According to their study, the people with the highest cognitive sovereignty scores were not the ones using AI the least. They were the ones using it the most: fully integrated, identity reorganized around AI and maximum engagement on every dimension measured. The researchers call them co-authors and I think it fits.

The lowest scores belong to those who kept AI at arm’s length and delegated work to it. Outsourcers.

I watched it expecting to find something that would challenge what I’d been feeling, namely that AI is augmenting me, not taking away from me. Instead, I found the beginning of an explanation to what I had been feeling.

Why this makes sense to me

My colleague Pascal Hamel likes to say “AI is only as good as your ability to tell it precisely what to do” and he’s right. An LLM is not a thinking machine. It’s a prediction machine. It predicts the most likely next word in a sentence, at enormous scale, over vast training data. Someone, and I beat myself up for not remembering where I saw this, framed it well: AI doesn’t write good code, it writes probable code. Why it matters? Because AI is not thinking outside the box, by definition and training it thinks inside the box.

When you engage deeply with AI, you learn exactly that. You learn what it does well and where it falls short. You develop a sense for when its output is just statistically average and when it’s actually useful. You bring the judgment it cannot have.

When you delegate to it, you never develop that sense. The box becomes invisible because you’re never inside the process. Don’t get me wrong, delegation is fantastic for some things, but not all things.

Think about how the best developers use AI coding tools. They don’t just prompt and accept. They architect the harness: context files, rules, test suites, feedback loops that shape what the agent can produce. The harness is human judgment made explicit and persistent. You can’t architect a good one if you’ve checked out of the process, which brings us back to the same variable: are you inside it or outside it and did you choose your position deliberately or by happenstance.

The three things worth protecting

The research also frames cognitive sovereignty around three components and as far as I’m concerned, all three are under pressure.

Awareness is the first to go when you’re moving fast. It’s easy to pawn off the work and reclaim a few minutes, or a few hours, but awareness is how you stay calibrated. When you stop noticing what AI gets wrong, you also stop noticing what you’re no longer contributing.

Agency is the balancing act. A fully autonomous agent maximizes throughput and minimizes your involvement. That’s useful in places, but it’s worth asking, for each workflow where human judgment would create value? The answer is rarely nowhere.

Accountability is the one I find easiest to hold. Everything I produce with AI is my work, full stop. I understand the pull toward the opposite when you’re under pressure, when AI was introduced into your process without much say, when you’re delivering against a deadline. The more engaged you are, the harder it becomes to disown the output.

What this means if you lead people

The outsourcer profile is not rare. It shows up as the person who treats AI as magic, who believes it will replace everything quickly by automating every function and make human input redundant. What I believe they fail to take into account is the value of human judgment and where it fits. Perhaps they skipped the hard work of staying inside the process?

Not everything gains from having a human in the loop. Summarizing the newsletters I skim each day? Delegated freely. For the work that carries judgment, the question for leaders is not whether to adopt AI or not, it is how to build teams that stay inside the process, because that is what separates a passing grade from an outstanding one.

So yes, in that sense, cognitive sovereignty doesn’t come from using AI less. It comes from using it with your eyes open but that does not mean co-authors are the golden standard, there is a trap lying in wait for them. As we get better and more fluent using AI, we must not stop questioning what the process is yielding or we risk losing the critical distance that forced us to get better in the first place.

I’ve felt the exhilaration of seeing a language model answer questions and flirt with reasoning, just as I felt the disappointment of realizing it was only predicting the next word with uncanny accuracy. The ups and downs of generating images when models were still finding their footing. Reading on about the ethics of it all, unsure whether to feel excited or unsettled, usually both at once.

Here’s the thing about following AI right now: there’s always something new to learn, and last month’s breakthrough is already assumed knowledge for some but will remain new for most. Under those circumstances, it’s easy to feel like you’re behind and the fact it moves fast in a way that makes it hard to hold onto what you’ve actually figured out. You spend so much energy keeping up that you rarely stop to notice how far you’ve come, how much you’ve learned and how much you may very well have to share.

That’s the first reason I’m writing this. Not to have answers, but to have a record of the questions I asked myself. A place to document what I’m experimenting with, what surprises me, what turns out to be noise and what ends up working at least for a time. The journey is worth tracking, especially when the destination keeps moving.

The second reason is simpler: I’ve been meaning to start writing again for years. AI gave me enough material to finally stop procrastinating, and more than that, it did the heavy lifting.

Three days ago I had no clue what Hugo was, Claude recommended it. I was aware Cloudflare had multiple services, but I did not know I could host a static blog there, published by a simple git push. Within 40 minutes, I had a local setup and a pipeline to publish online to my own domain. Claude did most of the work, I supervised and configured the tools. At some point, I also had to correct it because while AI is nice, it gets confused easily when UIs change. Bridging the gap between what they know, or used to know, and what you see is still a speed bump.

So, here we are: me writing on a Claude-suggested tool, published on a Claude-suggested host, using a Claude-suggested publishing pipeline, and you reading a text written by me from a draft written by Claude, based on a prompt by me.

Will you get value ouf of this or anything else I write? I hope so but I won’t guarantee it. If you do however, like and share as they say or just hit me up if you want to have a chat. Until then, keep being curious.