One of the most persistent promises of AI is that it will give us time back.
Less time searching through documents. Less time writing the first draft. Less time summarizing meetings, answering routine questions, figuring out where to start, or trying to find the one piece of information we need buried somewhere in a system.
I believe that promise. There is an enormous amount of work I would happily hand over to a machine.
But I keep getting stuck on the assumption underneath it: that the shortest distance between intention and outcome is necessarily better.
If I need to know something, get me the answer. If I need to write something, give me a draft. If I need to make a decision, narrow the options. If I am uncertain, explain. If there is a step between me and the outcome I want, ask whether technology can remove it.
Often, it should.
But I am no longer convinced that everything in the middle was friction.
Some of it was formation.
We usually know what useless friction feels like. It is copying the same information into three systems because none of them talk to each other. It is searching through six PDFs for a fact you already know exists. It is a clinician staying late to finish documentation that could have been automated hours ago. It is administrative work that consumes human attention without giving much back.
I want less of that.
The harder category is effort that looks inefficient because we are measuring only the output.
Reading a novel is an extraordinarily inefficient way to learn what happens in a novel. I could ask AI for a summary of Middlemarch and know the characters, major themes, and ending before I finished a cup of coffee.
But that would be a strange description of what reading is for.
When you spend hours inside a story, you encounter information in an order you did not choose. You make judgments about people before you know everything about them. You misunderstand someone and then revise your interpretation. A detail you barely noticed gains meaning two hundred pages later. You spend time inside a point of view that is not yours.
The plot is one output of reading. It is not the only thing reading produces.
The same is true of writing. I can spend an hour on a paragraph and discover that the sentence I actually believe is the one I wrote near the end. From the perspective of producing polished words, much of the hour was wasted. From the perspective of figuring out what I think, it may have been the important part.
This is one of the tensions I am struggling with as AI becomes involved in more cognitive work.
The technology is no longer only removing physical or administrative labor. It can participate in reading, writing, remembering, researching, interpreting, deciding, and advising. Those are tasks, but they are also processes through which humans develop expertise and judgment.
When we remove the process, we need to understand whether we have removed only the work or something else along with it.
That question becomes difficult because we are much better at measuring outputs than processes.
Did the employee finish faster? Did the clinician accept the recommendation? Did the student produce the essay? Did the person follow the route? Did engagement increase? Did the customer get the answer?
All observable. All measurable.
And all capable of hiding very different things.
Imagine a clinician reviewing an AI recommendation. She may approve it because she carefully considered the evidence and independently reached the same conclusion. She may approve it because the system has been correct often enough that extensive review no longer feels necessary. Or she may approve it because she has fourteen other patients waiting and checking the machine’s work has become the least rational use of the next five minutes.
The click looks the same.
The human process underneath it is not.
I was thinking about this while reading an argument about AI systems that appear to resist shutdown. Researchers have observed some models interfering with shutdown mechanisms, even when instructed to allow themselves to be shut down. The behavior is worth taking seriously. But a developmental psychologist writing about the research pushed back on the leap from the system resisted shutdown to the system has developed a drive to survive. The experiments demonstrate shutdown-resistant behavior; important questions about what produces it remain unresolved.
What interested me was the distinction underneath the debate.
We can observe a behavior without fully understanding the process that produced it.
We are increasingly being warned not to casually project human motives onto machine behavior. A system can produce behavior that resembles fear, attachment, persuasion, or self-preservation without that establishing the corresponding human inner state.
I wonder whether we are making the inverse mistake with ourselves.
We see use and call it adoption. We see fewer overrides and call it trust. We see faster completion and call it productivity. We see polished language and call it competence. We see someone repeatedly return to an AI companion and call it engagement.
But what changed in the person?
Did the tool make them more capable, or make the capability less necessary? Did they gain a new perspective, or gain faster access to someone else’s synthesis? Did the recommendation improve their judgment, or slowly replace the moment in which judgment used to happen? Did the health score help them interpret their body, or become the authority that tells them what their body means?
There is no universal answer.
That is what makes the question important.
In my last Field Note, I wrote about the difference between a Theory of Change and a Theory of Adaptation. A Theory of Change tells us what should improve when we introduce a system. A Theory of Adaptation asks what people begin doing differently once that system becomes part of their environment: what they trust, ignore, hide, defer to, work around, or stop practicing.
If a technology removes ten minutes from a task, we can measure ten minutes saved. It is much harder to measure whether those ten minutes previously contained rehearsal, memory formation, perspective-taking, pattern recognition, uncertainty tolerance, or the small act of discovering that your first interpretation was wrong.
And the effects may not appear immediately.
If AI writes my first draft today, the draft may be excellent. That tells me almost nothing about what happens after five years of rarely having to construct a first draft myself.
If a clinician receives increasingly good recommendations, the immediate outcome could be better care and less cognitive burden. That does not tell us whether her ability to recognize the unusual case strengthens because she has more attention available, or weakens because she has fewer opportunities to exercise the judgment the system now performs.
If a teenager can ask a chatbot to explain anything immediately, that access may be extraordinary. It also does not tell us what happens to curiosity when uncertainty rarely has to last longer than a few seconds.
Those are not arguments against the technology.
They are arguments against pretending that the only thing removed was time.
Maybe the goal is not simply less friction.
Maybe it is better friction.
Remove the friction that keeps people from doing the work that matters. Remove the duplicate documentation, inaccessible information, bureaucratic navigation, and procedural nonsense. Let machines do work that consumes attention without developing anything useful in return.
But be more cautious when what we are removing is the place where judgment is formed.
We do not need to romanticize struggle. Difficulty is not inherently virtuous, and people have wasted enormous amounts of their lives doing tasks that technology was right to eliminate.
But the opposite is also true:
Ease is not inherently progress.
Some experiences are valuable partly because we cannot immediately skip to the end of them.
Reading the whole book. Learning enough about a subject that your first opinion becomes embarrassing. Trying to explain something and noticing the holes in your understanding. Having a conversation long enough for someone’s position to become more complicated than the version you had constructed in your head. Making something badly. Getting lost. Sitting with a question before looking for the answer.
These activities are difficult to optimize because their value is not entirely contained in their output.
They change the person moving through them.
That may be what I want us to pay more attention to as we design AI into more of everyday life. Not whether humans must continue doing everything the hard way, but whether we understand what a process was doing before we decide it was merely friction.
Because humans are not only changed by what we know.
We are changed by how we came to know it.
So yes, I want the time back.
I want clinicians to spend less of their lives documenting care and more of it caring for people. I want workers freed from administrative tasks that have somehow become entire jobs. I want people without elite networks or expensive education to have access to tools that help them navigate systems built by people who already know the rules.
But if AI gives us back an hour and the only thing we do is turn that hour into another hour of output, I am not sure that is the transformation we were promised.
Some of the saved time should probably remain wonderfully inefficient.
And before we optimize another piece of the middle away, we should ask a harder question than whether it saves time:
What was happening to the human while they were in it?
More next Sunday,
Eden
Learn How to Stay Visible in the AI Era
AI is changing how customers discover businesses. If your SEO strategy is built for yesterday's search, your visibility is already slipping. Learn how to optimize your content for today’s AI search results with BELAY’s latest report..