People close to me have formed a view about what working this deeply inside AI systems does to a person. It is a reasonable view. It is also wrong.
I spend most of my working life embedded in AI systems. Not using them the way most people use them, to draft an email, to summarise a document, to answer a quick question. I am building the infrastructure. Designing how agents reason, how they receive context, how they hand off between systems, how intelligence flows through an architecture rather than sitting in a single model. The work requires me to think inside these systems, not just think with them. To understand how they behave when you change what they know, how they fail when the context is wrong, how decisions made downstream are shaped by the structure you built upstream.
By every conventional measure, I am exactly the person people worry about.
So when Helen Edwards of the Artificiality Institute published their Cognitive Sovereignty research earlier this month, I read it expecting confirmation of the concern. I expected the data to show that heavy users were eroding something. Losing sharpness. Becoming dependent.
The finding was the opposite.
The people most deeply integrated with AI scored highest on cognitive sovereignty. The people who kept AI at arm’s length, occasionally delegating tasks and accepting outputs without engagement, were the ones most at risk of losing themselves.
The variable that mattered was not how much they used AI. It was the character of the integration.
I was not surprised. I was relieved that someone had finally measured it.
What the research actually found
The Artificiality Institute analysed 1,250 transcripts of professionals discussing how they use AI in their work. Not whether they use it, or how much. The question was how: how AI sits inside their thinking, their professional identity, and their sense of what their work means. The sample covered workforce professionals, scientists, and creative practitioners. The methodology was AI-assisted qualitative coding at scale, each transcript assessed against a detailed framework with specific evidence cited for every classification.
What they built from that analysis is not a simple spectrum from light to heavy use. It is a three-dimensional framework.
The first dimension is Cognitive Permeability: is AI actually inside your reasoning? Not handling tasks while you think separately, but entering the iterative process itself, where the model’s responses redirect your thinking in ways you did not plan. The second is Identity Coupling: has your professional self reorganised around AI? Not just using it constantly, but reaching a point where you cannot describe your professional identity without reference to it. The third is Symbolic Plasticity: are the meanings and standards of your work in flux through AI engagement, genuinely open for renegotiation?
Three binary dimensions produce eight distinct integration roles. The data show where professionals actually sit.
The two most common roles together account for nearly two thirds of the sample. Framers, at 33.4%, let AI into their reasoning while keeping their professional identity and sense of what good work means entirely stable. Doers, at 30.3%, keep AI strictly outside their reasoning process, using it as an execution tool rather than a thinking partner. Scientists cluster heavily in the Doer category, reflecting a professional culture built around independent hypothesis generation and verification.
At the far end sits the role that carries the most important finding.
The number that should change your thinking
Only 3.8% of the sample had reached full integration across all three dimensions: AI inside their reasoning, professional identity reorganised around it, and the meanings of their work in active flux. The researchers call this the Co-author role.
Co-authors scored 8.54 out of 9 on cognitive sovereignty.
Outsourcers, people whose professional identity had coupled to AI but who kept it outside their actual reasoning, scored 6.43.
That is a 2.11 point gap driven by a single dimension: whether AI was inside the thinking or outside it.
The researchers define cognitive sovereignty through three components. Awareness: actively monitoring your cognitive relationship with AI in real time. Agency: your goals and delegation choices coming from your own judgment. Accountability: owning outcomes regardless of AI involvement. For most people, these three cluster tightly. 98.6% of participants scored within one point of each other across all three components. Sovereignty functions as a bundled capacity. You are either the author of your cognitive life or you are not.
What the Co-author data confirms is this: the integration itself requires sovereignty. When you are actively shaping how AI sits inside your thinking, choosing what to delegate, monitoring output quality, maintaining final say, you are exercising awareness, agency, and accountability by definition. The depth of integration does not erode those capacities. It demands them.
What builders already know
There is something that becomes obvious when you spend serious time designing AI systems rather than simply consuming their outputs. Presence is not optional. It is the work.
When you are building an agent, you cannot stand outside the process and collect results. You have to understand what the system is reasoning about, where it is making assumptions, what it does when the context shifts. You have to hold the architecture in your mind while the system runs. The moment you step back and treat the output as a black box, the quality of what you are building degrades immediately and visibly.
This is not a discipline you impose on yourself. It is what the work demands.
What the Artificiality Institute’s research confirmed is that this dynamic is not unique to builders. It applies to everyone. The Co-authors in the study describe iterative back-and-forth collaboration, not one-shot queries. They customise AI to their needs. They maintain editorial control and creative ownership. They use language like director, mediator, and creative partner to describe the relationship. A fiction writer in the study says: writing is a part of me and I want the work to be mine. A game developer describes going with their intuition when they disagree with the model. A manager describes using AI-mediated reflection to shift from a punishment-and-control style to something more considered.
Sovereignty is high because they stayed inside the process. The integration deepened their authorship rather than replacing it.
The two ways sovereignty fractures
The research also identified something the headline finding can obscure. Sovereignty does not just scale up or down uniformly. Under specific conditions it fractures, and the pattern of fracture tells you precisely what failure mode is in play.
The first fracture pattern is seeing without acting. This is concentrated among content workers, writers, editors, and others whose core output AI now produces cheaply at scale. Their awareness is intact: they can articulate exactly what AI is doing to their profession. But their agency collapses. They see clearly and feel stuck. In the researchers’ framing: they can read the story of what is happening to them but they cannot pick up the pen.
The second fracture pattern is owning without seeing. This appears in professionals who take full accountability for outcomes but are not actively monitoring how AI is shaping their cognitive process. They hold the pen, but they are not watching what the pen is doing to their handwriting.
These are different failures of cognitive authorship. They are not points on a single scale of more or less AI use. And they call for different responses.
What drift actually looks like
Drift is a thousand small surrenders that never feel like decisions.
It does not happen as a single choice to stop thinking. It happens as a sequence of frictionless handoffs, each one entirely reasonable, none of them feeling like a decision. You ask the model to help you think through a problem. It offers a frame. You accept it because it is good enough and you are busy. The next time you face a similar problem you reach for the same frame without noticing you are doing it. Three months later someone asks how you approach that kind of problem and you describe the model’s framework as if it were yours. You are not wrong exactly. But somewhere in that sequence the thinking stopped being something you did and became something you collected.
Discomfort is the signal we rely on to recognise loss. We feel the moment we can no longer remember a number we used to know. We feel the moment a word we once reached for easily is no longer there. Drift produces no such signal. The AI fills every gap before you experience the gap as a gap. The erosion is invisible until it is substantial.
The research makes the mechanism precise. The Outsourcer role, identity coupled to AI but reasoning kept separate, scored the lowest sovereignty of any group: 6.43. These are not people who avoided AI. They are people who let their professional self reorganise around it without ever engaging their thinking with it. Distance was not their protection. The absence of cognitive engagement was their exposure.
Presence is not a feeling
Presence does not mean feeling engaged. It does not mean typing more, spending longer on each prompt, or deliberately adding friction. It means your judgment is in the loop at every step, not just when you collect the result.
There is a moment most people who use AI seriously will recognise. You are working through something, going back and forth with the model, and at some point you realise you have stopped leading the conversation and started following it. The AI offered a frame. You accepted it. The AI suggested a direction. You went there. You are still typing, still apparently active, but the thinking stopped being yours several exchanges ago.
That moment is where the research lives.
The subtler layer is this: AI shapes the question before you have even asked it. Most people assume the influence runs one way, the output shapes what you believe. But the interface shapes what you think to ask. The defaults, the suggestions, the way a blank prompt feels, all of it quietly conditions what seems worth pursuing before you have typed a single word. You can be drifting before the conversation has started.
Staying present means noticing both. It means asking not just whether the answer feels right, but whether the question was yours in the first place.
You are not the author. You are the publisher.
Authorship implies agency over the process, not just the product. An author does not simply produce words. An author makes choices about what to say, how to say it, what to leave out, what angle to take. The product is the residue of a thousand small decisions, each of which reflects something about the author’s judgment and understanding.
If AI is making those decisions and you are collecting the output, you are not the author. You are the publisher.
That distinction matters because your mind is not shaped by the products you consume. It is shaped by the decisions you make. The thinking you do. The judgments you exercise. The moments of difficulty you push through rather than hand off.
The researchers frame it clearly. The critical finding from the Co-author data is not that integration is safe. It is that the integration itself requires you to be the author. When you are actively shaping how AI sits inside your thinking, you are exercising cognitive sovereignty by definition. The capacity and the practice are the same thing.
What determines the trajectory of your mind is not how frequently you use AI. It is whether each interaction has an author in the room.
Why this is harder than it sounds
I want to resist the temptation to make this feel achievable through a simple habit change. Stay engaged. Review the output. Ask more questions. These are reasonable practices and they are not wrong.
But the structural pressure runs in the opposite direction, always.
AI interfaces are built for frictionlessness. Every design decision in every major product is pointed at reducing the distance between your question and an answer. That is the product. That is what people want. And it is precisely what makes sustained presence difficult.
Think about what good management actually requires. You do not manage a person well by collecting their outputs at the end of the week. You stay inside the work with them. You understand how they think, where they make assumptions, when to redirect and when to let them run. The managers who drift into pure output collection lose touch with their teams, miss problems early, and find the capability they thought they had was never really theirs to rely on.
The parallel with AI is exact. Except that with a person, social friction keeps you engaged whether you choose to be or not. They push back, ask questions, have difficult days. That friction is protective. It keeps you in the loop involuntarily.
AI removes that friction entirely. The system will never demand your attention or signal that your presence has lapsed. It will simply do whatever you asked, whether your thinking was in the room or not. Which means presence with AI requires more deliberate intent than managing even the most self-sufficient person on your team.
Your brain is designed to offload whatever it can. That is not weakness, it is how cognition works. We built tools specifically so we could stop carrying things in our heads. The problem is that AI is a more complete offloading device than anything we have built before, and the completeness of the delegation is invisible until the capacity is gone.
Staying present is not a one-time decision. It is a repeated, quiet act of insisting that your thinking stays in the room. And that requires a different relationship with your own mind than most of us have been trained to have.
We were trained to value results. Outputs. Deliverables. Speed. AI delivers all of those more efficiently than we can. If results are what you optimise for, full delegation is rational.
But your mind is not a result. It is the thing that does the resulting. And that distinction does not appear on any dashboard.
What building taught me
I do not let AI close a thought for me. I will ask it to develop an idea, push back on a framing, generate alternatives I have not considered. But the closing judgment is mine. Every time. Not because I distrust the model but because the act of closing a thought is where my thinking lives. It is the moment of synthesis, of weighing, of deciding what I actually believe. Handing that off is handing off the thing that makes the thinking mine.
I notice when I am about to delegate from outside. There is a quality of attention, or inattention, that precedes it. A slight detachment. A sense of watching the process rather than being in it. When I feel that, I slow down. Not to reject the AI. To re-enter it.
I talk about what I am building with people who will push back. The AI will never tell you that your thinking has become lazy. A person who cares about your work will.
None of this is a methodology. It is closer to an orientation. The difference between someone who works with these systems and someone who is worked by them is not technical skill or frequency of use. It is whether they stay inside what is happening.
The Artificiality Institute’s research gave that orientation a name, a framework, and now a number. 8.54 out of 9. That is what it looks like when the integration goes deep and the author stays in the room.
The only thing that cannot be replaced
The debate about AI and the mind has been almost entirely about outcomes: are you becoming dependent, are you losing skills, is the technology making you worse at things you used to do well. These are real questions but they are downstream.
The research gives us the upstream answer. Deep integration with presence: 8.54. Identity coupled without cognitive engagement: 6.43. The gap is not about how much. It is entirely about whether the author was in the room.
But there is something the data cannot capture, something that only becomes visible when you have spent enough time inside these systems to feel it rather than measure it.
Your mind is the only thing in this entire arrangement that cannot be replaced. The models will improve. The architectures will change. The interfaces will evolve in ways none of us can fully anticipate. Everything in the system is in motion except one thing: the requirement for a human being to be present inside it, bringing judgment, meaning, and authorship to what would otherwise be very sophisticated pattern completion.
The question is not whether AI will change how you think. It already has.
The question is whether, the next time you open that conversation, you will be the one doing the thinking.
The research behind this piece
This article draws on original research published by the Artificiality Institute in March 2026.
Cognitive Sovereignty: Authoring Your Mind in the AI Age Helen Edwards, Artificiality Institute Read the full paper
The Artificiality Institute is a non-profit research organisation studying what AI actually does to people: not to jobs or industries, but to minds, relationships, and identity. Founded by Helen Edwards and Dave Edwards, formerly of Intelligentsia.ai (acquired by Atlantic Media) and now Visiting Researchers at UC Berkeley’s Center for Human-Compatible AI, their work is among the most rigorous and honest in this space. If this piece made you think, their journal will take you further.
Artificiality Institute · Artificiality Journal
If this landed somewhere real, subscribe. The next piece takes this to the organisational level: what happens when hundreds of people in the same company are drifting simultaneously, why nobody notices until the capability is already gone, and what it means to build intelligence infrastructure that compounds human authorship rather than replacing it.
Craig Hepburn is an AI strategist and Perplexity Fellow who builds advanced agentic systems. He spent years leading digital transformation at Art Basel and UEFA. Now he works on the harder question: not whether organisations adopt AI, but how they govern it when they do.
I think the findings come fairly directly from understanding that LLMs provide a new cognitive loop.
Think Socrates: ‘writing is bad as it stabilises thought’, but writing provided a new and useful cognitive loop. Writing didn’t replace speech, it complimented it. Ditto with LLMs which provide a new loop: interacting with sedimented human experience.
The risk isn’t depth of integration, it’s whether you understand what kind of thing you’re in a loop with. Socrates’ worry about writing wasn’t really about writing—it was about people treating fixed text as a substitute for living dialectic. The same mistake with LLMs would be treating the output as a conclusion rather than a move in an ongoing exchange.
The people with ‘low cognitive sovereignty’ in the research might simply be the ones who never understood what the loop actually is.