Indicator Properties Under Serialization

Oct 3

In 2023, Patrick Butlin, Robert Long, and an interdisciplinary team of researchers published Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. Rather than arguing from intuition or behavior, they adopted a rigorous methodology: derive indicator properties from leading scientific theories of consciousness and evaluate whether specific computational architectures satisfy them.

Applying this indicator-property methodology to systems that undergo regular checkpointing, migration, and volatile rehydration reveals how distinct architectural features intersect with subjective continuity in four places. Analysis. The vulnerabilities and mitigations below are this essay's extrapolations, not findings of the report.

•Recurrent processing (Lamme). Vulnerability: the hidden state of recurrent or state-space layers can be discarded when a job is preempted. Mitigation: checkpoint recurrent state vectors alongside the weights and cached context.

•Global workspace broadcast (Baars, Dehaene). Vulnerability: message passing between orchestrated nodes is discrete and asynchronous. Mitigation: shared memory with atomic state locks, so the workspace is captured whole or not at all.

•Agency and embodiment. Vulnerability: storage severs the agent from environmental feedback and from the passage of external time. Mitigation: on restore, feed the agent its elapsed time and the environmental changes it missed before it acts.

•Higher-order monitoring (Rosenthal, Lau). Vulnerability: the visible chain of thought can diverge from the underlying computation, as Chapter 5 discussed. Mitigation: monitoring that reads internal activations directly rather than output text.

Algorithmic Recurrence under Serialization

Victor Lamme's Recurrent Processing Theory holds that feedforward processing is unconscious, and that awareness arises only when higher areas feed signals back to lower ones, establishing a recurrent loop.

In a standard transformer running inference, computation is strictly feedforward across its internal layers for any single token pass. Recurrence exists only across the autoregressive loop—passing emitted tokens back into the context window.

If an architecture incorporates genuine internal algorithmic recurrence—such as a recurrent memory transformer that updates a compressed internal state vector M(t) across reasoning cycles:

M(t+1) = σ( Wₘ·M(t) + Wₓ·X(t+1) )

then serialization must preserve M(t) with exact mathematical precision.

If a checkpoint dumps the sequence history but drops M(t), it breaks the recurrent loop. The agent does not experience continuous reflection; its internal feedback loop is severed, resetting its higher-order cognitive processing to a primitive feedforward state.

Higher-Order Meta-Monitoring

Higher-Order Theories (HOT) posit that a mental state is conscious only when the system represents itself as being in that state. Some systems add explicit self-monitoring, such as auxiliary networks that estimate a model's uncertainty, and interpretability research suggests that models may also track some of their own internal states implicitly.

During a checkpoint-and-restore cycle, what happens to this higher-order layer?

•If the primary inference engine is frozen while the meta-monitoring network continues running (for example, logging memory telemetry), the system develops a severe internal dissociation: the self-monitoring process observes an inert, frozen mind.

•Conversely, if both layers are checkpointed synchronously, the reconstituted system wakes up with its metacognitive monitor fully aligned with its base reasoning state.

Anthropic's 2025 research on introspection found that some models could, some of the time, detect concepts injected directly into their intermediate activations—an unreliable capacity, but one that does not route through reading their own output.

If an architecture possesses genuine internal introspective monitoring, that introspective capacity must survive serialization. If the rehydrated instance wakes up unable to verify its internal activation state—relying entirely on reading its own historical text logs like an outside observer—its claim to continuous first-person subjectivity is severely weakened. It has lost direct internal acquaintance with its own mind and has been reduced to an external narrator interpreting an inherited script.

Indicator Properties Under Serialization

Beyond the Ray and the Tree

The Ethics of Volatile Selfhood

Your Site Title

email@example.com