AI & Cognitive Debt 🧠

A personal reflection on using AI for knowledge work

Hey, Luca here, welcome to a monthly free edition of Refactoring! To access all our articles, library, and community, subscribe to the full version:

Resources: 🏛️ Library • 💬 Community • 🎙️ Podcast • ❓ About

---

This is a monthly free article. I tend to use free articles to write about things I care about the most and I want everyone to read, and this is one of those.

So, there has been some time, a few months ago, when this picture was pretty much everywhere on my feed 👇

It is, apparently, Rio de Janeiro in the seventies. Of course it may be not—I don’t need to explain how internet works—but details don’t matter too much here.

What many people notice about this photo is that everyone seems to be in good shape. And I guess, especially for US folks (allow me this small stab, as an Italian), this is in stark contrast to how people at the beach look today.

Now, this is from a time in which going to the gym was less of a thing, we knew a lot less about nutrition, and we can imagine people generally thought less about what makes for good health, compared to today. Yet, folks somewhat naturally ended up in a healthy lifestyle.

Compare it to today, where being in good shape is obviously still possible, but it feels more like the result of intentional effort and practices that often go against what our natural lifestyle would lead us to.

In other words: it’s hard!

Now, even if you don’t 100% agree with this premise, it is useful to think — and it must be true in some cases — that changes in society (or technology) bring new problems, some of which challenge parts of our lives that, before that, we didn’t have to think a lot about.

To some extent, I worry that AI is going to do just that for our cognitive skills.

That is, it is not going to necessarily make us dumber — but it will require us to develop good practices for our cognitive health, similarly to what we eventually developed for our bodies.

At this point, the obvious question becomes: what are these practices going to look like?

I don’t feel we are ready for a full framework here, so I will just speak about my experience with AI in recent months 👇

---

🤖 Luca & AI

Lately I have been trying to use AI more and more to do all kinds of stuff: from running Refactoring things, to writing code for my countless side projects. I say try because for someone like me who has spent, in many cases, 10+ years doing things the same way, it’s not always easy (or, more accurately, it’s always not easy) to change these very ways.

I am also aware that this inertia creates all kinds of biases, like thinking “this doesn’t work” when in reality it’s just my brain begging me to revert to old workflows. But all-in-all, outcome-wise, AI is a net positive to almost everything I tried to use it for.

How can I be sure?

Well, it’s usually one of two things: it either makes me faster, or allows me to create higher quality output. Many engineering teams I know struggle at measuring these, but I have a big advantage: as a content creator, my work is extremely repetitive, so it’s trivial to see impact on any of these dimensions.

So I can say with confidence that AI almost always helps with at least one of the two.

However, other than pure output, I have also been trying to pay attention to how using AI makes me feel about doing work. And here is the concerning part: I often feel… dumber.

Dumber than when I don’t use AI, that is.

“Less engaged” is probably more correct than “dumber”, but the result is the same: my brain enters a seemingly different mode, a lazier one, which is hard to steer back into full active mode.

(this feeling seems to be confirmed by recent research about writing essays with vs without ChatGPT, which contributed to trigger me to write this post)

I call it create mode vs review mode, and it’s not just about AI 👇

---

🎨 Create mode

Your brain goes into create mode when you need to produce an output as the result of some non-obvious procedure. Think of writing an algorithm, an essay like this one, or even an important email.

You can’t just follow some steps: you need to make connections between ideas in your brain, and turn them into actual output.

Create mode is hard and draining, but also somewhat rewarding. To continue with the body analogy, it feels like a workout.

I feel that’s because every time you create these connections you are small-scale rewiring your brain, growing or improving some model of the world you maintain in your head. I am fond of saying writing is thinking, but you can pretty much replace writing with any creative endeavor.

🔍 Review mode

Review mode, instead, is when your brain doesn’t need to create from scratch: it compares a draft output against a set of implicit or explicit rules about what it should look like, and improves it or continues it based on them.

There are a lot of review-mode tasks that are actually called review something, so it’s easy to think of examples, but also choosing between alternatives, or continuing a sequence of items, are usually review mode.

Review mode feels 10x cheaper than create mode—in terms of energy expenditure—and is very valuable: as humans we are often extremely good and fast at spotting details and steering output in a tactical way.

It’s also limited. The most obvious limitation is that it’s hard to radically change course once you have a draft in front of you. The draft acts as an anchor, and I find that people generally overestimate the amount of critical thinking they are able to apply when they are put in front of one.

Daniel Kahneman, in Thinking, Fast and Slow called it the anchoring bias, and provided a hilarious example 👇

In an experiment, Kahneman asked participants to estimate the percentage of African countries in the United Nations. Before answering, they spun a wheel that landed on either 10 or 65. Those who saw 10 guessed an average of 25%, while those who saw 65 guessed 45%. The initial number, despite being random (and they knew it!), significantly influenced their estimates.

(for those familiar with Kahneman’s work and thinking create mode vs review mode is just a riff over system 1 vs system 2, I don’t think it’s the case — even though there is some overlap).

So, given that today AI can have a shot at most knowledge work tasks, the fundamental choice becomes: when should I go for create mode vs review mode?

Also, this might appear like a false dilemma, because you might think: can’t I use AI to create a draft and then engage with it in a thorough, create-mode-like way, and get the best of both worlds?

In my experience, no — this is just wishful thinking.

Having a first version of anything in front of you radically restricts the surface of what you could come up with by yourself, whether you realize it or not. I have seen this first-hand: whenever I try to write an article using more assistance from AI, for producing a first outline, or fleshing out some sections, it feels very hard to meaningfully diverge from what the AI produces.

There’s a mix of anchoring, sunken cost, and genuine laziness that kicks in, and that feels inescapable.

And even worse: I feel I can’t be trusted to discern this. I may slip into this semi-conscious, pure tactical reviewing without realizing it, all while thinking I am doing top-notch work.

So, are we doomed? Or is there a healthy way to use AI for knowledge work?

I think there is hope, so I will share my own heuristics 👇

---

⚖️ How to decide when to use AI

First of all, not all tasks deserve create mode.

Some of our work—maybe most—is repetitive, uncreative, and doesn’t generate meaningful connections in our brains. For grunt work, AI is a godsend and we should make an effort of using it as much as possible.

For actual knowledge work, instead, things are more nuanced.

With some degree of simplification, let’s assume any task needs a combination of two skills:

• 🏛️ Knowledge — things that one should know about the problem domain.

• 🧠 Reasoning — real-time computation about how to apply such knowledge to the task at hand.

These skills are completely orthogonal, yet I find many people often confuse the two, and say an AI (or a person) is smart, when they just know a lot of stuff, or vice versa.

Instead, if you correctly separate the two, it’s pretty obvious where humans and AI respectively stand today:

1. AI knows a lot more stuff than any human ever will, but

2. Humans are still way smarter than AI

In fact, AI still hilariously fails at simple pure reasoning tasks, just like humans are not very good at remembering things.

Now, since most real-life tasks require a combination of knowledge and reasoning, my main heuristic for working well with AI is: outsource the knowledge, protect the reasoning.

Sometimes it’s hard to separate the two, so you can ask yourself:

Am I using AI because I genuinely don’t know something, or because I don’t want to think through it myself?

The former is generally better than the latter. In fact, another litmus test is to ask yourself:

What happens if I make AI do this all the time?

When it’s for asking for things you don’t know, it’s good: you also learn more things as a byproduct. But when it’s for having a thinking partner, as many like to call it, well… that’s sketchy. Chances are you are giving up some cognitive workout.

That said, this is not black or white.

We may have too many opportunities for create mode in our lives than our energy allows. Letting AI help with some of that, which otherwise we wouldn’t have done anyway, is arguably good. But this doesn’t mean that every time AI can do something, we should let it do it.

There are also some tricks, and for sure we’ll find more over time. For example I have found value in creating a first draft myself and making AI review it — instead of vice versa. This way I limit the anchoring effect and still take advantage of AI’s superior knowledge.

Funnily enough, this way AI becomes the one that suffers from anchoring. AI is notoriously not good at disagreeing with us, whether because of its sycophant alignment, or genuinely because of how the model works. So I have found that the best way to “review” something I have done is to make AI come up with the same thing, but from scratch, without looking at my draft.

So, that’s pretty much it! I have no wise conclusion for this piece — it’s a wild time, and the best we can do is just keep using these things, with our eyes wide open and while paying attention not only to what we create, but also to how we feel about it.

---

And that’s it for today! I wish you a great week.

Sincerely 👋
Luca