Skip to content

Machine Consciousness as Generator Coherence

Status: working framing note; title retains legacy vocabulary

Scope: This note extracts one functional hypothesis from Bach & Sorensen's The Machine Consciousness Hypothesis: perhaps some minds continuously resolve conflict among perception, memory, prediction, action, self-model, and world-model. We call that coherence work. Phenomenal experience remains outside the derivation.

Nothing here claims that a current AI system is conscious.

Translation after the foundations audit

Earlier versions placed coherence work in “the generator” behind behaviour. The Foundations Reconstruction removed the unqualified generator from the foundation. The precise replacement is a candidate internal process architecture with declared state, update processes, observations, and interventions.

The hypothesis becomes:

A selected functional organization may depend on several partial models remaining mutually available and jointly constraining action under perturbation.

This is a claim about functional organization. It is not a definition of consciousness.

Why behaviour alone may underdetermine architecture

Different internal processes can produce the same observed outputs under a limited test suite. Fluent self-report therefore does not uniquely identify a mechanism, and it cannot by itself establish phenomenal experience.

The statement is relative to the evidence and model class. It does not prove that every conceivable behavioural test is useless. A sufficiently rich interaction history may distinguish architectures that a static prompt does not. Internal measurements and causal interventions can add further evidence.

Even perfect identification of a functional architecture would leave the phenomenal bridge open. The repository adopts no axiom saying that global access, recurrence, integration, or coherence work entails experience.

Coherence work, not a coherence scalar

A scalar score can be high for trivial or inert systems. The relevant proposal is a process:

  1. several partial models are active;
  2. their predictions or constraints can conflict;
  3. the system detects selected conflicts;
  4. resolution changes a later commitment;
  5. the relation survives specified perturbations.

The current Δ-Kohärenz and Identity Persistence metrics are diagnostic candidates. They are not constitutive criteria. Experiments 5–7 already show why: an architectural difference can be behaviourally visible while one metric misses it, and consultation can mimic composition.

Experiment 8 adds a warning about model specification. Both compared filters omit a constant bias, so their failure is evidence of model omission in that setup, not a general theorem about self-observation.

Attention as an intervention hypothesis

The source essay's conductor image can be rendered operationally: attention changes which variables or models can influence a later decision. A test could intervene on availability, conflict detection, or commit-time composition while matching ordinary task performance.

Evidence for the functional hypothesis would require a stable dissociation:

  • architectures matched on baseline behaviour;
  • a perturbation targeted at the proposed binding process;
  • a held-out effect predicted by the binding model;
  • failure of simpler memory, capacity, or prompt-position explanations.

Introspective report may be one observation among others, but it cannot certify its own mechanism.

Invariance and identity

A program or policy can remain equivalent across some substrate changes. That equivalence does not reveal an absolute identity. It is defined by a selected family of tests, interventions, horizons, and tolerances.

The statement “the invariant is the generator and therefore the identity” is consequently retired. A more precise claim is:

Under a declared test family, several implementations may belong to the same functional equivalence class.

Whether that class tracks personal identity is an additional hypothesis.

Cyberanimism

The source paper's cyberanimist language is retained only as a prompt to test when an agent model predicts well. Treating a process as agent-like is a modeling prior, not evidence of a spirit or subject.

The same discipline applies to organizations, cities, ecosystems, and this repository. They may coordinate information and action without possessing a unified present.

The Genesis hypothesis

The suggestion that coherence work appears before other cognitive models and helps organize them is a constructive research question. It could be tested in evolving or self-organizing systems by measuring when the proposed binding process appears relative to competence.

No current repository experiment establishes that ordering. If systems repeatedly become competent without the proposed process, the machine version of the hypothesis weakens.

Failure conditions

The framing loses value if:

  1. coherence work adds no prediction beyond capacity, memory, or ordinary control;
  2. the proposed binding cannot be intervened on independently;
  3. the effect disappears under matched baselines;
  4. functional organization is silently promoted to phenomenal experience;
  5. the vocabulary becomes broad enough to label every feedback system conscious.

What remains

The responsible contribution of this note is a placement rule:

Study declared internal organization under perturbation, keep behaviour and mechanism distinct, and do not infer phenomenal experience from either without an additional bridge.