Using AI Without Losing Your Own Insight: Vision, Hypothesis Space, and the Limits of the Machine

This website, and the book it supports, wouldn't exist in its current form without AI. I've spent years capturing ideas — on my phone, in scattered notes, in the middle of teaching days — and tools like Claude and Gemini have genuinely helped me organize years of that raw material, find themes I hadn't consciously noticed in my own thinking, and bridge separate pieces of writing into something more coherent. I say this up front because it matters for what follows: every field now requires people who can work with AI well. But working with it well is not the same as handing it the wheel. This reading is about the difference — what AI is genuinely good at, what it still cannot do on its own, and why the part it can't do is exactly the part that will keep you valuable in whatever you end up doing.

What AI is actually good at

Given enough raw material and context, AI tools are remarkably good at finding patterns across a large, messy body of work — connecting an idea from one file to a related idea in another, noticing a recurring theme across years of scattered notes, or drafting a first pass at organizing something that would otherwise take a person far longer to sort through by hand. Used this way, AI functions like an extremely fast, tireless research assistant: one that can hold an enormous amount of context at once and surface connections a single human mind might miss simply from the sheer volume of material involved.

What AI still can't do

Recent work in the philosophy of mind and language has made a specific, well-argued case that today's AI language models lack genuine intentionality — the capacity to originate their own goals and direct their own attention toward a purpose they've chosen. Researchers describe current systems as fundamentally responsive rather than agentive: they answer, elaborate, and connect within whatever direction a person points them toward, but they don't independently decide that a direction is worth pursuing in the first place. In practice, this matches what I've experienced directly: AI can give me back genuinely useful context connected to my own thinking, but it needs to be pointed. It can find connections I hadn't seen. It cannot supply the bridge to what comes next — the sense of which of those connections actually matters, and why. That judgment is still coming from me.

Hypothesis space, revisited

If you've read this site's earlier piece on finding your own research question, you've already met the idea of a hypothesis space — the enormous, mostly unexplored territory of questions a researcher could choose to pursue at any given moment. Most of that space goes unexplored not because it's unreachable, but because nobody has pointed their attention there yet. AI is genuinely excellent at helping you explore a hypothesis space once you've picked a direction into it — summarizing what's already known, drafting variations, checking your logic, generating a wider set of options than you'd think of alone. What it doesn't do is tell you where, in that vast space, is actually worth walking toward for you specifically, given what you care about and why you're doing the work at all.

Pasteur's Quadrant, and where AI fits

Political scientist Donald Stokes's 1997 book Pasteur's Quadrant offers a useful case study in exactly this kind of human vision. Stokes argued that "basic" and "applied" research aren't opposite ends of one line, but two separate dimensions: how much a project pursues fundamental understanding, and how much it's driven by a real-world use. Physicist Niels Bohr's quest to understand atomic structure, with no application in mind, sits in one quadrant; Thomas Edison's relentlessly practical invention work, without much interest in underlying theory, sits in another. Louis Pasteur's own research — on fermentation, on disease, on vaccination — sits in a third quadrant entirely, because he was pursuing deep scientific understanding and a specific, felt sense of real human need at the same time, inseparably. Pasteur didn't arrive at that combination by mechanically searching a hypothesis space. He arrived there because he had a vision — a value-driven sense of which questions were worth his life's attention — and that vision is not something any tool, then or now, could have generated for him.

Being the master of your own domain

Here is the practical version of all of this, for whatever field you end up in: the people AI displaces won't primarily be people replaced directly by the tool. They'll be people who never became good enough in their own domain to direct it well — to recognize a genuinely good AI-generated idea from a mediocre one, to catch its mistakes, to know which question is actually worth asking before you ever open the tool at all. Becoming a real master of your own domain — not instead of using AI, but so that you can use it well — is what keeps a human genuinely necessary in that loop. I didn't write this website or this book by asking AI what I should think about. I brought years of my own ideas, my own values as a teacher and a scientist, and a specific direction I wanted to build toward — and the tool helped me get there faster than I could have alone. That order of operations is the whole point.

Intentionality

The capacity to originate one's own goals and direct attention toward a chosen purpose; current AI language models are widely argued to lack this on their own.

Hypothesis space

The full, mostly unexplored range of questions a researcher could choose to investigate at any given time.

Pasteur's Quadrant

Donald Stokes's framework describing research driven by both fundamental understanding and real-world use at once, as distinct from purely basic (Bohr's) or purely applied (Edison's) research.

Vision

A person's value-driven sense of which questions, among all possible ones, are actually worth pursuing — the part of research direction this reading argues AI cannot supply on its own.

Human-in-the-loop

A way of working with AI in which a person retains judgment, direction, and final responsibility rather than deferring decisions to the tool.

Think about it

Sources: Stokes, Donald E. Pasteur's Quadrant: Basic Science and Technological Innovation. Brookings Institution Press, 1997; recent philosophy of mind and language scholarship on large language models and intentionality, including work published in Synthese and Phenomenology and the Cognitive Sciences (2025–2026). Personal reflection: J. Reid Schwebach, on the process of using Claude and Gemini to organize and connect years of personal notes into this website and its companion book. See also this site's companion piece, "Finding Your Big Idea: A Real Method for Navigating the Hypothesis Space." DRAFT — verify current Virginia Science SOL alignment (if any) with the current Curriculum Framework before publishing.