The Human Stance: What “Human in the Loop” Should Actually Mean
The Master, the Emissary, and the Machine: How AI is completing a centuries-long drift in thought, and how to keep thinking well alongside it. Article 5/5.
“I do not underestimate the importance of the left hemisphere’s contribution to all that humankind has achieved, and to all that we are, in the everyday sense of the word; in fact it is because I value it, that I say that it has to find its proper place, so as to fulfil its critically important role. It is a wonderful servant, but a very poor master.”
Iain McGilchrist, The Master and His Emissary, p. 437
The series:
1. The Diagnosis: AI as the Emissary’s Triumph
2. The Structure: Decontextualization and the Loss of the Whole
3. The Unsaid: Language-Dependence and the Loss of the Tacit
4. The Author: Misplaced Confidence and the Loss of Accountability
5. The Human Stance: What “Human in the Loop” Should Actually Mean (this article)
This is the article that tries to chart a way forward as we continue with AI.
A brief recap, for readers arriving here from article 1/5.
The series has argued that AI is the most powerful instance yet of a particular mode of thinking that has been quietly dominant in our culture for centuries. Iain McGilchrist calls it the left-hemisphere mode. It works by stripping context, by operating only on what can be articulated, by producing confident outputs without the answerability that would let those outputs be checked against the world. Three traits explain how AI is left-hemisphere thinking run wild: decontextualization, language-dependence, misplaced confidence. Articles 2/5, 3/5, and 4/5 developed each one, with examples from my professional exchanges with AI (that readers likely have encountered themselves). The cumulative argument is that AI cannot fix these limitations from inside its own mode. The limitations are inherent to what AI is.
These limitations are not a bad thing. In fact, they are the opportunity for humans to leave their mark.
The two failures of the current discourse on human collaboration with AI
The current discourse on human-and-AI collaboration is lacking in at least two ways.
The first is thinness. A common word for the human contribution is “judgment.” We are told that AI does the heavy lifting and humans bring “judgment.” But judgment is a thin word for what humans actually have to do. It is a placeholder, not a description. It doesn’t tell you what to cultivate, what to protect, what to watch for in yourself. It is especially difficult to instill in junior contributors who are most at risk of lazy thinking attributable to AI use.
The second failure is positioning. The dominant frame for the human role is “human as a check,” which sits the human at the end of the AI’s work, inspecting outputs for failure. This is the frame baked into “human oversight,” “human review,” and much of the guidance organizations are currently writing about AI use. It positions the human as quality control: reactive, late-stage, watching for errors after the fact.
Both framings make the human’s role passive and opaque. Judgment is what you apply at the end; oversight is what you do at the end; review is what you do at the end. The human is the last filter before the work goes out. The AI is the source; the human is the check.
The whole argument of this series points the other way. If decontextualization, language-dependence, and misplaced confidence affect everything AI does, the human contribution cannot be a final filter. By then the work has already been done using an incomplete perspective. The human has to be present from the beginning, holding the perspective the AI cannot hold.
ACTS as the human stance, in action
McGilchrist gives us a richer vocabulary, and the right-hemisphere faculties described in articles 2/5, 3/5, and 4/5 give us a usable framework. I’ve summed them here using the acronym of ACTS:
• A — Attention. The kind of attention you bring determines the kind of world that shows up.
• C — Context. Holding the situated whole the system cannot see.
• T — Tacit. Bringing into the work what cannot be put into a prompt.
• S — Stake. The answerability to outcomes that gives advice its weight.
The human ACTS. The framework names a stance, an action, a way of attending.
Two paragraphs on what the ACTS stance represents, before I get into practice.
McGilchrist’s right hemisphere is, among other things, the seat of how we attend to the world. His central claim about attention is that it isn’t neutral. The kind of attention you bring to anything determines what shows up. Narrow, instrumental attention shows you a world of objects to be used. Open, vigilant attention shows you a world of situations, relationships, and contexts in which the objects are placed.
The human contribution to AI-assisted work is, in this sense, a way of attending that the AI cannot perform. The AI brings narrow, instrumental attention to the tokens in front of it. That is what it is for. The human brings the broad, situated attention to what those tokens mean in the world they came from, what is in this situation that wasn’t in the training distribution, what is not yet visible because it cannot be put into words. The stance is to keep that broader attention present throughout the work, holding the wider field that the AI cannot see. It is the master’s stance from McGilchrist’s fable. The emissary’s narrow attention has become the architecture of AI; the master’s stance is what stays with the human.
ACTS in practice: the individual
The most practical form ACTS takes is self-questions. Not prompts you give the AI, but questions you ask yourself while working with AI. Questions to the AI just relocate the work back to the AI, which is the problem we’re trying to solve. Self-questions keep the work on the human side, which is the point.
Questions such as:
Attention. Where is my attention actually shaped right now: by what I came here to do, or by the AI’s framing? Am I bringing attention, or outsourcing it?
Context. What about this client, moment, or situation is not in the training distribution? What is only legible from where I am sitting?
Tacit. What do I know about this that I cannot put in a prompt? Where does this output not quite ring true to me, even before I can say why?
Stake. Am I willing to be held accountable for this as if I had drafted it myself? Where am I leaning on the AI’s authority rather than my own answerability?
These are not a checklist. They are habits of awareness as you work. The practice is recognizing what you bring, not what you can extract. If you answer the attention question and notice you have been outsourcing attention, you stop. You read the AI’s output more slowly. You ask what you actually came here to do. If you answer the context question and notice the situation has features the training distribution didn’t, you make those features explicit (to yourself, not to the AI) before you decide how much weight to give the AI’s output.
There is a corollary to this practice that becomes clearer once you have started doing it. ACTS applies not just to how you work alongside AI but to which tasks you ask it to do in the first place. The highest-value AI uses I have found are not final products but highly-falsifiable research processes built around controlled datasets and specific questions I can review for myself. The AI does the legwork; the practitioner holds the question, the context, and the stake. Where the work in front of you can be structured that way, ACTS holds naturally. Where it cannot, the framework is asking you whether AI should be doing the work at all.
The faculties take years to develop in any practitioner. They are vulnerable to AI in a specific way: the AI is fast, the AI is fluent, the AI is always available, and the path of least resistance is to let the AI’s framing become yours. The thing you watch for, in yourself, is the slow ceding of these faculties. You watch for the moment you stopped reading the source document carefully because the AI summarized it. You watch for the moment you stopped trusting your own felt sense of a piece of work because the AI’s confident gloss made yours seem like an objection. You watch for the moment you let the AI’s draft be the version you defended, without quite noticing that you no longer felt able to say what it said with your own voice.
What you don’t delegate is the part of the work that can only be done by someone with stake. That is most of what makes the work yours.
ACTS in practice: the organization
The same framework changes shape at the organizational level.
Most organizations are absorbing AI as a productivity play. Efficiency, automation, cost reduction, throughput. It is itself a left-hemisphere framing, and the same set of problems follows from it as follows from any left-hemisphere framing applied to a situation that requires right-hemisphere holding: the parts get optimized, the whole is disregarded, and the costs land somewhere downstream that no one is measuring.
A right-hemisphere-respecting integration of AI looks different.
First, there are roles whose job is to hold what AI cannot hold. These differ from AI oversight roles, which are quality control at the end. The roles I mean have attention broader than what AI can articulate: people who hold the client relationship, who hold the institutional memory that isn’t documented, who hold the felt sense of where the firm is heading. These roles need to be staffed by people with stake and protected from being recategorized as bottlenecks when the productivity numbers get tight.
Second, work has to be structured so the right-hemisphere faculties get exercised rather than eroded. The faculties were developed, in previous generations of practitioners, through apprenticeship. The junior person did the work themselves, made mistakes, got corrected by someone more senior, and built up a felt sense of the work over years. AI threatens that apprenticeship by giving juniors a tool that produces work products that look polished from day one. The juniors don’t develop the muscles. The seniors lose the apprenticing relationship. Five years from now, there is no one with the right-hemisphere faculties the firm needs, because no one had to develop them.
Third, training has to be redesigned to take this into account. The relevant training cultivates the faculties AI cannot supply: structured reading of original documents (not AI summaries), client interactions held by juniors, the kind of slow attention that builds the felt sense over time. This needs to be a deliberate investment because the path of least resistance points the other way.
Organizations face a choice they are not currently treating as a choice. The choice is whether to absorb AI as pure productivity gain (and quietly let the right-hemisphere faculties atrophy across the workforce) or to absorb it as a tool whose limitations require structural compensation. The first looks cheaper in year one and more expensive in year five. Most organizations are heading for the first by default.
ACTS in practice: the advisor
This is where the argument gets sharpest, because the advisory craft is largely right-hemisphere work, and AI is more disruptive to it than to most kinds of work.
What advisors are actually paid for, when you strip the surface activity away, is context-holding (the felt sense of this client, this situation, this moment), the tacit (the craft sense that develops over years of advising), and stake (the answerability that makes advice trustworthy because it can be challenged later). These are the ACTS faculties precisely. The advisory craft is ACTS made into a job description.
AI is good at the things that look like advice from outside: producing articulate analysis, organizing material, framing options, drafting recommendations. It is bad at the things that make advice valuable from inside: knowing what this particular client needs that another client wouldn’t, knowing when a recommendation that looks correct on paper would fail in this room, knowing what you are answerable for.
The result is that AI use in advisory work is not a neutral productivity question. It is a question about what the advisor is still bringing that makes the advice theirs. As article three argued, tacit knowledge is what makes an expert an expert; careless AI adoption erodes that knowledge, and with it the substance of expertise the client is paying for. If the advisor adopts AI without changing anything else, what they end up bringing is the prompt, the editing, and the relationship management. The substantive part of the advice has been produced by something that has no context, no tacit feel, and no stake. What ends up reaching the client is branded AI output. The client may not notice immediately, but they will notice eventually, and the trust the advisor’s name has accumulated over years gets quietly burned.
The senior advisor and the junior advisor need to use AI differently, for reasons the framework makes explicit. The senior has the faculties; the AI is a tool that extends them, and the senior’s attention, context-holding, tacit sense, and stake are all in play. The junior does not yet have the faculties; the AI does for them what they need to be doing themselves in order to develop the faculties. If juniors use AI the way seniors do, the path to becoming a senior is broken. They become permanent juniors with a powerful tool, and the firm’s bench is hollowed out.
A closing thought
To close the series, I want to come back to where I started.
Years of sustainability work taught me to notice when a single mode of thinking has stopped serving the problem it is supposed to solve. In corporate sustainability, the dominant left-hemisphere thinking reduces whole-systems questions to metrics and calculations and then wonders why the calculations don’t deliver the outcomes the system actually needs. The pattern is familiar. The pattern is the same one McGilchrist describes happening across centuries of cultural history. The pattern is the one I have been arguing, across this series, is now showing up most powerfully in AI.
The framework I have offered, ACTS, is small. It won’t fix everything. What it does is give people who already sense something is wrong with the current discourse a place to stand. There’s a better way to articulate “judgment”. Human as a check was the wrong position. The work is bigger than that, and it is yours to do, before the AI gets to it.
McGilchrist’s line of caution is that the left hemisphere is a wonderful servant and a very poor master. The same is true of the most powerful thing the left hemisphere has ever built. AI is a wonderful servant. It cannot be the master. The role of the human, in the era of AI, is to be a master worth serving.