I was working on a detailed research project recently, using Claude to support the process. The work was moving fast, faster than it had any right to. I was making connections, surfacing themes, building arguments. And somewhere in the middle of all that, I caught myself thinking: “I really understand this material now.”
Then I paused. Did I? Or did I understand what Claude had organized for me?
That distinction is small, almost imperceptible. It’s also, I’d argue, the most consequential question facing learning leaders today. Because if I can fool myself after 25 years in this field, every employee in every organization can fool themselves too. And they probably are.
Welcome to unconscious incompetence at scale.
The virus we’re not talking about
Noel Burch’s classic competence model indicates that we start unconsciously incompetent: We don’t know what we don’t know. We become consciously incompetent when we recognize the gap. We then work our way to conscious competence, where we can do a thing with effort. Eventually, with enough reps, we reach unconscious competence, where it’s second nature. The important distinction here is the difference between I have read, I have understood and I have performed.
Unconscious incompetence is different and more dangerous: I don’t know that I’m wrong or that my capability falls short. I swear the plural of octopus is octopi, or I’m sure I can weld commercially because I watched my father weld in our garage. I’m not aware there’s anything to learn.
What I am calling the fifth stage of competence is really a hyper-enabled state of unconscious incompetence catalyzed by collaboration with artificial intelligence. What gets a little funky between the traditional unconscious incompetence stage and this new fifth stage I’m proposing is the journey. The end state is actually the same, but the path there is different.
First, I realize I don’t know something. Then I work with AI to generate work that indicates I know something. It may be technically accurate, but do I really understand it? I have likely produced high-quality work without having developed the underlying capability necessary to truly understand it. It sure feels like competence. It performs competently. But pull the tool away, and the capability isn’t there anymore.
This is the virus I want learning and development leaders to take seriously. AI is extraordinary, but the same features that make AI such a powerful accelerator—synthesis, summarization, pattern recognition, instant analysis—bypass the cognitive work that creates real learning. When AI delivers the conclusion, we lose the muscle that knows how to build arguments.
What we end up with is a workforce that’s increasingly productive and decreasingly capable. That math doesn’t work for any business. And it especially doesn’t work for ours.
The loss of friction
Real learning requires friction and effort, and I think we underestimate just how much. It requires the productive struggle of working through something hard, getting stuck and finding your way back out. The brain encodes what it has to work for. Easy in, easy out.
AI removes friction by design. That’s its whole value proposition. And, in plenty of contexts, that’s exactly what I want. Eliminating low-value friction frees us up for higher-value work.
Here’s the trap, though. Not all friction is wasted effort. Sometimes the friction is the work. The struggle to find the right framing for an argument, the wrestling with conflicting data, the slow process of building a mental model. That is not an obstacle to expertise. That is the path to it.
What we have to build, in ourselves first and then in our organizations, is the ability to tell the difference between when to use AI to remove friction and when to preserve friction because removing it would remove learning along with it. Essentially, the question is how to introduce positive friction into the creative process with AI.
How I preserve friction in my own work
Here’s what this looks like in practice for me right now. My recent research project involved reading through tons of source material, identifying themes and building out schema that would hold up across dozens of texts. It’s exactly the kind of work AI is extraordinarily good at accelerating. It’s also exactly the kind of work where letting AI accelerate me too aggressively would mean I never actually learn the material.
So I’ve developed a layered approach. The first pass through any source is mine alone, no AI allowed. I read it, take notes and write out my own logic or sense-making before I open Claude. It’s slow and often frustrating. That’s the point: I need to know what I think before I let a model influence what I think. Only then do I bring AI in, and not to do the work but to challenge mine, to find what I missed and argue against my scheme. The friction stays intact because I’m still doing the lifting.
This isn’t efficient, by the way. A pure AI-first approach would be faster. But faster isn’t the goal. Faster while still learning is the goal. Those are very different things.
Anchoring AI to how learning moves
On a strategic problem like positioning a new offering or responding to a market shift, I write my own frameworks and form my point of view before reading what the AI models have synthesized. I use AI to sharpen my thinking, not stand in for it.
What holds all of this together is anchoring the work to how learning actually moves rather than to whatever the tool produces first. Bob Mosher and Conrad Gottfredson’s 5 Moments of Need offer a useful map here: The moment we encounter something new, the moment we need more, the moment we apply it, the moment we have to solve a problem and the moment something changes. The point is to move through those moments myself rather than let a model move through them for me.
In practice, that means starting with my own questions:
- What should I be asking?
- Which sources matter? Who thinks so? And why?
- How do the positions compare, and where do they break down?
- How is this applied, and where does it commonly fail?
Posing those questions well is already part of the learning. From there, AI helps me build an expression of my own understanding rather than simply supplying one, and I keep asking it to challenge both itself and me.
As my thinking takes shape, I bring in human experts and give them the reasoning behind the work, not just the result. That way, they push on the logic rather than polish the surface. Then comes the step most easily skipped: I apply it, against reality or a case example built to test it. That is where I have read, I have understood and
