#70 | Did you get better, or did AI?

TL;DR: Has your AI work gotten better over the past year? Probably. But there’s a question most AI users never ask — was it you, or was it the model?

👋 Hello,

Last week, I reviewed a piece I’d written with AI assistance.

I genuinely couldn’t tell where my thinking ended, and the model’s began. Ok, it wasn’t unsettling in a dramatic way. More like realizing mid-conversation that you’re not sure which ideas or opinions you actually formed yourself.

Do you know that too? It’s a strange feeling. And it turns out to be pretty common.

Looks like there’s a problem with getting better at AI tools.

At some point, you stop being sure whether you got sharper or whether the model is just doing more of the work. Your brain strongly prefers the first story :), so it goes with that one, quietly, without flagging it.

Sure, your outputs have improved over the past year — that part is probably true. Client feedback is better, or simply fewer revisions, when you create content.

But here’s what you likely haven’t measured: did you improve, or did the model? I never thought about it in detail, to be honest.

Researchers at DDai Inc. published a ​preprint this month​ naming this exact problem.

They call it the LLM Fallacy. You produce something good with the model’s help. Your brain quietly files it under things I’m good at. The model’s contribution disappears from your mental accounting.

Fine in isolation. The problem is, it keeps happening.

Worth noting: the paper hasn’t been through peer review. The researchers say so themselves and call for more testing. But it puts precise language on something AI users are already sensing — and that makes it worth reading.

The mechanism

You write a prompt, and the model returns something polished. You scan it, adjust a word or two, and send it. Your brain logs: great work [your name here] today. (Get a drink!)

But there was no clean line between what you did and what the model did. You brought the direction and judgment. The model handled the execution. That distinction gets blurry fast — and stays blurry.

Part of this has a name in learning research: the fluency illusion.

When something reads smoothly — confident, well-paced, contextually right — your brain takes that as a sign of competence. The output sounds like your best work. So it feels like your best work.

That feeling is real. It’s just not a reliable signal.

But beware. Something quieter builds alongside it.

Every time the model does the heavy lifting, you skip doing it yourself. The skill — whatever shape it takes in your work — goes unpracticed.

Friction is how expertise knows it is working.

Struggling through a hard argument is not inefficiency. It’s what keeps the skill sharp. Finding your own way to a conclusion you’re not yet sure of — that’s the exercise.

Remove it, and you remove what maintains the capability. Outputs still get produced. The underlying skill just stops being used. Let’s look at this example.

What’s already been measured

A ​2020 study in Nature​ tracked GPS use and spatial memory over three years. The more participants used GPS between check-ins, the larger the drop in unaided navigation.

Steady decline, not a sudden cliff. The more use, the more loss.

This is the part that surprised researchers: weak navigators didn’t decline most. People who’d had strong spatial skills lost more. They had more to lose.

Aviation found the same thing. The FAA concluded in 2013 that the problem was structural:

“Over-reliance on automation may weaken a pilot’s ability to maintain aircraft control manually.”

The International Air Transport Association confirmed measurable skill loss across a significant share of surveyed pilots. Monitoring a system that flies the plane is real work. Flying the plane is different work.

Same principle on your desk. When the model produces what you used to produce yourself, the output looks identical.

What’s changed is invisible — until the moment the tool isn’t there.

The score you’re keeping

There’s a ​consistent pattern​ in how practitioners account for AI-assisted work. Good outputs get attributed to your judgment. Missed outputs get attributed to the tool.

Run that across months, and you build a self-image that keeps tilting — regardless of what your actual capability is doing underneath it.

What makes this hard to see is that good output is exactly what makes the illusion convincing. If the model produced mediocre work, you’d catch it. You’d rewrite it, right?

You’d notice the gap. The problem is the model doesn’t produce mediocre work — not when you’ve spent real time learning to use it well.

The better the output, the less occasion there is to find your floor.

AI literacy research makes this sharper. Among regular AI users, the classic pattern — where the least skilled overestimate most — disappears entirely.

Everyone overestimates. More striking: the more AI-literate someone is, the more overconfident they become about their own performance.

The people who know the tools best are most exposed to the illusion. Great future ahead!

That includes me. I’m aware of the irony of writing this in a newsletter I drafted with AI assistance.

The researchers who named the LLM Fallacy used LLMs to write it—and disclosed this in the paper itself. Something almost admirably honest about that.

Also, something quietly uncomfortable for anyone working this way, myself included.

What calibration looks like

The research doesn’t argue for using AI less. The question it leaves you with is narrower: when did you last test what you can actually do without it?

GPS spatial memory declines with use — and comes back with deliberate practice.

Pilots who returned to manual flying recovered their skill. Not permanent damage. A maintenance problem. The difference between something lost and something that just hasn’t been used.

Calibrating isn’t about proving you can work alone. It’s about locating the gap — finding out whether the distance between what you bring and what the model brings has grown without you noticing.

Here’s the practical version. Pick one task you use AI for every week — a specific deliverable, an analysis, something you’d normally think through with the model.

Do it without the model. Produce something you’d actually send. (Something real, not a practice run.)

Do it once. Then ask: Is this meaningfully different from what I’d produce with AI assistance?

If the gap is small, good news — the model is amplifying your capability. Your judgment, its execution, together produce something neither does alone.

If the gap is larger than you expected, the model has been carrying more weight than you realized.

The capability you thought you were building has already been built elsewhere.

Neither answer is a verdict. Both are better to find now — quietly, on your own terms — than in front of a client.

That line moved. Most of us haven’t gone back to check where it is now.

Cheers,

Mark
The AI Learning Guy
👋⚡😎

Interesting Sources

• ​The LLM Fallacy preprint​ — arXiv / ddai Inc.
• ​GPS use and spatial memory​ — Nature Scientific Reports
• ​Manual flying skill degradation​ — CARI aviation journal / FAA
• ​GenAI and critical thinking​ — Microsoft Research
• ​AI speed versus skill​ — INNOQ

Note: No single website has all the answers. This list serves as a starting point for those who want to explore or satisfy their curiosity about AI. Links: Links with * are affiliate links. See disclosure below.