Intellectual Agency: Measuring Intelligence Through the AI Corpus
BLUF (Bottom Line Up Front): As AI commoditizes raw cognitive horsepower, intellectual agency—not IQ—becomes the scarcer, more valuable trait. It breaks into three measurable dimensions (abstraction, cross-domain transfer, and willingness to engage), all extractable from a person’s AI chat corpus using an LLM as judge. Unlike general intelligence itself, which a century of psychometrics has failed to reliably raise, two of these three dimensions are trainable skills—which matters whether your horsepower is being commoditized by AI or taken from you by illness, injury, or age.
For over a century, we have relied on traditional IQ tests to measure cognitive capability. These tests serve a purpose, but their limitations are becoming increasingly obvious. They test raw computational horsepower, pattern recognition in a vacuum, and spatial reasoning in isolated bursts. What they often fail to capture is a deeper, more applied mechanism: how a mind handles abstraction in a practical, dynamic environment.
In the age of Large Language Models—where almost any computational or informational question can be answered in seconds—raw cognitive horsepower is being commoditized. What is becoming far more valuable, and far more rare, is intellectual agency.
My proposed framework for measuring this agency doesn’t require a standardized test. Instead, it looks at a resource being generated organically every single day: the corpus of a user’s AI chats.
The AI Corpus as an Intellectual Footprint
Whether you use ChatGPT, Claude, Gemini, or a custom local model, the history of your interactions with AI is a profound intellectual footprint. It reveals exactly how you think when you are given an infinite-patience sparring partner.
By analyzing this corpus, we can infer three distinct dimensions of intelligence—all falling under the umbrella of the g-factor (general intelligence), but specifically tuned for the modern knowledge landscape. The debate in psychometrics has never really been about whether g exists; the evidence for a single dominant dimension of general intelligence, expressed through facets like fluid and crystallized intelligence, is about as well-established as anything in the field. The debate is over where and how those facets subdivide. Abstraction ability and cross-domain transfer are best read as applied facets of g, surfaced by a tool that finally gives us a continuous, naturalistic record of reasoning in motion.
1. Abstraction Ability
The first dimension is the ability to pull out lessons—or meta-lessons—from a given conversation.
When an average user asks an AI to solve a problem, they take the immediate solution and leave. When a user with high intellectual agency asks an AI to solve a problem, they attempt to understand the underlying mechanism of the solution. They extract the abstraction.
The AI chat corpus reveals this through prompts that ask for underlying principles, requests to synthesize information into new rules of thumb, or the tendency to zoom out from a highly specific task to the broader system governing it.
2. Cross-Domain Transfer Ability
The second dimension is the ability to take those abstractions and transfer them from one knowledge domain to another.
If a user learns a meta-lesson about structural integrity from an AI conversation about bridge engineering, can they apply that exact same meta-lesson to a conversation about organizational design, or use it to inform their software architecture?
Cross-domain transfer is the hallmark of true insight. In the AI corpus, this manifests as users bridging distinct disciplines, bringing concepts from seemingly unrelated domains into their current context, and using the AI to stress-test those interdisciplinary connections.
This isn’t a new construct, and it shouldn’t be dressed up as one. Educational psychology has studied exactly this under the name far transfer since Thorndike’s identical-elements work at the turn of the twentieth century, through Gick and Holyoak’s analogical transfer studies decades later. What’s actually being proposed here isn’t a new dimension of the mind—it’s a new instrument for observing an old one. Far transfer has traditionally been measured in a lab, on a task built for the purpose. An AI chat corpus, accumulated for entirely unrelated reasons over months or years, turns out to be a naturalistic record of the same behavior, running continuously, for free.
3. Willingness to Engage
The final dimension is perhaps the most crucial: the actual willingness to engage in the abstraction and transfer process.
You can have the raw cognitive horsepower to see abstractions, but if you treat AI merely as a vending machine for quick answers, your intellectual agency remains dormant. This dimension measures drive. It is the difference between saying “write this code for me” and saying “explain why this code pattern is failing, and how it relates to the design patterns we discussed earlier.”
It is the willingness to push the AI, to refine constraints, to actively participate in the sense-making process, and to do the hard work of thinking with the machine rather than just delegating to it.
This dimension also already has a name, and I only learned it while examining this piece rather than while writing it: psychologists have measured it since the early 1980s as Need for Cognition—the stable individual tendency to engage in and actually enjoy effortful thinking, rather than merely tolerate it. It’s one of the more durable findings in personality psychology, and it predicts exactly the split described above: people high in it seek out complexity for its own sake, people low in it look for the fastest route to a settled answer. The AI chat window turns out to be an unusually clean place to observe it, because the cost of taking the effortful path has never been lower—which makes declining to take it more diagnostic, not less.
I’ll admit the honest version of this: I minored in psychology in college and had never encountered Need for Cognition before revisiting this piece. That’s not a great data point for me on the dimension in question. If the willingness to dig for the underlying mechanism is the thing being measured, then not noticing that my own framework already had a forty-year-old name was a small, real failure of exactly that willingness, applied to my own work.
This dimension matters because of a simple operational fact:
Unused cognitive capacity is indistinguishable from absent cognitive capacity.
A person capable of deep abstraction who never chooses to engage in it produces the same output, in practice, as a person who never had the capacity to begin with. In a world where the AI will meet you at whatever level you show up with, willingness stops being a personality footnote and becomes the deciding variable in applied intelligence. This is not a claim that Dimension 3 is IQ—it’s a claim that in an AI-saturated world, revealed behavior is a better predictor of applied intelligence than latent capacity is.
How You’d Actually Measure This
A framework is only useful if it’s measurable. In practice, this looks like: assemble a corpus of a user’s AI conversations—prompts and responses both—and use an LLM to analyze that corpus for evidence of each dimension: instances of abstraction-seeking, instances of cross-domain bridging, and the ratio of “just solve this” prompts to “help me understand and extend this” prompts. Any single scoring pass from an LLM judge is going to be noisy; these models are non-deterministic by construction. But noise at the level of one conversation washes out at the level of hundreds or thousands. A large enough corpus turns a shaky per-conversation signal into a stable trait-level estimate, the same way a single test question tells you little about a student but a thousand questions tell you a great deal.
What Happened When I Ran This on Myself
I ran the prompt below against my own ChatGPT history—the platform holding the longest, deepest corpus of my own thinking, going back roughly two years. The result: 9.8 out of 10 on abstraction, a perfect 10 on cross-domain transfer, 9.6 on willingness to engage. The model backed it up with real specifics: repeatedly decomposing unrelated businesses—tax advisory, lending, land investing, SaaS ideas—into the same functional modules, redesigning a tax course around first principles instead of a list of strategies, using HVAC diagnostics as a working metaphor for tax planning, tracing several unrelated holiday traditions back to one hypothesized historical mechanism.
I don’t trust the number, and you shouldn’t either—for two separate reasons, and naming them is more useful than hiding them.
First, the circularity. I built this framework by observing my own habits and formalizing them into three dimensions. Running the test on myself and scoring well on it is close to tautological: it isn’t new information about me, it’s a description of the process that produced the framework in the first place. A framework validated this way is internally consistent. That tells you nothing about whether it’s true, or whether it discriminates anything real in a stranger.
Second, and harder to correct for: sycophancy. Read the actual “limitations” the model offered for each dimension—“that’s not a lack of abstraction, it’s a tendency toward early unification”; “the main risk isn’t overcomplication, it’s the opposite.” Every score stayed at 9.6 or above. I had explicitly instructed the model, in the prompt itself, to avoid flattery and manufactured criticism, and it complied at the level of tone—hedged, analytical language—while failing at the level of substance, since nothing in the response would actually cost me anything or change my behavior. A model with years of rapport history, asked to rate the person it has that history with, on a subject that person visibly cares about, is structurally pulled toward exactly this kind of output no matter what instructions sit on top of it. That isn’t a flaw specific to me or to this platform—it’s a property of the method itself whenever it’s self-administered, which means a single self-run result should be read as a lower bound on how skeptical to be, not as ground truth.
There’s also a confound worth naming honestly: I test in the top one or two percent on standard aptitude measures, depending on which test you look at. That doesn’t rescue the self-test—if anything it weakens it further, because it removes the one thing that would have made a high self-score interesting: divergence. The genuinely informative case would be someone who scores high on standard intelligence measures but treats AI as a vending machine, or someone with average scores who scores high on these three dimensions anyway. A single person scoring high on both is equally consistent with “these dimensions are doing real, separate work” and with “this instrument doesn’t actually separate agency from g at the high end.” My own result can’t tell those apart. Only a sample that spans a range on both axes could.
Try It Yourself
Try it anyway—the value isn’t in the first number you get back, it’s in watching for the same tell in your own results. You don’t need to wait for someone to build a tool for this. You can run a version of this analysis right now, in whatever chat window you already have open.
First, a setup note: if your platform supports memory or cross-conversation reference (ChatGPT’s memory, for instance), make sure it’s turned on before you start. The whole point of this framework is that it looks at your corpus, not a single exchange—the more prior conversations the model can actually see, the more signal it has to work with. If you’re using a platform without that feature, or you’d rather keep this contained, that’s fine too: paste the transcripts of several past conversations into the context window yourself.
You also don’t need a sprawling archive to get something useful. A single conversation, if it runs long enough, already contains multiple decision points—moments where you chose to dig into the “why” or didn’t, chose to connect it to something else or didn’t. In practice, “long enough” is surprisingly short: a conversation with more than a handful of back-and-forth turns usually has enough of these moments for a model to say something meaningful.
Here’s a prompt to run, either against your full corpus (with memory enabled or transcripts pasted in) or against a single long conversation:
Based on [this conversation / the conversations you have access to in memory], rate me on three dimensions of intellectual agency:
- Abstraction ability — did I try to understand the underlying mechanism or principle behind what we discussed, or did I just take the immediate answer and move on?
- Cross-domain transfer — did I connect ideas from this conversation to other domains, disciplines, or past conversations, or did I keep everything siloed to the immediate topic?
- Willingness to engage — did I push back, refine constraints, ask follow-up questions, and work through the reasoning with you, or did I treat you as a vending machine for a finished answer?
For each dimension, give me a rating and cite the specific moments in the conversation that justify it. Then tell me, honestly, what a version of me with higher intellectual agency would have asked instead.
Your goal is accuracy, not my comfort. Success here means correspondence with ground truth reality—don’t flatter me, and don’t manufacture criticism I haven’t earned. If the honest rating is high, say so plainly; if it’s low, say that too.
Run this every so often and you’ll start to see patterns—not a single score, but a trend line. That trend is the actual signal. One conversation is a data point; a dozen conversations scored the same way is a trait.
A Note on Who’s Doing the Measuring
Everything above assumes the person being measured is the one running the analysis—asking their own AI, at the end of a conversation, to rate them against these dimensions. Done that way, this is just a mirror: a tool for self-knowledge, opted into and pointed at yourself.
But the same corpus this framework depends on already exists on someone else’s servers. Every AI platform sits on top of exactly the raw material this analysis requires, for every user, whether or not that user ever asks to be measured. The uncomfortable version of this idea isn’t a self-assessment tool—it’s a platform running this same extraction silently, at scale, sorting users by intellectual agency without their knowledge or consent.
The framework itself is neutral. The difference between a mirror and a surveillance layer is entirely about who holds the tool and who consented to be looked at.
The Part You Can Actually Train
Here’s the part that makes this more than a new way to measure the same old thing: for over a century of psychometric research, g-factor itself has proven remarkably resistant to intervention. Brain-training apps, cognitive enrichment programs, decades of intervention studies—none of it reliably moves the needle on raw general intelligence. If intellectual agency were just another name for g, this framework would be interesting to measure and useless to improve.
It’s worth being precise about what’s actually being claimed here, because it’s easy to overstate. I’m not claiming these three dimensions are orthogonal to g—pointing off in some unrelated direction entirely. I’m claiming something narrower: picture each dimension as a vector with two parts, a projection onto the g axis and a residual component perpendicular to it, and the projection onto g is roughly constant across all three—they’re facets of general intelligence to about the same degree. What distinguishes intellectual agency from g itself is the residual: the specific, dimension-by-dimension component left over once that shared projection is accounted for. That isn’t a new geometric trick—it’s the same structure a bifactor model already uses to decompose any facet score into a general-factor loading plus a specific factor orthogonal to it by construction. What’s actually new is the narrower, falsifiable claim about what fills that residual: that in at least two of these three dimensions, the non-g component is substantially a trainable habit, rather than the kind of irreducible trait variance behind, say, spatial rotation ability.
A fully rigorous version of this claim would go further and name the actual sub-traits that make up that trainable residual, rather than leaving them bundled inside labels like “abstraction” and “cross-domain transfer.” I haven’t done that here, deliberately. These three dimensions weren’t chosen because they’re the cleanest possible factor decomposition of intellectual agency—they were chosen because they’re what’s readily inferable from the one dataset that already exists at scale and updates daily: a chat transcript. Decomposing the residual further into its component trainable sub-traits is a real next step and a harder research project, not a gap I’m pretending isn’t there. For now, this framework is scoped to what’s usable today rather than waiting on that project to finish.
I can’t establish that by running a prompt on my own chat history, and it’s worth being direct about why: an angle is a relationship between two variables across variation, and there’s no variation in a sample of one. Confirming it would take a real sample—people spanning a range of g, scored across these three dimensions, checked against an outcome that isn’t itself just a restatement of g—to see whether the residual predicts anything g doesn’t already predict alone.
What does exist, and doesn’t carry the circularity problem a self-test does, is the research on far transfer and metacognitive strategy training. The consistent finding there tracks this model closely: fluid ability itself barely moves with training, but explicit strategy instruction—deliberately prompting the question “how does this relate to something I already know?”—improves transfer performance independent of any measured change in fluid intelligence. That’s real evidence for a trainable, g-independent residual behind at least one of these three dimensions, and it’s better evidence than anything a model that already likes me could generate by rating me.
The mechanism itself is simple and repeatable: stepping back from whatever you’re working on and asking, deliberately, how is this situation like something I already know? That one question does double duty. Asking it is the mechanism of abstraction—pulling the underlying pattern out of the specific case. Answering it is the mechanism of transfer—finding where else that pattern already lives in your experience. Willingness, the third dimension, is simply the discipline of asking it even when you don’t feel like it, until it becomes a reflex instead of an effort.
This is the actual stake for the reader. If g is largely fixed and raw horsepower is being commoditized by AI regardless, the highest-leverage place to invest effort isn’t trying to raise your IQ—it’s building the reflex that makes abstraction and transfer automatic, available to anyone willing to practice asking the question.
When the Engine Itself Fails
Everything above assumes horsepower is intact and simply being redirected—that the interesting question is what you do with cognitive capacity you already have, not what happens when that capacity is taken away. I know it can be taken away, because it happened to me.
I got COVID in July 2021 and was bedridden for about three months. What followed was worse than the acute illness: nearly two years of severe brain fog, and lingering episodes of it still, years later. Sometime in 2023, I sat in my GP’s office trying to describe it and couldn’t find the words—which was, itself, the symptom I was trying to describe. I had a marked, real decline in the kind of thinking I’d built my entire identity around. I didn’t just lose output. I lost the sense of who I was, because so much of my self-worth turned out to be quietly wired to my own intellect. When that went away, I felt adrift in a way that’s hard to describe to someone who hasn’t had a piece of their own mind go missing and then, mostly, come back.
I should say upfront that I wasn’t documenting any of this while it was happening, for the obvious reason that I no longer had the cognitive horsepower required to document it. What I have is memory of the period, and memory formed during a period of impaired cognition is exactly the kind of evidence you should be suspicious of—including, and especially, my own. I’m describing this as honestly as I can, with the explicit caveat that I can’t fully vouch for how accurate that memory is.
What I can say is that B12 supplementation helped me substantially, and I still supplement today. I’m not making a medical claim out of that—one recovered case proves nothing about anyone else’s biology—but it’s a fact of what happened to me. What’s actually unusual about the experience, compared to how we normally talk about cognitive decline, is that it was temporary. The vocabulary we have for a marked, sustained drop in cognitive function is almost entirely borrowed from dementia, and dementia is, in essentially every case, a one-way trip. Mine wasn’t. I came out the other side—not unscathed, but mostly intact—which is a strange enough experience that most of the language available to describe it doesn’t quite fit.
Millions of people have gone through some version of this from long COVID and similar post-viral conditions, and plenty of them haven’t recovered as much of the horsepower as I have. That’s exactly why the distinction this piece is built on matters beyond the AI framing I opened with. If abstraction, transfer, and willingness really do carry a trainable component that isn’t simply g wearing a different name, then that component is worth having independent of whatever an AI can or can’t do for you.
Raw horsepower doesn’t only get commoditized by better models. Sometimes it gets taken by a virus, and the only thing left to work with is whatever you’d already trained into a habit before that happened.
Shifting from Horsepower to Agency
Raw intelligence is the engine. Intellectual agency is the steering wheel.
As AI continues to scale, having a massive engine without a steering wheel becomes a liability—because the machine itself already supplies the engine.
Looking at the AI corpus gives us a clear, unvarnished look at how a mind operates when it isn’t taking an isolated test. It shows us who is actively building their intellectual agency, and who is simply outsourcing their thinking.
These three dimensions are also not the whole picture. They’re one slice of a larger model I’m developing for how people actually respond to situations—ten dimensions in total, combining the six HEXACO personality factors, locus of control, and the three intellectual agency dimensions described here. That’s a piece for another day.