Step 4 of 11
Evaluative Agency Theory
The claim underneath the four layers: what you produce with AI and what you keep hold of while producing it are two different quantities, and only one of them compounds.
The previous step ended with a problem rather than an answer. If the results disagree because the engagement varies, then a framework that names the engagement has to say what it thinks is varying, and why that particular thing matters more than the others it could have picked.
Evaluative Agency Theory is that answer. It is one claim in three parts.
First, separability. Producing an artifact and evaluating it are different acts, and AI changes their relative cost by very different amounts. Production gets dramatically cheaper. Evaluation gets slightly cheaper at best, and in some conditions more expensive, because now you are evaluating something you did not build and cannot fully see inside. Second, locus. Of the two, evaluation is the one that improves the person doing it. Judgment is built by exercising judgment, and delegating production does not touch it, while delegating evaluation removes exactly the repetitions that would have built it. Third, cultivability. Evaluative agency is a disposition rather than a trait: it grows with practice under principles, and it decays without them. That growth is the loop the model calls Becoming, and it runs on a timescale no single conversation can show.
Put together: the useful question is not how much you delegated, but whether you kept authorship of the evaluation while delegating the production.
What the theory commits the model to
Each layer of the AI Engagement Model exists because the theory needs something at that position. The chain reads upward, and it closes with the loop most frameworks leave off.
Mechanism
If the person and the tool form one coupled system, then "did AI help?" compares that system to itself minus a part. The unit of analysis has to be the interaction.
Dispositions
Reliance on others is the ordinary condition of knowing, not a failure of it. So the durable tendency worth having is not independence. It is calibration.
Principles
Delegation is graded, not binary, and the level you pick has consequences you can reason about in advance. Principles are the standing policies for picking it.
Behaviors
Conversation has a turn structure, so what a person did is recoverable from the record. The eight modes are characterizations of runs of turns, which is what makes them codable.
Outcomes
What the engagement strengthened or weakened: learning, judgment, agency, voice. This is the layer the theory makes claims about and the research program has to earn.
Becoming — the loop back, not a sixth layer
Across many engagements, repeated reflective behavior can strengthen the dispositions the chain started from. That return arrow is what the model calls Becoming. It is a loop over time, not a further layer, and no single conversation, mode count or score demonstrates that it has happened.
What it is built on
None of the three claims is new on its own. What is new is the arrangement. Each source below established something in its own field; the second sentence in each entry says precisely what that licenses in this model, so you can see where the borrowing stops.
Supports Mechanism
The extended mind
What it established. An external resource that is reliably available, easily accessed, and automatically trusted can be part of a cognitive process rather than merely an input to it.
What it licenses here. The mechanism layer's insistence that the unit of analysis is the person-and-tool system. It does not license the stronger metaphysical reading of the thesis, which remains contested; the model uses the three criteria as a description of when coupling is tight, and the third one — automatic trust — is exactly what evaluative agency withholds.
Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. Source
Mixed-initiative interaction
What it established. The useful designs are not full automation but a negotiated division of labor, in which either party can take the initiative and the system reasons about the cost of acting versus waiting.
What it licenses here. Reading an interaction as co-produced, and reading the modes as answers to who took the initiative for what. It is a design principle for interfaces, not an empirical finding about learning, and nothing about outcomes follows from it.
Horvitz, E. (1999). Principles of mixed-initiative user interfaces. Proceedings of CHI '99, 159–166. Source
Supports Dispositions
Epistemic dependence
What it established. Much of what any person rationally believes rests on the testimony of others whose reasons they cannot personally check. Dependence is the normal condition of knowing, not a lapse in rigor.
What it licenses here. The model's refusal to treat reliance as a deficit, and its choice of calibration rather than independence as the disposition that matters. Hardwig is arguing about human experts, who can be held responsible in ways a model cannot; the model borrows the structure of the argument, not a verdict about AI.
Hardwig, J. (1985). Epistemic dependence. The Journal of Philosophy, 82(7), 335. Source
Supports Principles
Metacognitive monitoring and control
What it established. Cognition has two levels. An object level does the work, and a meta level monitors it and issues control decisions back down. Monitoring and control are separate functions and can fail separately.
What it licenses here. The claim that engagement is controllable rather than merely observable, and the shape of a principle as a standing control policy. The separation also explains a specific failure: a person can monitor an AI answer accurately and still not act on it, and the model has to treat that as different from not noticing.
Nelson, T. O., & Narens, L. (1990). Metamemory: A theoretical framework and new findings. Psychology of Learning and Motivation, 26, 125–173. Source
Types and levels of automation
What it established. Automation is not on or off. It applies in degrees, and independently across information acquisition, analysis, decision selection, and action, with consequences for the operator that differ by stage.
What it licenses here. Calibrated trust as a graded choice rather than a binary one, and the claim that the same tool used at a different level is a different engagement. It is also why the modes are functions rather than a ladder: a high level of automation on a settled factual question is not a worse choice than a low one.
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics — Part A, 30(3), 286–297. Source
Supports Behaviors
The organization of turn-taking
What it established. Conversation is locally managed through a describable turn-taking system, which makes what participants are doing analyzable from the record of the talk itself rather than from what they say about it afterwards.
What it licenses here. Coding conversations into modes at all. The turn is the unit and a mode is a characterization of a run of turns. What it does not license is treating a mode label as a fact about the person; it is a description of an episode.
Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696–735. Source
Language use as joint action
What it established. Language use is joint action built on common ground: what participants take to be mutually established shapes what any given utterance accomplishes.
What it licenses here. Why the same words are different behaviors depending on what has already been established, and why classification reads runs of turns rather than isolated prompts. A single prompt read out of context is not enough to name a mode.
Clark, H. H. (1996). Using Language. Cambridge University Press. Source
Supports Becoming
The reflective practitioner
What it established. Competent professionals think in action and reflect on action afterwards, and it is that reflection, not the accumulation of experience by itself, that turns practice into expertise.
What it licenses here. The Becoming loop. Doing a thing repeatedly does not by itself change the person doing it; reflecting on it is the mechanism, which is the reason the model is teachable rather than merely descriptive. Schön was writing about human practice without AI, and whether the same conversion works when the practice is mediated is an open empirical question, not a settled one.
Schön, D. A. (1983). The Reflective Practitioner: How Professionals Think in Action. Basic Books. (Routledge reissue, 2017.) Source
Quick check
Evaluative Agency Theory says delegating production is usually fine. What does it say is costly?
What would count as evidence against it. Three results would be genuinely hard to absorb.
If people who delegate evaluation as well as production show the same downstream capability, over a meaningful delay, as people who kept the evaluation, then the locus claim is wrong and the framework is measuring something that does not matter. If forcing evaluation — through reflection, verification, or having to defend a choice — reliably fails to narrow the gap, then the cultivability claim is wrong and evaluative agency is closer to a trait than a disposition. And if mode patterns turn out to be stable across tasks and contexts for the same person, that would favor reading them as personality after all, against the model's explicit claim that they are functions.
Each of those is measurable. None of them is settled. The next step lays out the four layers in full, and the one after it reports what the research has actually found so far, including the parts that do not help.