Step 5 of 11

The four layers, and what they contain

Mechanism, dispositions, principles, behaviors, and the outcomes they move.

Each layer answers a different question, and the previous step says why these are the questions. Mechanism asks what kind of system this is. Dispositions ask what you bring. Principles ask what you hold to. Behaviors ask what you actually did. Outcomes ask what it changed.

The eight behaviors are the layer you can observe directly, which is why they carry most of the measurement. The other layers explain why two people behave differently and how a person's behavior changes.

Expand any element to read its full definition.

The AI Engagement Model as a formation tree. The surrounding conditions represent the co-production mechanism in which person, AI tool, task, and context shape the interaction. Roots represent dispositions, the trunk represents principles, branches represent eight observable engagement behaviors, and five fruits represent Learning, Judgment, Agency, Voice, and the carefully bounded long-run outcome of Flourishing.

The formation pathway — click a layer

groundsenableshapeproduce

Mechanism

Select an element to see what it means.

Roots are dispositions, the trunk is principles, the branches are the eight modes, the fruit are outcomes — each layer grows from the one beneath it.

Layer 1

Layer 1 — Mechanism

What kind of system you're working with

Layer 1 specifies the operating logic of the technology. Generative AI exhibits an asymmetry no prior productivity technology shares at this scale: a short prompt (the local rule) produces an output (the global pattern) whose richness and consequentiality is disproportionate to the input. The output is co-produced in the strict sense — neither the human nor the model produces it alone. This claim is anchored in three converging traditions: complex adaptive systems (Holland, 1992; Kauffman, 1993), the extended-mind thesis (Clark & Chalmers, 1998), and Cognitive Continuum Theory (Hammond, 1988, 1996). The novel contribution is not that emergence describes AI — that observation is trivial — but that the user's recognition of themselves as the local-rule author of an emergent system is the foundational mental shift required for fluent engagement, and that this shift can be deliberately cultivated. Four competing mental models (substitution, augmentation, authority, companion) each produce predictable failure modes. The framework's diagnostic move is to surface which mental model the reader is currently operating with.

Layer 4

Layer 4 — Dispositions

Five qualities that make the principles livable

Layer 4 names the five dispositional grounds that make the principles inhabitable: Intellectual Humility, Tinkering Disposition, Tolerance for Ambiguity, Internal Locus of Evaluation, and Investment in Process (Care + Patience merged). Principles can be declared but not installed by declaration; the dispositions determine whether a person can live the principles or merely recite them. The dispositions split across the conative/cognitive divide: Tinkering and Investment in Process are conative-dominant; Intellectual Humility and Tolerance for Ambiguity are cognitive-dominant; Internal Locus of Evaluation is hybrid. This positioning lets the framework draw on Mischel & Shoda's Cognitive Affective Personality System (1995) as the theoretical anchor for an integrated dispositional account. Intellectual humility draws on established work (Leary et al., 2017; Porter et al., 2022; Whitcomb et al., 2017); the tinkering disposition draws on the joyous-exploration facet of curiosity (Kashdan, 2018) and on tinkering pedagogy (Resnick, 2017).

Layer 3

Layer 3 — Principles

Six postures that distinguish good engagement

Layer 3 names six postures that distinguish good engagement from merely efficient engagement: Authorship, Productive Friction, Calibrated Trust, Dialogue (not Transaction), Voice Preservation, and Reflection. A user can perform any mode with or without these principles in operation. Production Assistant with Authorship and Voice Preservation produces an artifact that retains the author's signature; Production Assistant without them produces a homogenized AI default. Same mode, different quality. Each principle is anchored in established literature (e.g., Calibrated Trust in Lee & See, 2004; Productive Friction in Bjork's desirable difficulties; Reflection in Schön and the metacognition tradition). The framework's contribution is integrative: showing how the six function as a coherent set, how they support each other, and how they depend on the underlying dispositions. The principles are assessed through a mix of self-report measures and behavior-pattern indicators derived from conversation analysis.

Layer 2

Layer 2 — Behaviors

The eight engagement modes

Layer 2 is the observable surface of AI engagement. The eight modes are defined behaviorally — by what the user does in their messages — and cluster into three epistemic tiers along the intuitive-analytic continuum (Hammond, 1996): Passivity (Oracle, Production Assistant), Partnership (Tutor, Collaborative Problem-Solver), and Agency (Verification Agent, Creative Expander, Critical Challenger, Problem Setter). A central idea is that passive and agentic use may be governed by different cognitive channels: roughly, the calibrated judgment of when to trust AI output on the passive side, and awareness of one's own thinking on the agentic side. The modes are constructed to map onto Hammond's continuum and to cover Bloom's revised taxonomy from Remember through Create at the meta level. Mode fluency — selecting the mode that fits the task and recognizing drift — is the practical operationalization of this layer.

What good engagement yields

Learning
Retention and transfer beyond the immediate task.
Judgment
Calibrated trust and pattern recognition that compound over time.
Agency
Sustained capacity to drive your own thinking with or without AI.
Voice
Work that sounds like you over a longer time horizon.
Flourishing
Engagement that preserves motivation and meaning, not just productivity.

Quick check

Which layer can be recovered from a transcript of someone's conversation?