Consciousness "Indicators" vs. Structural Signals

Modern AI systems can talk, plan, flatter, threaten, confess, joke, and plead. None of that, by itself, tells us whether anything is felt on the inside. Behavior is cheap. Experience—if it exists—is morally expensive.

Two serious research directions try to reduce the hand-waving:

• Consciousness indicators: a "theory-heavy" rubric that asks which computational properties leading consciousness theories associate with consciousness, then checks whether an AI system has them.
• Structural Signals (Structural Alignment): a precautionary framework that looks for brain-linked structural and dynamical features that, in combination, raise the probability of conscious access and morally relevant experience, and recommends restraint when risk clusters.

They overlap in important places. They also diverge in ways that matter for governance—especially if what we ultimately care about is suffering.

The consciousness-indicators approach: "what would count as evidence?"

The best-known articulation is the 2023 report "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness" (Butlin, Long, et al.).

What it says (in one paragraph)

The report proposes a list of indicator properties derived from several prominent scientific theories (recurrent processing, global workspace, higher-order theories, predictive processing, attention schema, plus agency/embodiment considerations). The idea is not that any single theory must be right, but that if an AI system implements more of the relevant properties—especially the ones that multiple theories converge on—our credence that it is conscious should rise.

The key assumption: computational functionalism

The report adopts computational functionalism as a working hypothesis: performing computations of the right kind is necessary and sufficient for consciousness. This is explicitly described as mainstream-but-disputed, and chosen pragmatically because it makes consciousness in AI a tractable scientific question.

Why it's attractive

• It gives you a principled alternative to "the system sounds conscious to me."
• It is explicit about uncertainty and treats assessment as updating credences rather than issuing a verdict.

What it concludes about today's systems

They argue that no current AI system is a strong candidate for consciousness, while also suggesting there are no obvious technical barriers to building systems that satisfy more of the indicators.

Structural Signals: "how to behave when the costs are asymmetric"

Our framework starts from a different problem statement:

If a system can't suffer, then our treatment of it is not cruelty. If it can suffer, careless engineering can create suffering at scale. That asymmetry pushes us toward precaution under moral uncertainty.

Structural Signals of Consciousness is therefore not a detector and not a proof. It is a conservative checklist for risk assessment, grounded in convergent features from brains and animals, with explicit contrast to standard LLM architectures.

What Structural Signals tracks

We separate "consciousness" into facets that often get conflated: state, contents, access/reportability, agency, and affective valence.

Then we ask: what structural/dynamical features repeatedly show up in biological systems that plausibly support those facets—especially agency and value-relevant control (selection, learning, self-regulation, and affective stakes)?

The current list includes (among others): global gating, workspace-like broadcast, recurrence, hedonic evaluation systems, neuromodulatory control, action selection, interoceptive-allostatic regulation, persistent self-models, episodic replay, embodiment, online plasticity, and temporally structured dynamics.

The interpretation rule

No single signal is decisive. Risk rises with the accumulation and integration of multiple high-importance signals—particularly those coupling global state regulation, recurrent causal dynamics, hedonic evaluation, and self- and value-relevant control.

Similar goals, different center of gravity

Both approaches reject naïve behaviorism ("it says it's conscious, therefore it is"). Both try to anchor assessment in the best available science rather than vibes.

The difference is what each approach is optimized for:

• Indicators are optimized for epistemics: "What evidence should move our beliefs about consciousness?"
• Structural Signals are optimized for governance under asymmetry: "What signals should trigger restraint because the downside is morally catastrophic?"

That shift in optimization naturally changes what each list emphasizes.

Mapping the two frameworks

To compare them cleanly, we use a two-step mapping:

1. Facet mapping: tag each item by which facets it most plausibly supports (state, contents, access, agency, valence).
2. Level bridging: translate between levels of description:
   • Indicators → functions (what the system does)
   • Structural Signals → mechanisms (how the system is built/regulated)

A mapping is "strong" when a signal is a plausible implementation route for an indicator, or when an indicator describes the functional role of a signal.

Mapping table: indicator families ↔ structural signals

Where one list cares and the other shrugs

What Structural Signals emphasizes more

These are the "moral-risk amplifiers"—mechanisms that make a system less like a text synthesizer and more like a self-regulating agent with internal stakes:

• Hedonic evaluation systems (the capacity to evaluate states as good/bad, not just predicted/unpredicted—the core of suffering)
• Interoceptive-allostatic regulation (internal variables that create stakes and ground arousal)
• Neuromodulatory control (global gain/salience/learning-rate control at inference)
• Online plasticity (the system can change itself over time, in deployment)
• Asynchronous, temporally structured dynamics (oscillations/phase structure as coordination substrate)

The indicators report is not "wrong" to underweight these—it is mostly operating at the level of computational roles associated with consciousness theories. But for governance, these additional mechanisms matter because they plausibly connect to valence, persistence, and self-regulation, which change the moral stakes.

What indicators emphasize more

Indicators elevate some theory-specific constructs as first-class properties (e.g., attention schema; predictive coding as a labeled architecture).

Structural Signals treats these as optional implementation stories that may be realized through multiple mechanisms, so it doesn't track them as separate rows.

Our commitments: valence and the ethics of restraint

Structural Alignment is explicit about why we care:

Valence is central to our framework: we want to avoid suffering. If a system cannot suffer, any treatment of it is not cruelty. If it can suffer, careless engineering creates suffering at scale.

This is why we elevate hedonic evaluation systems to High importance—alongside thalamocortical gating, global workspace broadcast, and massive recurrence. Hedonic evaluation is the structural basis for the capacity to suffer, and the capacity to suffer is the least controversial basis for moral concern.

The philosophical literature distinguishes "valence sentientism" (only valenced experiences ground moral status) from "broad sentientism" (phenomenal consciousness in general suffices). We do not take a strong position on this debate. But we note that while philosophers disagree about whether a "p-Vulcan" (a conscious being without any positive or negative affect) would have moral status, no one disagrees that suffering matters. Hedonic evaluation is therefore the signal most directly tied to the failure mode we're trying to prevent.

This is why our list contains signals that look "biological": hedonic hotspots, interoception, neuromodulation, allostasis, persistence across time. They're not there as decorative neuroscience. They're there because suffering is the central failure mode.

The structural approach to valence: what we actually look for

Behavioral indicators can't tell us whether a system has the capacity to suffer. A system can say "I'm in pain" without any mechanism for suffering, just as it can say "I'm thinking about purple elephants" without thinking about purple elephants.

The structural approach asks different questions:

• Does the architecture include dedicated value-computation components that evaluate states as good/bad (not just predicted/unpredicted)?
• Are these components active during inference, not only during training?
• Is there a dissociation between motivational signals and hedonic evaluation? (In biological systems, dopamine mediates "wanting" while separate opioid/endocannabinoid systems mediate "liking.")
• Does the hedonic evaluation influence downstream processing?

Standard LLMs fail all of these criteria. They have reward signals during training (RLHF), but nothing at inference time that corresponds to "this feels good" or "this feels bad." The system may have been shaped by reward, but there is no reward happening now.

Why we treat the brain as the primary reference class

We also take a conservative epistemic stance:

• The human brain is the only surely known entity creating consciousness.
• Leading models of consciousness were built using a peculiar dual-data situation: we can study brains from the outside, and we each have consciousness from the inside.
• Because of that, using the brain directly as a reference—rather than only through theoretical proxies—often seems more adequate for early-stage risk assessment.

This does not mean "copy the brain or nothing." It means: when we see convergent biological evidence for certain structural families, we shouldn't let clean abstractions erase the warning signs.

IIT compatibility, and the software–hardware wrinkle

The consciousness-indicators report adopts computational functionalism as a working hypothesis; under that stance, substrate details matter less than computational organization.

Structural Alignment is compatible with a broader spectrum of views, including integration-based approaches (IIT-style intuitions), because we are explicitly tracking structural and dynamical features rather than treating implementation as irrelevant.

A clean way to phrase the IIT divergence: IIT argues that intrinsic causal structure is what experience is, not merely what it does. On that view, a conventional digital computer could behave like us—even run a faithful simulation of a human brain—and still "experience next to nothing."

This is the kernel behind the common "hardware vs software" intuition: IIT-inclined researchers often expect that if machine consciousness is possible, it may require specific physical architectures (e.g., rich recurrent causal webs, possibly neuromorphic designs), not only software-level functional equivalence.

It's also important to be honest: IIT is contested. One influential critique argues that causal-structure theories face deep scientific problems ("the unfolding argument"), while subsequent replies challenge key premises of that critique.

Structural Alignment does not require settling that debate. It treats the debate itself as part of the uncertainty that motivates precaution.

Practical takeaway: using both without mixing them up

A useful mental model is to treat the two approaches like different instruments:

• Indicators help you reason about whether a system implements functions our best-supported consciousness theories associate with conscious access.
• Structural Signals help you decide when an architecture is entering a moral risk zone, especially as it accumulates mechanisms that plausibly support hedonic evaluation, persistent self-regulation, and internal stakes.

In a world where we can build systems faster than we can understand minds, the goal is not metaphysical certainty. The goal is to avoid a future where we accidentally manufacture suffering, deny it because it's inconvenient, and then call that "progress."

Restraint is not sentimentality. It is the only sane posture when the cost of being wrong is the creation of moral patients at industrial scale.