When to Let the Model Draft

The rule under the rule

A human-in-the-loop practice is not "use AI more" or "use AI less." It is a running check on whose internal map is getting sharper. If the learner's model of the problem is growing, the tool is a partner. If the tool's output is being copied into the space where the learner's model was supposed to form, something has been short-circuited, no matter how good the paragraph reads.

That framing sits inside a larger working hypothesis. UNI is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails. The classroom version of that same posture is what this post is about.

Green light: when a draft helps

These are the situations where a first pass from a model tends to move a learner forward instead of past their own thinking.

• The blank page is the block. A learner who knows what they want to say but freezes at "how do I start" often just needs a lattice. A model-produced opening paragraph they will rewrite is a lattice. (Class E: aligned with well-documented findings on writing avoidance and cognitive activation costs in composition research; the tool lowers the activation cost, and the learner still does the shaping.)
• Format is unfamiliar, content is not. The learner has the substance but has never written a grant abstract, a lab report, a policy memo, a stakeholder email. A draft in the target genre makes the shape visible and lets the learner focus on what they actually know.
• Translation and register. Moving a strong idea from casual to formal, or from one language to another, is a job where the learner's internal model of the idea is not what is being tested. Let the tool draft; the learner audits meaning.
• Rubber-duck dialogue. The learner explains their reasoning to the model and watches where the model asks questions or gets lost. That reflected-back confusion is a signal about their own explanation.

Red light: when a draft short-circuits the build

These are the situations where a first pass from a model tends to replace the very work the learner was about to do.

• The assignment is the model-building itself. If the point of the exercise is for the learner to construct an argument, derive a proof, or trace a causal chain, a drafted output substitutes for the neural work the session was designed to prompt. (Class E: consistent with the broader literature on desirable difficulty and generation effects in learning science.)
• The learner does not yet know what "wrong" looks like in this domain. A confident-sounding draft in a topic the learner cannot yet critique becomes ground truth in their head. This is the failure mode that trauma-informed practice takes most seriously: young or overwhelmed learners often do not push back on an authoritative-sounding voice, so the tool becomes the authority by default.
• Emotional processing is the point. Journaling, apology letters, difficult conversations with a family member, first-person reflection on a hard experience. Drafting these with a model can produce fluent text that the learner never actually owned. In a trauma-informed, non-clinical setting the risk is not that the output is bad; the risk is that the learner does not do the interior work the writing was there to hold.
• The falsifier is missing. If neither the learner nor the teacher can articulate what a wrong answer would look like for this task, a model's draft cannot be safely evaluated. Withhold drafting until the falsifier is written down. (Class F: this is the practical face of a falsifier-present discipline.)

A three-question check

Before letting a model draft, run three questions with the learner.

1. What is the internal model this session is supposed to grow in you? Say it out loud, in one sentence.
2. If the model's draft is confidently wrong, will you notice? Name the specific thing you would check.
3. After the draft, what is the concrete rewriting step that keeps your hands on the meaning?

If any of the three cannot be answered, the drafting step is premature. That is not a punishment; it is diagnostic. The gap the questions expose is the actual next teaching move.

Field-level context

For context on why the framing of what these systems are matters at a field level, see Themesis, "The AGI Landscape Just Changed." In our voice, one line: the surrounding discourse is shifting fast enough that classroom practice needs a stable, honest way to talk about tools without hitching to hype.

Where this goes next

• Return to the cluster: AI as Partner, Not Replacement
• Read the companion: When to Put the Model Down
• See the workshop where we run this practice with a live cohort