In June I took 2,704 em dashes out of my own website, across seventy files. Two days earlier, a review of all fifty-three essays then on the site had found the same few constructions running through nearly every one of them: the snap-back contrast, the aphorism parked alone at the end of a section, the list of three, and the em dash. Any one of those is a device, the review said. Across fifty-three essays they become a fingerprint, and readers in 2026 read that fingerprint as machine-assisted.
The review was run by a model, which I will come back to. If a fingerprint like that could settle on essays with my name on them, I wanted to know what it was doing to work that nobody reviews at all.
- The popular name for this is the wrong one, and I got it wrong in print myself.
- Why the correction that used to arrive on its own has stopped arriving.
- A test for any piece of work that starts with what you could safely delete.
- How the essays here get made, including where I let the ratio slip.
The Pattern in the Feed
For a few months now I have watched people appear in my timeline who were never there before. Quiet engineers. Colleagues I have known for years who did not post twice in half a decade. Now there is a post every few days, in clean paragraphs, with a tidy lesson at the end.
What is missing is anything only that person could have written.
The easy version of this essay is a sneer, and the sneer gets it wrong. The people writing those posts are telling the truth as they experience it, and many of them genuinely are more capable than two years ago, in some narrow ways. The trouble is what that feeling is attached to.
What This Is Actually Called
Everybody reaches for Dunning-Kruger, and I have reached for it twice on this site: carefully in an essay on mastery, and less carefully in an essay on interviewing, where I repeated the version everybody repeats, that the people who know least are the most confident. The 1999 paper is more precise. Its lowest scorers sat around the 12th percentile and placed themselves near the 62nd, the most wrong about themselves of anyone in the study, yet they still rated themselves below the people who had actually done well. I have corrected that essay.
The cartoon curve with Mount Stupid on it is nowhere in the study either. Kruger and Dunning saw the statistical objection coming in 1999, and it sharpened in 2020, when Gignac and Zajenkowski made the case that the effect is mostly a statistical artefact.
A more precise instrument comes from Yale. In 2015, three researchers ran nine experiments on a simple question: does searching for an answer change how much you think you know? People searched the internet to confirm the details of everyday explanations, such as how a zipper works. Then they rated how well they could answer questions from entirely different subjects, and a comparison group did the same ratings without searching first.
The searchers rated themselves higher, on subjects they had never looked up. The effect held when time and content were matched between the groups, and it held when the search itself was unsuccessful. Matthew Fisher, Mariel Goddu and Frank Keil named it plainly: people mistake access to information for their own understanding of it.
That mechanism is kinder than the popular story. It describes a memory built to treat a reliable outside source as part of itself, one that never labels which part lives in your head. Now put a model where the search box was. It writes the answer in your voice, structures it, softens the objections and hands it back finished. The confusion gets much harder to see, because what comes back looks like something you made.
I have started calling this borrowed competence. You hold something real and genuinely useful. It belongs to somebody else, and the tell is that you cannot put it down and rebuild it.
In Your Kindness Is Their Rework I use one test for any shortcut: does this only work if someone doesn't know something? Borrowed competence fails that test from the inside. The person who doesn't know is you.
Why Nobody Corrects You Any More
The old kind of overconfidence used to break, because reality kept an appointment with you. You wrote bad code and it fell over. You presented a design, and someone senior asked the one question you could not answer. Each of those was a calibration event, and it was cheap because it came early.
Borrowed competence cancels most of those appointments, because the output is good enough to pass inspection. A fluent post reads as informed. Code that compiles and passes its tests reads as understood. Nothing fails loudly enough to show you what you are missing. What arrives instead is engagement, which measures something else: nobody in the comments will tell you your mental model is hollow, and a ranking algorithm will happily reward a post that is fluent and empty.
The deeper version has been measured. In 2025, researchers from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using generative AI at work. The more confidence people had in the AI's ability to handle a task, the less critical thinking they reported doing on it. Confidence in their own ability ran the opposite way: people who believed they could do the task themselves engaged harder with what came back.
That inverts a comfortable assumption. The person best equipped to catch a model's mistakes is the one who least needs the model, and the one who needs it most is the least likely to check. I made the same argument about the payments years from experience; now there are numbers under it.
Where It Breaks: Software
Production does not read your prose. It runs your assumptions. Writing can carry borrowed competence for years, and a system under load finds it much faster. Generating code got cheap while verifying it stayed expensive, and that gap is where review queues now pile up.
Borrowed competence is older than the models, and I have produced it myself. The first design document I ever wrote as an architect was rejected in fifteen minutes. My chief architect read it, looked up and told me it was not an architecture, it was a description of code. Every word was mine. It had the shape of a design document and none of the answers one exists to hold, and six or seven questions about fraud parameters, scale, data classification and peak load took it apart.
In a codebase today it looks like this. The service has a circuit breaker because circuit breakers are good practice, and it trips on a healthy dependency during a traffic spike because nobody sized the threshold against real throughput. The architecture decision record lists four alternatives nobody actually rejected, so when a constraint moves next quarter the decision gets regenerated rather than re-derived, and a regenerated decision carries none of the reasoning.
I review designs by asking about throughput, load and capacity before anything else, and it is the fastest borrowed-competence detector I own. A model will produce a handsome design with no numbers attached, because the shape of a design is learnable and the numbers belong to your system.
The cleanest test I know is deletion. Ask someone what they could safely remove from their own system. Understanding shows up as knowing what is load-bearing. Borrowed work has no idea, so nothing gets deleted, and the codebase grows a layer that everyone is afraid to touch. When it breaks at night, the question from You Cannot Prompt Ownership arrives on schedule, and borrowed work has an owner on paper with nobody able to answer it.
The Tells
These are the signs I notice, roughly in the order I notice them. I name them from no high ground at all, because the review found most of them in my own catalogue.
Receipts are also, on my own numbers, what readers act on. Of the LinkedIn posts I have measured, the one that sent the most people to this site was about that fifteen-minute rejection: 47 clicks from 1,910 impressions. A technical explainer I paid to boost was shown 23,248 times and sent 15.
None of these is proof, and polish is not the crime; absence is. Good writers use these tools, and plenty of people writing in a second or third language use them to fix grammar that would otherwise get them dismissed unfairly. English is not my first language either.
In Defence of the Tool
The version of this argument that turns into gatekeeping does more harm than the problem it describes. At twenty-two I was a computer science student in Lahore. A student in the same city today can ask for an explanation of consensus protocols at midnight, from something that will never make them feel slow for asking. That is a real redistribution of access, and I will not pretend otherwise for the sake of a tidier essay. The review that found my fingerprint was run by a model, and I kept using the tool afterwards.
What matters is direction. Using a model to compress the search is a gift. Using it to skip the struggle is expensive, because the struggle is where learning happens. Psychologists call the useful kind of friction a desirable difficulty: the version of you that comes out of the hard path can judge an answer, while the version that was handed one can only hold it.
Compress the search. Do not skip the struggle.
How the Seventy Percent Works Here
An essay like this owes you an account of how the essays on this site get made. Many of them start as a voice note. The story, the receipts, the argument and the position I could turn out to be wrong about all come from me, often spoken before anything is typed. That is the seventy percent, and no model can supply it, because none of it is in anyone's training data.
A model turns that material into a draft. What I ask for after that is amplification: tighten this, find the hook buried in paragraph nine, tell me what a hostile reader would say, fix the phrasing that gives away English is not my first language. The passes after that are mine. I take back the sentences that sound like a machine being helpful, and if a paragraph survives, I can defend it in a room. I also publish my usage, so you can hold me to all of this: in one thirty-day stretch this summer, 140.5 million tokens across 70 sessions.
I have not always held the line. When that June review came back, the fingerprint was thickest in my AI essays, the pieces where I had let the model carry the most. Taking out the em dashes was the cheap part of the fix. Rewriting those essays until they read like a person again was the real work.
In another essay I argued that when an agent writes thirty percent of a codebase, the other seventy percent is the knowledge that never made it into the code. This is the same seventy percent, pointed at whoever signs the work. And the rule alone is not enough, because the shapes get in through what you read as well as what you write.
The technical writers I trust were doing all of this before any of it existed. Martin Fowler, whom I count among my virtual mentors, has spent decades revising his published thinking in public. Julia Evans draws what she has just learned with the confusion still visible. Dan Luu publishes the measurements and lets them be inconvenient. Gergely Orosz names companies and prints numbers. Charity Majors tells you what broke. None of them competes with a model on fluency, because their work is anchored in things that happened to them.
The Ten Minutes
Somebody gives you ten minutes of a life they do not get back. That is the whole transaction of an essay, and I do not think it survives being automated.
So the question I try to hold before publishing is whether anything in the piece could only have come from me, and whether the reader ends up somewhere they could not have reached by asking the same question themselves. When the answer is no, the honest move is to leave it unpublished.