AI Intuition: The Discernment of Where AI Belongs

The question that matters about AI is where it belongs.

I have come to call the skill that answers that question AI Intuition. It is the calibrated capacity to know, in a particular situation, whether AI should be absent, advisory, collaborative, or in command. It is trainable. It is also, at the moment, almost no one's training priority. The default move is to reach for the tool, then reason about it. The skill is to reason about it, then choose whether to reach.

That default is starting to leave a measurable trace.

A team at MIT Media Lab ran an EEG study across three groups writing essays. One used ChatGPT, one used a search engine, one had no aid at all. The ChatGPT group showed the lowest brain engagement of the three and consistently underperformed at neural, linguistic, and behavioral levels. A separate randomized trial put a group through a 45-day delay and then tested them on what they had supposedly learned. The two groups had performed about the same in the moment. The traditional learners retained eleven points more afterward.

A Harvard Business School field study of 244 consultants at Boston Consulting Group put numbers on the human side of the same pattern. Even while being formally evaluated, twenty-seven percent of highly motivated, highly capable professionals defaulted into what the researchers called Self-Automator mode: handing the whole task to the model and lightly editing the result. They knew they were being watched and slid into it anyway. The pull toward over-delegation is strong enough to overcome incentive.

My claim is about positioning. The cost of frictionless delegation is real, and it compounds quickly. The judgment that should be deciding when to delegate is exactly the judgment that gets quiet first.

The cognitive scientists have a sturdy frame for when human intuition can be trusted. Daniel Kahneman and Gary Klein, in a famous adversarial collaboration in 2009, concluded that expert intuition becomes reliable in two conditions: the environment provides valid cues, and the practitioner has had repeated practice with prompt, unambiguous feedback. Fireground commanders have it. Stock pickers do not. Your gut about where AI belongs is reliable only in the domains where you have done the reps and honestly looked at the outcomes.

Most of us have done many uses without doing the reps. The reps include the moments where you notice afterward that you would have learned more by doing it yourself, or that the draft you accepted made the meeting weaker, or that your own thinking became sloppier in proportion to how much you offloaded. Without that review loop, you are doing the same thing every day and slowly believing you are getting better at it.

The contemplative traditions had a word for this skill long before the cognitive scientists needed one.

In the Yoga Sutras, Patanjali named it viveka: discrimination. Sutra 2.26 calls uninterrupted discriminative discernment the means of liberation. In the Pali Canon, the Buddha named it yoniso manasikara, attention to the origin, wise reflection. Aristotle, in Book VI of the Nicomachean Ethics, called it phronesis, practical wisdom, the capacity to see the right thing to do in the particular case. The Desert Fathers called it diakrisis. Ignatius of Loyola devoted the Spiritual Exercises to it under the name discretio spirituum, with two sets of rules for distinguishing the inner movements that lead toward life from the ones that lead away from it. The Quakers built a group protocol around it: the clearness committee, four to six people whose only job is to ask the focus person open, honest questions for two or three hours, with no advice and no leading.

These are detailed protocols for the cognitive skill of knowing what fits. The skill is old. The stakes are new.

The most stable contemplative practice for building this skill is the Examen. Ignatius treated it as the one prayer Jesuits were instructed never to skip. The shape is simple. At day's end, you become quiet. You review the day with gratitude. You notice your emotions, where there was consolation (clarity, energy, a deepening) and where there was desolation (numbness, restlessness, a sense of being more shallow than you started). You take one feature and reflect from it. You set a concrete intention for tomorrow.

I run this on my AI use. Where today did AI help me think better? Where did it help me think less? Which uses produced consolation, which produced desolation? When the same tool keeps producing desolation, I stop using it. When a use keeps producing consolation, I keep doing it. The Examen is, structurally, the contemplative ancestor of the modern decision journal. It works.

The other half of the skill is design. How you position AI in a workflow is upstream of how it affects you.

Researchers at HBS, after the BCG study, formalized three stable postures. The Centaur keeps a clear division of labor: human owns strategy and judgment, AI owns calculation and execution. The Cyborg interleaves iteration deeply within each step. The Self-Automator delegates wholesale. The right posture depends on the work. Centaur fits high-stakes, judgment-laden decisions. Cyborg fits exploratory, iterative work where you want the friction of comparison. Self-Automator fits routine work where skill development is not the point. The trap is using Self-Automator mode on work where the effort itself is the mechanism: learning, formation of judgment, writing that is supposed to clarify your own thinking. The model finishes faster. You learn less.

Cal Newport's three-screen filter from Digital Minimalism is the cleanest tool I know for admitting any new technology. Does this AI use serve something I deeply value? Is it the best way to serve that value? Can I bound it with clear operating constraints? Three yeses, the tool earns its place. Anything less, it is optional, which means it is clutter.

The most actionable habit I have built around all of this is the pre-prompt pause. Before I reach for a model, I name the cognitive task in one sentence. Am I generating, evaluating, deciding, or learning? Generation and evaluation tolerate AI well. Learning tolerates it poorly, because effort is the mechanism. Decision is mixed and depends on the stakes. Naming the task takes five seconds. It changes the next twenty minutes.

The deeper claim under all of this is that production is now cheap and taste is the bottleneck. When anything can be made, the scarce question is what should exist. Taste is cultivated discernment under a contemporary name. You build it the same way Aristotle and the Buddha and Kahneman agree you build any expert intuition: deliberate practice in a high-validity environment with prompt feedback. Read excellent things. Refuse slop. Articulate why an option is wrong before generating alternatives. Every day. The reps build the filter.

I think of AI Intuition as a formation practice. The Examen, the pre-prompt pause, the monthly review of where you delegated and what you got back, these are the practices by which a person becomes the kind of person whose intuition about AI can be trusted. The output is a more accurate sense of where your own attention belongs.

How we use our tools shapes who we become. The Ignatian tradition argues this at the soul scale. Shoshana Zuboff argues it at the civilizational scale. The MIT study argues it at the EEG scale. They agree.

The question is whether you are training the skill that decides, or only the skill that delegates. They are different skills. They live in different parts of you. And only one of them gets to choose what kind of mind you have on the other end of this.