Why Smart People Understand AI but Don’t Change How They Work

Part 1 of 3 in The Friction in AI Progress

Humans have a pretty horrible track record of changing our behavior just because we know we should. I see that in myself every night before bed. I’ve heard and read countless times that looking at my phone can negatively affect my sleep (my tracked sleep scores confirm this), yet I still look at it. That’s a habit I know is bad and keep anyway. My way of working, and where AI fits into it, is both similar and different. Similar, because knowing AI could make it better hasn’t changed it either. Different, because it’s a series of good habits that got me here, so the promise of a better way has even less pull.

In the last article, I named the AI Assimilation Gap, a period when someone is already using AI while familiar work still moves through much the same process. I considered whether unclear expectations or missing skills could explain why that period persists. Both can play a part, yet clarity and capability can be present while the work carries on as before. Just like my phone-before-bed pattern, my inability to “behave” and simply incorporate AI into how I do my job has been nagging at me for quite some time, and in parts of my work it still does. I know it can improve the way I work, yet something I can’t quite see keeps steering me toward leaving things as they are. This makes sense, I guess, and it’s why I want to look at my defaults, the ways of working I stopped noticing long ago.

In the data work I’ve done for nearly ten years, I trust conclusions more when I can reconstruct the path that produced them. Work like this is often called deterministic. If a figure shifts, I want to inspect how records were cleaned, matched, or filtered until I can account for the difference and reproduce the result. That gives me a way to check the answer against the underlying material and correct a faulty step. I’ve spent years perfecting this level of visibility. Being able to explain the result is one reason the process earned my confidence. But despite AI entering the picture and despite seeing its value quite early on, the way I produced familiar work stayed… well, familiar. It took about two years before it changed how I work day to day.

Smart people (and I’m not calling myself one) can understand AI and keep producing work in much the same way, while still using the technology and finding value in it. AI enters a system with its own logic, built around work that has proven to be dependable. Understanding the tool’s capabilities can show us what might be possible, but it says little about the system the tool is entering.

The system we stopped questioning

The way I work, and presumably the way we all work, rests on judgments we’ve made often enough that they’re now hard to notice. And that’s the purpose. Once something is settled, we no longer need to pay it much attention. In David Foster Wallace’s 2005 Kenyon commencement speech, an older fish asks two younger fish how the water is, and one later turns to the other and asks, “What the hell is water?” The same can be true of our working defaults. They’re present in so much of our work that we stop seeing them as choices. When a task arrives, we know where to begin, what to check, when to stop, and which parts are ours to handle. Over time, the need to revisit our approach grows quiet because, most of the time, what we’re already doing works well.

If I had to distill what two decades of refining how good work gets done have taught me, I’d name four core tenets: focus, habits, standards, and identity. Most of us rely on all of them daily and question them rarely. Each reflects professional judgment developed over time, and each helps us produce work we can trust and take responsibility for.

AI doesn’t care about how carefully those judgments have been formed. It presses on them anyway.

When AI enters familiar work, our attention may move differently, our steps may change, our confidence in an answer may weaken, and our sense of what belongs to our role may shift. A task we know well can begin to feel less settled, even while the assignment itself remains recognizable. I’d like to say I can take that in stride, but often I can’t. When AI reaches into familiar work, I find myself reconsidering what I’ve learned to count on and what I’m responsible for. Each tenet raises its own question about the work, and knowing more about AI doesn’t necessarily answer it. The first question concerns where we direct our attention.

Focus: Where does my attention belong?

For much of my working life, productive attention meant giving one task my full focus, protecting that time from interruption, and staying with the work until I could call it finished. Then I moved to the next thing. Attention was something to guard, and concentration meant keeping it in one place. There was good reason to work that way. When a task needs careful judgment, shifting away can break the thread of what we’re doing. Staying with one task makes it easier to notice details and carry a line of thought through. We aren’t multitasking behind the wheel, and we know this. We’re trying to do two things at once, badly. Human attention mostly switches between tasks, and those switches can slow performance.¹

AI can run parts of a task in parallel while I supervise. In one stretch of work, I had an AI audit running across a long document while I worked through a single word elsewhere. The audit kept moving while I shifted attention between its progress and the line I was revising. The work can happen in a shorter span, with decisions gathered around the moments I check, redirect, or accept a result. I’m still attending to one thing at a time, but now productive attention includes deciding when to switch and what deserves the next look.

That changes what productive attention looks like. It can mean keeping several moving parts in view and focusing closely when judgment is needed, rather than staying with one task from beginning to end. Parts of the work can continue while I inspect another, but I still have to choose what to watch and when to intervene. At any given moment, which part of the work needs my attention? Supervising parallel work can feel more scattered and demanding than the single-task focus it replaces. For many of us, that feeling is enough to keep the old way of focusing in charge. And often, before I decide, a familiar routine is already shaping what I turn to first.

Habits: Does this routine still get to run?

Charles Duhigg describes a habit loop of cue, routine, and reward. A cue prompts a behavior, and the reward that follows helps reinforce it. A routine can become automatic through repetition, without telling us much about willpower or character. One study found that roughly 43 percent of daily behaviors were habitual, often occurring in familiar contexts.² That pattern is easy to recognize at work. Month-end arrives, a recurring template is open, or the same file lands in front of us, and a known sequence follows. When I pick up work I’ve done many times, I tend to move through its steps without reconsidering each one. It’s a sort of work autopilot, and we all know it feels amazing to knock out good work without spending precious brainpower on it. The routine carries the benefit of experience, while also narrowing which other approaches come readily to mind.

Part of the reward—beyond praise, saved effort, or, my personal favorite, dopamine—is confidence in the result. With a process I’ve run many times, I know how long it takes, where it tends to break, and which parts deserve a second look. Being able to show my work matters to me; the answer alone doesn’t provide the same reassurance.

AI keeps reopening processes we thought we’d settled. I rebuilt my company’s brand system three times before AI entered the process and four times since. Because I keep letting the process change, at any given moment I’m not sure how I’d actually produce a marketing asset. The routine never gets to settle. At this pace, I should stop writing before the brand gets two versions ahead of the article.

Habit can carry the work forward, and AI can introduce another possible way through it. More often than not, the usual routine keeps winning even after AI has opened another path. I know how easily I return to mine. When should I stop and reconsider a familiar way of working? When the result is ready, I still want to know whether I can explain how it came about and why I trust it.

Standards: Can I claim authorship of this?

I love asking an LLM to help me synthesize something long or complex, and an AI-generated analysis can arrive looking ready to use. When AI handles part of the analysis, some of the work happens out of my sight. The explanation can be smooth and the conclusion plausible, while the decisions that shaped them are harder for me to trace. AI lets us increase our output without increasing our input, but my name still goes on the result. That changes what it feels like to author the work. Standards shape what evidence we expect, what uncertainty we’re prepared to carry, and what we can explain if someone asks how we reached a conclusion. We bring that bar to every analysis, long before any particular answer arrives.

AI can make this harder because the tool that helped produce an answer can also offer reasons to trust it. If I ask the same system to review its analysis, I haven’t gained an independent judge. Its explanation may sound coherent and reassure me that the output is fine, but that reassurance comes from the process I’m trying to assess.

Some outputs can be checked against a result outside the answer. A script either runs or errors out. Numbers reconcile or they don’t. I can inspect those signals. An analysis or recommendation may sound just as plausible, though, without an external check that tells me whether its conclusion has moved beyond what I can judge. I can stand behind work without certainty about every detail, as long as I have enough basis to answer for the claim and its limits. Can I put my name on this? Often, a plausible answer can still feel too uncertain to put our names on, so we fall back on the version we can defend.

Identity: Does this belong in my role?

“I’m not a programmer” did two jobs for me. It described a real limit on what I could do, and it also marked a boundary around the work I understood as part of my role. Even when I can vouch for the quality of a piece of work, that doesn’t automatically make producing it part of my job. Programming sat outside my practical abilities, which made it easy to treat it as outside my responsibilities, too. A role also organizes what people are expected to contribute. For much of my career, the limit in what I could do gave that boundary a clear edge.

AI changed that limit by making programming accessible without the expertise it used to require. I built a small HTML productivity app for myself, a cool-as-hell browser tool with two drop zones. One took a changing list of addresses; the other took a static table mapping about ten thousand ZIP codes to a few hundred urban areas. It returned the list enriched with each record’s metro area, replacing a VLOOKUP chore I did constantly.

What I could create with generative AI now raised a different set of judgments, because the old boundary had been built around ability and ability no longer answered the whole question. I could write software with an LLM’s help, but should building it become part of my work just because I can? Even if a project fits within my capabilities, I still have to decide whether it belongs in the job I’m trying to do. And I have to be honest about whether I want to spend my attention on that kind of work, or whether I was drawn to it because it had once seemed out of reach. Those are separate decisions, each shaping how I understand the role I’m willing to inhabit. Is this mine to do? I can program with AI now and still hesitate over whether that belongs in my job. That hesitation is often enough to keep the familiar boundary in place.

When the friction shows up

Familiar work is the perfect proxy for watching AI press on all four core tenets at once. A task arrives, asking us to decide where and how to focus, whether to follow the usual routine, what evidence is enough to trust the result, and which parts of the work belong to us. The assignment may look much as it always has, and expectations largely haven’t changed either. The deadline hasn’t moved, the expected quality is the same, and the finished work still carries our name. But with pressure to bring AI into our work, the judgments that once made the task feel straightforward may now be in play together. Orchestration between several new systems means our attention shifts as parts of the work run in parallel. A settled routine, perhaps one implemented with AI no more than a month ago, gets reopened. A plausible answer asks for scrutiny, while the boundaries of our knowledge, or even the role itself, come into view.

Friction in changing how we work doesn’t show up at one shared moment, and it doesn’t show up evenly across departments or teams. Some friction is there before anyone tries to change anything, in the moment we merely consider it; some only appears once we’re attempting to work differently. The difference in timing is especially visible in an organization where people aren’t all at the same point. One individual may be weighing whether to alter a process while another has already shifted how they produce familiar work. An organization can be trying to make sense of a change through experiences that don’t line up neatly, even among people working toward the same outcome.

Grouping the four by when their friction appears helped me see things differently. Difficulty incorporating AI into our work based on standards and identity seems to travel with us, largely independent of any particular task. Standards and identity shape what we count as good enough and what we see as ours to do, before the work even shows up or any change has taken place. Frustrations tied to focus and habits arrive with the work itself, once we’re actually attempting to do it differently. Changing how we work doesn’t happen all at once, and where it’s worth pursuing, we have some say in which parts move first. Seeing that split lets me look a level deeper, and brings the water I’d stopped noticing back into view. It makes me think two different things may be happening here, though each deserves a closer look before I name it. I suspect both have a lot to do with how long the AI Assimilation Gap lasts.