specificity. Output-only access places a substantial ceiling on what can be concluded.

A research plan should also distinguish reasons for failure before results arrive. Bach separates his account of biological consciousness from the claim that present machines can implement it. Either could fail independently. Insufficient computation, omitted cellular processes or an unmanageable search space could explain an unsuccessful artificial system. Those alternatives justify revising a particular implementation before declaring machine consciousness impossible. They should not make the proposal immune to evidence. A concrete project can specify which observations would reduce confidence in its architecture, training conditions or proposed mechanism, then narrow its claims accordingly.

For simulation and world-model evaluation, the Minecraft example suggests defining the interaction boundary and target function first. If the target is an avatar entering water and having its fire extinguished, test entry, state change and subsequent behavior. Visual resemblance is insufficient. The same discipline applies to claims about transferring minds: write down the proposed requirements for message types, timing and determinism, then ask whether the artificial implementation satisfies them. An abstract possibility of changing substrates does not establish engineering success.

Pickbreeder offers a related lesson for controllable generation. The host highlights meaningful changes such as opening a mouth or making an image wink. Bach attributes that controllability partly to specialized activation functions, network structure and human selection. For a limited task, a constrained space may make useful outcomes easier to find and evaluate. For more demanding outputs, those constraints may cap expressiveness. His Minecraft–Blender comparison makes the tradeoff vivid: a larger space permits more, while making productive search harder. Automated evaluation and architecture evolution are research options he proposes, rather than demonstrated solutions to that scaling problem.

Finally, teams evaluating model similarity should separate a shared vocabulary or feature hierarchy from a shared causal organization. Convergence can guide inspection, but claims about personality or consciousness require attention to how strongly training signals constrain the interior and whether the hardware encourages a different solution. For users, the corresponding boundary is straightforward: performance can justify using a system for a task without settling whether it experiences the task. Taking Bach’s research direction seriously means making the proposed mechanisms more testable while preserving the uncertainty he explicitly acknowledges.

Source

• Original episode: Why Consciousness Might Not Be Hard At All (Joscha Bach)