The more time I spend working with AI output, the more I notice the patina. It’s hard to describe. There’s no single tell, no obvious watermark. The grammar is fine. The structure is fine. The vocabulary is fine. But something is off. It’s almost accurate without being true. It overembellishes things that don’t need emphasis and glosses over the parts that actually matter. It sounds like a person who read about the thing but never did the thing.
I’ve been working on my resume with AI assistance, and this is where it becomes impossible to ignore.
The resume coach
A good resume-writing coach does something specific. You sit down, you talk about what you did, and they listen. Then they take your words and find better ways to phrase what you actually said. They don’t add things. They don’t invent accomplishments. They don’t assume that because you worked on a database migration, you must have also improved query performance by forty percent. They reshape what’s real into something that reads well on paper.
AI doesn’t work like that. You give it the same conversation, the same raw material, and it comes back with something that’s close to what you said but isn’t what you said. It fills in gaps with plausible assumptions. It inflates scope. It adds metrics that sound right but aren’t real. It uses phrases that no human being would say about their own work. And it does all of this confidently, fluently, in perfectly formatted bullet points.
So you go back and correct it. You say “no, I didn’t lead a cross-functional initiative, I just talked to the other team.” You say “remove the thirty percent improvement, I don’t have that number.” You say “that’s not what happened.” And it adjusts, apologizes, produces a new version. Better. Closer. Still not right. Still carrying some residue of assumptions you didn’t make and emphasis you wouldn’t have chosen.
You go back again. And again. Each round gets closer but never quite lands. The output keeps having this slight drift from reality, like a translation that’s technically correct but misses the tone. Eventually you get tired. Eventually you say: “Okay. This is close enough.”
The surrender
That moment — “close enough” — is what scares me.
Not because the resume is wrong. It’s not wrong, exactly. It’s approximately right. If you squint, it’s fine. Nobody reading it would catch the inflation because the inflation is subtle and plausible. It’s the kind of thing you’d get away with. And that’s the problem. Getting away with it becomes the standard.
I see this same surrender happening everywhere AI generates content. A marketing email that doesn’t quite sound like the brand but close enough, ship it. A product description that overstates a feature but close enough, it’s live. A code comment that describes what the function sort of does but not exactly what it does, close enough, merge it. Documentation that explains the system in a way that’s mostly right but misses a subtle constraint, close enough, publish it.
Each instance is small. Each one is defensible. Each one is a tiny relaxation of the standard for what “right” means. And they compound.
Craftsmanship
There’s a word for caring about whether something is right even when nobody would notice if it’s wrong: craftsmanship. It’s the cabinetmaker who finishes the back of the drawer even though it faces the wall. It’s the engineer who names variables well even in a script nobody else will read. It’s the writer who rewrites a sentence four times because the rhythm is off, not the meaning.
Craftsmanship has always been inefficient. That’s the whole point. It’s the choice to spend time on things that don’t strictly need to be spent on, because the act of caring about the details is part of what makes the work good. Pageantry. Decorum. Ritual. The stuff that gets stripped out when you optimize purely for outcomes.
AI is the ultimate optimizer. It produces output that achieves the stated goal in the fewest possible steps. Need a resume? Here are bullet points. Need an email? Here’s professional language. Need documentation? Here’s a clear explanation. The outcome is met. The craft is absent.
And when you optimize communication the same way, you lose something real. The hyper-efficient version of a sentence that conveys the information without the texture. The perfectly structured paragraph that says nothing the reader didn’t already expect. The answer that’s technically responsive without being actually helpful. Few words do trick.
The interpretation layer
Here’s where it gets weird for me, and I’m aware this is the part where my brain works differently than most people’s.
I tend to think about the world as if everyone has a state object. A JSON blob with all their parameters as scalar values. Beliefs, preferences, knowledge, context — all reducible to discrete, queryable fields. I know that’s not actually how it works. People are messy and contradictory and contextual. But my default mental model is the clean one, the structured one, and I have to actively remind myself that the territory isn’t the map.
AI has the same problem. It builds a model of what something should look like based on patterns in its training data, and then it generates output that matches the model. The model is close. It’s often very close. But it’s an interpretation, not a reproduction. It’s the AI’s understanding of how resumes work, not my resume. It’s the AI’s understanding of how technical writing reads, not my technical writing. The gap between the model and reality is small enough to miss and large enough to matter.
What unsettles me is that humans do this too. We operate on narratives and interpretations and approximations of truth. We tell ourselves stories about what happened and why, and those stories are close enough to reality that we don’t usually notice the gap. We round off the edges, simplify the causes, flatten the complexity into something our brains can hold.
The AI is doing the same thing back to us. It takes our messy, complicated reality and produces a clean version that’s close enough to pass. And if our version was already an approximation, and the AI’s version is an approximation of our approximation — how many layers deep before “close enough” stops being close at all?
The blurry line
I think what bothers me isn’t that AI gets things wrong. It’s that it gets things wrong in a way that’s hard to distinguish from the way humans get things wrong. We’re both working with interpretations. We’re both filling in gaps with plausible assumptions. We’re both smoothing over uncertainty with confident language.
The difference is supposed to be that I know what’s real. I was there. I did the work. I know which parts of my resume are precisely true and which parts are rounded up. The AI doesn’t know that because it can’t know that. It’s pattern-matching against what resumes look like, not against what I actually did.
But when the AI produces something and I accept it because I’m tired of correcting it — when “close enough” wins — I’m giving up the one advantage I had. The knowledge of what’s real. If I let the AI’s version of my experience replace my version, then we’re both just working with interpretations, and neither one of us is tethered to the ground truth.
That’s the patina. It’s not that AI output is detectably wrong. It’s that it’s subtly untrue in ways that accumulate until nobody in the chain — not the writer, not the editor, not the reader — is working with actual facts anymore. Just approximations of approximations, close enough all the way down.
I don’t have a fix for this. Maybe I need a better skill for resume writing, one that constrains the AI to reshape my actual words instead of generating new ones. Maybe the tooling gets good enough that the patina fades. Maybe I’m overthinking it and everyone else’s resume is already approximate and I’m the one being unreasonable.
But I don’t think the answer is to stop noticing. The moment you stop noticing the gap between “close enough” and “right” is the moment you start producing slop and calling it acceptable. And I’d rather fight with the AI through ten more revision cycles than publish something that isn’t true just because it looks like it could be.