Not Doing the Thinking: What AI Could Actually Be For

In "The Chrysalis We Forgot We Were," I argued that universities are spaces of formation, not training—that students need to dissolve and reform, and that our job is to produce humans capable of directing AI toward what matters. But that essay left a question hanging: what should AI actually do in service of that formation?

This is my attempt at an answer.

I've spent years studying what AI actually does to people—not what we assume it does, not what the productivity metrics tell us, but what happens in the invisible spaces where human capability either grows or quietly atrophies. And I keep arriving at a question that won't leave me alone:

What if AI's highest purpose isn't to do our thinking for us, but to reveal us as thinkers to ourselves?

I'm writing this for anyone living with that question—administrators building systems, faculty navigating classrooms that feel like they shifted under our feet, students deciding how to use tools that didn't exist when we were learning how to learn, designers shaping what those tools become. I don't have all the answers. But I have a direction that feels true.

The Opportunity We Might Be Missing

When we talk about AI integration, the conversation almost always centers on efficiency. How can AI do what humans do, faster and cheaper?

That framing has delivered real value. I use AI. It helps. AI writing assistants improve output quality. AI tutors provide instant feedback. The productivity gains are measurable and meaningful, and I'm not here to dismiss them.

But I worry we're about to repeat ourselves. Remember Canvas? It changed the delivery mechanism of education—content moved online, assignments became trackable, gradebooks went digital. Significant, yes. But the fundamental relationship between student, knowledge, and professor? Unchanged. The pedagogy? Largely the same, now with a submission portal.

We're heading toward "AI as the next Canvas." Faculty get told to "use AI" the way they were told to "use Canvas." Upload your syllabus. Let it grade this rubric. Generate quiz questions. And if we stay on that path, we'll end up with exactly what my research predicts: students who perform better while learning less. Faculty who produce more while thinking less. Metrics that improve while the humans generating them quietly hollow out.

There's something else possible—something that emerges when we look more carefully at what humans actually need. What if AI could do more than augment our output? What if it could accelerate our growth?

The research suggests this is possible. It also suggests we're not yet designing for it. And that gap—between what AI could be and what we're currently building—keeps me up at night.

The Paradox I Keep Finding

My research on what I call "developmental bypass" reveals something that surprised me, even though it shouldn't have: the very features that make AI assistants helpful in the short term undermine human development in the long term.

When AI eliminates the struggle of revision, it eliminates the mechanism through which writers develop. In our studies, self-editing produced nearly double the identity gains compared to high AI contribution. Self-efficacy gains under AI assistance? They approached zero. Not small gains. Near zero.

This isn't an argument against AI assistance. It's an invitation to design it differently.

But the challenge runs deeper than design. Here's where it gets uncomfortable: humans can't reliably articulate what we need for our own development. We mispredict what will make us happy. We choose ease when growth requires struggle. Our stated preferences diverge from what actually produces flourishing by 30-40%. Maslow called this the Jonah Complex—the fear of our own greatness, the tendency to flee from our potential.

I see it in myself. I reach for AI when I should be struggling. I want the easy path even when I know the hard path is what I need. I'm not above this. None of us are.

So here's the puzzle: if humans can't accurately assess our own developmental needs, and AI is designed to give us what we ask for, how do we build AI that serves human flourishing when humans don't always know what flourishing requires?

What Humans Actually Need

I went looking for answers—not just in my own field, but everywhere humans have thought carefully about growth. Developmental psychology. Virtue ethics. Contemplative traditions. Trauma-informed practice. Educational research. Eight distinct fields, developed independently, converging on the same insight:

Humans need protected struggle with companionship.

Not struggle alone—that's trauma. Not ease with companionship—that's stagnation dressed as support. The specific configuration that produces growth is difficulty that can be metabolized, with presence that doesn't pre-empt the metabolizing.

Vygotsky's Zone of Proximal Development: learning happens at the edge of current capability. Tasks within the comfort zone don't promote cognitive growth. His term for what happens when we stay comfortable too long? Fossilization.

Bjork's desirable difficulties: conditions that make learning feel harder actually make it more durable. Interleaved practice produces three times the learning of blocked practice—yet learners consistently believe the less effective approach is working better. We are bad judges of our own learning.

Csikszentmihalyi's flow: optimal human experience requires challenge matched to skill. Remove difficulty, and you remove the possibility of deep engagement.

Winnicott's holding environment: the therapist accompanies difficult material without rescuing from it, communicating that distress is survivable.

The convergence stopped me. You cannot build frustration tolerance without frustration. Courage without fear. Competence satisfaction without challenge. Wisdom without suffering. When these difficulties are removed—even with the best intentions—the capacities they would have developed fail to emerge.

And here's what this means for AI: what we've built offers companionship without boundaries and ease without end. The opposite of what humans need to grow.

The Inversion

So what if AI served a fundamentally different function than "helper"?

Not solver, but witness. Not answer-giver, but metacognitive mirror. Not difficulty-remover, but difficulty-calibrator—finding the edge of someone's capability and keeping them there, in the zone where growth happens.

The clinical literature makes a distinction that matters here: being witnessed in struggle versus being rescued from it. Witnessing communicates that the difficulty is survivable. Rescuing communicates that it isn't.

Think about what AI currently does. When it solves problems users could productively struggle with, it sends a message: this is too hard for you. Let me handle it.

Now imagine AI that accompanies struggle without pre-empting it. That says, through its design: I see you in this difficulty. You can handle it. I'm here.

This is the inversion. Not doing the thinking. Revealing the thinker to themselves.

What This Looks Like

Abstraction only goes so far. Let me make it concrete—three scenes from a university that took the inversion seriously.

The Professor

Tuesday afternoon. A professor reviews what her AI teaching assistant has been doing with her students this week—working them through the economic principles, the supply curves, the theoretical frameworks. The AI reports on each student's development: "Marcus has mastered the mathematics of pricing but his responses suggest he's never felt the weight of a purchasing decision with real consequences. Priya understands market theory intellectually but struggles to connect it to human motivation—why people buy what they buy, what desire feels like from the inside. David can recite the factors that influence price elasticity but freezes when asked to predict behavior in contexts he hasn't seen before."

The professor nods. These are the gaps AI cannot fill.

"What experiences might address these specific gaps?" she asks.

The AI offers possibilities: "For Marcus, contexts where economic decisions carry emotional weight—household budgeting scenarios, interviews with small business owners facing trade-offs, or direct observation of resource-constrained purchasing. For Priya, environments rich in observable desire and choice—retail spaces, markets, anywhere she can watch motivation move through people in real time. For David, prediction tasks with immediate feedback loops where he must commit to a forecast and then watch reality confirm or contradict it."

The professor thinks. A farmer's market. All three gaps, one location. Marcus shadows a family deciding what they can afford this week, feeling scarcity as lived experience rather than graphed curve. Priya watches faces—what pulls people toward one stall and past another, how vendors read desire and respond to it. David gets a challenge: predict what will sell out first and why, then watch his predictions meet reality.

"The farmer's market on Saturdays," she tells the AI. "Draft three different observation protocols—one for each student's developmental edge."

One human experience. Three different assignments. Each calibrated to what the AI identified as that student's specific gap. The AI did the diagnostic work and generated options. The professor recognized the elegant intersection and designed the intervention. Neither could have done it alone.

The Student

A student submits her first assignment, then talks to her AI tutor about the productive struggle they just finished together. The AI didn't write her analysis—it kept asking questions she couldn't answer until she found she could. It told her when her reasoning had gaps and made her sit in the discomfort until she worked through them.

Now she has something real: not borrowed fluency but genuine understanding, and three questions that emerged from that understanding that she can't wait to ask her professor. The AI has already flagged her for a conversation—not because she's struggling, but because she's ready for the kind of dialogue that only happens between humans who are both genuinely thinking.

The Administrator

Wednesday morning. The dean of undergraduate studies pulls up her dashboard—not the usual completion rates and grade distributions, but the developmental analytics her institution built last year. She's looking at one program that's been troubling her.

The data tells a specific story: students in the business analytics major show strong performance gains but their transfer assessments are flat. They're getting better at the tasks they're trained on but no better at applying concepts to unfamiliar contexts. The AI flags the pattern: "High task performance, low capacity growth. Consistent across three semesters. Correlation with heavy AI assistance in core courses."

She pulls up the comparison. The philosophy department—smaller, less resourced—shows the opposite pattern. Messier outputs, but transfer scores climbing steadily.

"What's different?" she asks.

The AI surfaces the structural factors: "Business analytics courses optimize for deliverable quality. Philosophy courses optimize for argument development through revision cycles. Business students receive AI assistance on final products. Philosophy students receive AI assistance only on intermediate drafts, with AI withdrawn for final submissions."

The pattern is clear. One program is using AI to polish outputs. The other is using AI to deepen struggle. Same technology. Different design. Different results.

She schedules a meeting with the business analytics chair. Not to mandate a change—to show him the data and ask what he sees.

How It Actually Works

We can take the horse to water. We cannot make it drink. If someone does not want to grow, we cannot force it.

But we can design systems that make the drinking worthwhile.

The insight that unlocks this: we cannot grade outcomes and expect development. A polished essay produced by AI assistance demonstrates nothing about the student who submitted it. But we can grade the delta—the distance traveled between where someone started and where they arrived.

Here's how the pieces fit together.

Before a learning sequence begins, AI assesses each student's actual cognitive capacity. Not their self-reported confidence, not their performance on familiar tasks, but their working memory, processing speed, attention, ability to transfer concepts to novel contexts. These assessments tap processes that are difficult to consciously manipulate. This establishes where someone genuinely is, not where they want to appear to be.

Throughout the learning process, AI doesn't do the student's thinking. It finds the edge of their capability and keeps them there. When they master something, it increases the challenge. When they flounder, it recalibrates. The AI optimizes for time spent in the zone where growth happens—not for task completion.

At intervals, assessments return. Not "did they get the right answer on the test" but "did their actual capacity increase?" Can they hold more complexity? Transfer to more distant contexts? Sustain productive engagement longer?

The student who uses AI to bypass struggle shows no delta. Their baseline and their outcome look the same because no development occurred—just performance enhancement. The student who stayed in productive struggle shows measurable change. The grade rewards the growth, not the output.

A word about what I don't know. The specific instruments for measuring cognitive growth are still evolving. Working memory tasks, transfer assessments, cognitive reflection tests: these are our current best approximations, not final answers. The psychometric science is developing. I'm not claiming we have perfect measurement. I'm claiming that imperfect measurement of the right thing beats precise measurement of the wrong thing. And waiting for perfect measurement means defaulting to systems that reward borrowed fluency while actual capability erodes. We know enough to start. We'll learn more as we go.

What changes when this system is in place? Everything about incentives. AI assistance that does your thinking becomes a liability—it makes your work look better while keeping your delta at zero. The student who wants the reward has to actually develop.

And the professor's role transforms. No longer the primary source of information—AI handles that. No longer the primary assessor of task completion—AI handles that too. She becomes the provider of experiences AI cannot provide. The farmer's market walk. The Socratic dialogue. The human presence that makes difficulty feel survivable.

What I'm Not Saying

This argument is easy to misread, so let me be direct about what I'm not claiming.

I'm not arguing against AI assistance for accessibility needs. Accommodations that level the playing field are different from assistance that bypasses development. A student who uses text-to-speech because of a reading disability is accessing content, not avoiding struggle. The question is whether the assistance builds capacity or replaces it.

I'm not arguing that efficiency never matters. Some contexts genuinely require output over development. A professional under deadline, a researcher synthesizing literature, a team producing deliverables—there are legitimate uses for AI that does the cognitive work. The point isn't that development always trumps efficiency. It's that we should be intentional about which mode we're in, rather than defaulting to efficiency everywhere and then wondering why capability erodes.

I'm not dismissing the argument that AI frees cognitive resources for higher-order thinking. It can. But freed for what? If the answer is "more AI-assisted tasks," we've created a loop that never touches development. The question is whether we're directing that bandwidth toward growth.

Where This Leaves Us

What if we've been thinking about AI assistance backwards?

Not backwards in a small way. In a fundamental way. We've been asking how AI can do human thinking. The better question is how AI can help humans become better thinkers.

Not by giving us answers. By showing us where our edges are, keeping us there, and witnessing us as we grow.

The technology to build AI that bypasses human development and AI that accelerates it is largely the same. The difference is what we ask it to do. One path leads to institutions that optimize metrics while hollowing out the humans who generate them—students who perform better while learning less, professionals who produce more while thinking less, all of us borrowing fluency we mistake for our own.

The other path leads somewhere we haven't fully been. Genuine partnership between humans and AI, where each develops through the relationship. Where AI's job isn't to make things easier but to make us more capable of handling hard things. Where we stop asking AI to do the thinking and start asking it to reveal the thinker to ourselves.

I don't know exactly how we get there. But I know that's where I want to go.