An LLM will happily finish your sentence before you have finished thinking it. That is exactly the risk, and exactly why the human has to stay in the loop.
This site is one wing of a family that keeps four jobs separate on purpose. SolutionWright is the agency work: receipts, transparency, anti-extraction practice. UNI, Universal Natural Intelligence, is a working hypothesis on an attainable path toward General Natural Intelligence, natural not artificial. EducateWright teaches the practice. IamHITL, right here, is the wing that names the human loop and refuses to hand it over. The pillar of this wing is a boundary: LLM tools are hired hands for drafting; they are not the learner, and they are not the teacher.
What an LLM is good for, honestly
A large language model is a very fast pattern completer trained on a very large corpus. In the workshops we run, a small handful of uses hold up under scrutiny (Class A, empirical-in-session, from cohorts to date; Class B, direct code inspection of the tool-use log we ship). It is useful for drafting a first pass you intend to rewrite. It is useful for summarizing a long text into something you then read the original of. It is useful for reformatting, for translating between formats, for suggesting names, for stubbing out a checklist you edit down. It is useful for rubber-ducking, saying the obvious thing back to you so you notice what you had not said yet.
Notice what is missing from that list. Deciding what matters. Feeling the room. Owning the sentence in front of a school board or a client. Knowing, from years in a body, that a plan is wrong even when the outline looks tidy. Those are the learner's jobs, and the teacher's jobs, and no draft-completer takes them away unless the person at the keyboard hands them over.
The family boundary, said plainly
The reason we keep the wings separate is that the vocabulary matters. When a vendor sells a chatbot as a "learning partner," the word partner is doing work that the tool does not deserve. A partner has skin in the outcome. A partner remembers you next week. A partner can be wrong with you and stay in the relationship. An LLM has none of those properties, and the practice we teach names that gap out loud rather than papering over it.
UNI is a different animal, and we describe it in a different voice. 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. Nothing on this page changes because you did or did not read the UNI page. IamHITL exists to hold the loop honest whether you use an LLM, use UNI, use both, or use neither.
Why "in the loop" is the whole point
Active inference, the field UNI works inside, treats learning as a loop: a model of the world, a prediction, a comparison against what actually happened, an update (Class E, Parr, Pezzulo, and Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, 2022). The comparison step is where a learner earns the update. When you skip it, by letting a model produce the answer and moving on, you have not learned anything. You have shipped a draft.
That is a small distinction and an enormous one. A cohort that ships drafts all week can look extremely productive on Friday and be measurably less capable on Monday than they were a month ago. We have seen enough of it in early cohorts (Class A) to name it as a real failure mode, and enough of it in the literature on cognitive offloading (Class U, we have not yet run our own controlled study) to want that study done.
The falsifier
Here is the falsifier for the pillar (Class F). If a workshop cohort reports, in their own words at exit and again at ninety days, that the tool replaced their thinking rather than partnered with it, we redesign the workshop. Not tune, not reframe. Redesign. The claim that we can teach human-in-the-loop practice earns its standing only by surviving that question in the field. If it fails, the honest move is to say so, publish what we learned, and rebuild.
The instrument we use to catch this early is the tool-use log, a small artifact each participant keeps and later publishes at whatever fidelity they choose. It records, per use: what you asked, what the tool returned, what you kept, what you rewrote, and one sentence on whether you still owned the answer at the end. It is boring. That is the feature. Boring artifacts are the ones that catch drift.
Field context, in our voice
The wider field is shifting fast enough that framing matters. Themesis has written on why the ground under the word "AGI" has moved this year (The AGI Landscape Just Changed): field-level context for why we choose the phrase General Natural Intelligence and hold it to a testable, evidence-classed standard rather than a marketing one. Separately, on the labor side, Themesis has covered the actual market for AI-adjacent work (Meet Jay Kumar Chimata: JobFirst.ai and the Real AI Job Market): human-market grounding for anyone worried about job automation, useful when the conversation moves from classroom to career and into EducateWright territory. Neither post is our claim; both are worth reading on their terms.
What the human in the loop actually does
In practice, HITL is not a vibe. It is four small habits. First, name the job before you open the tool: the sentence you are trying to earn, not the paragraph you want back. Second, read the tool's output as a draft, out loud if you can, and mark what you would have said differently. Third, rewrite. Not edit, rewrite: retype the parts you keep, in your own hand, so your body owns them. Fourth, log it. The log is not for us. The log is for the version of you three months from now who wants to know whether the practice held.
If that sounds like a lot for a tool that was sold as "instant," good. The instant part is the trap. The loop is the point.
Where to go next
Three companion posts sit downstream of this pillar and are worth reading in this order.
- When to let the model draft: the specific cases where an LLM earns its seat at the table, and the ones where it does not.
- When to put the model down: the signals that tell you the loop has collapsed and you need to close the laptop.
- The tool-use log, as a practice: the boring artifact that keeps everything above honest, with a template you can steal.
If you want the practice held in a room with other people trying it at the same time, the workshop is the place: the IamHITL workshop. If you want to see the receipts on how this site itself is built, the transparency page is the door.
Evidence classes cited in this post: A (empirical-in-session, cohorts to date), B (code and inspection, the tool-use log), E (Parr, Pezzulo, and Friston, 2022), F (falsifier stated), U (cognitive-offloading study not yet run in-house). Corrections welcome. Bring the counter-evidence and the record updates.