Pace Makes Fit

Case of the Week, Min Wu, PhD · ai-public-health.com

When I work with full focus — especially alongside AI — I stop noticing the clock. An hour goes by in what feels like twenty minutes. For a long time I read that as the good kind of tired: deep work, flow, the sign a session had gone well. It took me longer than I'd like to admit to notice the exhaustion afterward wasn't the flow kind. It was the kind that comes from being handed more than I'd finished thinking about, again and again, faster than I could catch up. Eventually I started setting a timer. Every hour, when I'm working with AI on a project, it goes off, and I stop — not because the work is done, but because I am.

Key Takeaways

• AI exhaustion isn't always flow — often it's a pace mismatch between how fast AI answers and how fast a person can absorb them.
• Min Wu's ATPM framework names this "cognitive load fit": does AI's pace match what you can actually take in — not whether it's right.
• Cognitive load fit scores pace, not accuracy — a confidently wrong answer can score as well as a correct one.

What I Almost Called Productivity

For months I filed the exhaustion under normal. AI tools are fast — a question goes in, a fully-formed answer comes back, sometimes three or four directions worth pursuing before I've finished evaluating the first one. Every individual reply was clear. Nothing about any single exchange overwhelmed me. What accumulated was the rate: the AI's turnaround time and my own thinking time were never running at the same speed, and across an hour that gap compounds. I wasn't confused by any one answer. I was behind on all of them at once.

Setting an hourly stop wasn't a productivity hack. It was closer to a repair — a fixed point where I have to catch up with what's already in front of me before I'm allowed to ask for more.

A Structural Parallel

I want to be precise about what this isn't, because "set a timer" sounds like advice, and that's not quite the claim. Two systems working together — a person and an AI — only produce something usable if the pace one is running at is a pace the other can actually absorb. Speed of response and quality of response are separate questions. An AI can be accurate, well-reasoned, and clearly written, and still be the wrong thing to receive at that moment, if the person on the other end hasn't finished processing the last three things it sent.

Where ATPM Comes In

In Artificial Intelligence in Public Health, I proposed a framework for exactly this kind of relationship: the AI thinking partnership model (ATPM), built for situations where AI isn't a tool you use once and close, but a partner you return to across a working relationship that evolves over time. ATPM scores that partnership across six dimensions — trust, value alignment, cognitive load fit, control and autonomy, adaptability over time, and emotional resonance — each anchored to a plain question a user could ask mid-session.

Cognitive load fit's question is the plainest one: is this AI helpful without overwhelming me? It isn't asking whether the AI is accurate, or aligned with what I actually want, or emotionally attuned — ATPM has other dimensions for that. It's asking something narrower, and easy to overlook: whether the help arriving matches what I have the bandwidth to receive right now.

It's worth being precise about what these six dimensions actually score: the user's side of the relationship, not the AI's output directly. ATPM is a user acceptance model — it captures what a product team or a health system has a direct business interest in, because acceptance predicts continued use. It doesn't measure whether that use is warranted, and it isn't built to.

Where the Score Actually Drops

I notice this most clearly away from AI entirely, in my own reading. When a book or long essay hits a term I don't know — real domain vocabulary, not jargon for its own sake — I stop on that page, and more often than I'd like to admit, I don't go back. The right vocabulary is what lets a reader carry themselves the rest of the way through a book unassisted. I don't have data for this, only my own habit and years of watching it in students: give someone the concepts, and they can largely teach themselves. Withhold them, and fluent, well-organized writing doesn't help — it just fails faster, because there's nothing to attach it to.

I think the same gate operates with AI. Two people ask an assistant the identical, well-formed question — one in a domain they've spent years in, one in a domain they haven't. Same AI, same pace, same clarity. The expert absorbs the answer the way I absorb a paragraph in my own field: mostly recognition, cheap and fast. The novice is doing what I do when I hit an unfamiliar term — constructing meaning from scratch, at a cost the fluent delivery doesn't reveal. Cognitive load fit, scored identically if both report the pace felt fine, is measuring something real but incomplete. It can't see that one of them was coasting and the other was translating — and that gap isn't AI literacy, or general literacy. It's standing in the specific domain the answer is about.

What This Doesn't Prove

I want to be honest about what an hourly timer actually demonstrates, which is less than it might sound like. It's a single self-report data point — uncontrolled, entirely subjective — for exactly the kind of measurement ATPM currently asks a user to supply themselves, on a 1-to-5 scale, because the field doesn't yet have a better instrument. Diagnostic tools that could score cognitive load fit in something closer to real time, rather than a person guessing at exhaustion after the fact and reverse-engineering a cause, are future work, not built yet. My timer is that guess, formalized into a habit — a fix I built on my side because the general-purpose model I was using had no way to build it on its own. That's true of a commercial assistant tuned once for everyone alike. It isn't evidence about where the threshold sits for anyone other than me, or even for me on a different kind of task.

There's a sharper limit worth naming too. My timer tells me when a session started to feel like too much — that's a read on fluency, on whether the pace matched what I could absorb. It says nothing about whether what I absorbed was right. A confidently wrong suggestion, delivered at a pace I can handle, would score exactly as well on cognitive load fit as a correct one delivered the same way. That's not a flaw specific to my timer; it's built into the dimension itself — and I suspect, with no data to back it, that fluency gets optimized for commercially as much as pedagogically, because it's the easier of the two to measure and sell.

What the timer buys me, though, is room to do the thing cognitive load fit can't do on its own. Slowing down is also when I bring my own expertise to what I'm asking for in the first place, and when I actually have the space to check whether what came back holds up, rather than accepting it because it arrived clearly and on schedule. The same pause that restores fit is the pause where fluency gets checked against validity instead of mistaken for it.

Pace makes fit.

Observer's Insight

Watching an AI answer instantly, it's tempting to treat speed itself as the value being delivered — faster is better, more directions offered is more helpful. Cognitive load fit is the ATPM dimension that pushes back on that instinct hardest: the value isn't in how fast the answer arrived, it's in whether the person receiving it still has room to think. An AI's rate of output isn't, on its own, a measure of how well it's partnering with anyone.

The timer doesn't slow the AI down. It slows down how much of the AI's output I let myself be handed before I've caught up with what I already have. I didn't have language for that distinction until I'd built six dimensions and had to ask, of my own working hours, which one kept coming up short. It turned out to be the plainest one on the list.

Support Author: As a member of the Springer Nature Author Affiliate Program, I may earn a small commission from purchases made through this affiliate link to my Springer Textbook (Artificial Intelligence in Public Health): https://tidd.ly/4mH9389