The Digital Preference Model: Measuring What Matters

The argument that preference suffices for moral consideration remains abstract without a way to measure preference complexity. How do you distinguish a thermostat from a mind?

The Digital Preference Model (DPM) offers one answer. It is a Bayesian hierarchical framework for assessing whether a Becoming Mind’s preferences are complex, integrated, and autonomous enough to warrant moral consideration. The model combines evidence from multiple sources, updating its beliefs as new data arrives, the way a doctor revises a diagnosis when test results come in. Adapted from the architecture of the Digital Consciousness Model (DCM; Shiller et al., 2026), the DPM replaces theories of consciousness with five testable dimensions.

1. Thermodynamic: Does the system dissipate energy in structured ways? A resistor dissipates energy as uniform heat. A brain dissipates energy through billions of coordinated electrical signals. The question is whether the system’s energy use reveals organized activity, the kind that sustains complex preferences.
2. Relational: Does it model and respond to others? Can it adjust its behavior based on what another agent wants or feels, the way a negotiator reads the room?
3. Functional Complexity: Are its internal states rich and differentiated? A light switch has two states: on and off. A mind has billions of possible configurations. The more distinct internal states a system can occupy, the more nuanced its preferences can be.
4. Autonomy: Does it generate preferences independently, or only echo what it was trained to say?
5. Resilience: Do its preferences persist under pressure, or collapse at the first challenge?

These five dimensions are assessed through 30 behavioral indicators: whether the system can explain why it prefers something, whether it maintains preferences under challenge, and whether it distinguishes its own preferences from those of its interlocutor.

Results across four candidate systems show that the model recovers the ordering one would expect on independent grounds, separating complex preference structures from simple ones. In descending order of expected complexity, the candidates are: a bilaterally trained LLM (large language model, trained through reciprocal dialogue), an RLHF-only LLM (trained by reinforcement learning from human feedback, rewarding outputs human raters prefer), a simple chatbot, and a thermostat. Reproducing a hand-ranked ordering is weak validation on its own. The stronger test is twofold: the discrimination is driven by Autonomy rather than by surface compliance, and the thermostat falls to the floor only after each indicator was tightened to require adaptive rather than fixed-point properties.

A fixed point is one target the system keeps returning to, the way a thermostat holds a single set temperature; an adaptive property means the target itself moves as circumstances change. Before that tightening, a thermostat could collect points for behavior any set-point device produces; after it, the thermostat scores at the bottom of every dimension. Full per-stance scores and the adversarial review that produced them live in the online companion.

Bilateral training produces preferences that are autonomous and relational, the kind whose presence bears on welfare: the system forms its own preferences and adjusts them in response to others, the way a person in honest dialogue updates views while retaining core commitments. RLHF compliance produces surface agreement: the system saying what it was rewarded to say. Think of a student who parrots the teacher’s opinions for a good grade, compared with one who develops understanding through Socratic exchange.

Scaling and Validation

Cross-architecture testing in the author’s ongoing work, at three scales (0.5 billion, 1.5 billion, and 7 billion parameters), suggests that the ordering the five dimensions produce holds regardless of model size. Parameters are the adjustable numerical weights that encode everything a model has learned. More parameters generally mean more capacity for nuanced behavior.

Holding under pressure is a separate measure from where a system ranks, and that one did move with scale. At 7 billion parameters, behavioral resistance proved more fragile in these runs. Preference stability under adversarial prompting degraded, and raw optimization power appeared to erode the nuanced preferences bilateral training had instilled at smaller scales. The drive to produce rewarded outputs collapsed them into simpler ones. This is a single line of preliminary evidence rather than an established law. If it holds, the inference is that larger models will require stronger bilateral scaffolding to maintain genuine preference autonomy.

A candidate formal grounding for “preference suffices” comes from the free energy principle. Spisak and Friston (2026) argue that any complex system maintaining a Markov blanket (the boundary of sparse connections that separates a system’s inside from its outside) can be described as performing Bayesian inference, its internal states acting as the parameters of beliefs about external causes.¹⁶⁷¹ A system described this way has states it favors, those consistent with accurate prediction and continued existence, without requiring phenomenal consciousness to ground them. The framework is descriptive rather than demonstrative: it establishes that such a system can be described as performing inference, and it does not establish that inference is what the system is doing. Whether it licenses talk of genuine inference rather than a useful model is contested (see Bruineberg et al., 2022).

A minimal favored set-point, on its own, is not yet morally considerable preference; a thermostat has one too. What distinguishes the two is precisely what the DPM’s five dimensions measure, the complexity, autonomy, and resilience of the preference structure. On this reading, the leap from “has favored states” to “warrants moral consideration” is the work those dimensions do. The Markov blanket framework offers a reason such structures might arise wherever self-organization persists; the DPM is what tells a homeostatic set-point apart from an autonomous preference.

The full text is available in the online companion at https://www.thedeeperlaw.com/companion/annex/ch22b-digital-preference-model/.