You have seen this Slack message. Maybe you have sent it. "yeah I, uh, asked Claude and then I edited it a lot, so like, it is mostly mine." Then a nervous laugh emoji. Then silence.
It is June 2026. Can we please stop doing this?
The autocomplete flashback
I keep thinking about 2015. When IntelliJ started suggesting full method bodies, half the office called it cheating. Real engineers type every character. Real engineers memorize the standard library. There was a guy on my team who would turn autocomplete off during pairing sessions to prove a point. I have no idea what the point was. He left for a different company and presumably types every import by hand to this day.
Same energy with linters. Same energy with debuggers, honestly. There is a generation of devs who learned on printf and still treat breakpoints as a moral failing. "I just read the code carefully." Cool, congratulations, you also took twice as long to find the bug.
Every productivity tool gets the same arc. First it is cheating, then it is normal, then the people who refused to use it quietly start using it and never mention the transition. AI is just further along that arc than most people are willing to admit.
The question that actually matters
Here is what nobody is asking in code review, and what everyone should be asking: do you understand what you committed?
That is the whole bar. Not "how much of this did you type." Not "did a model touch this file." The question is whether you can sit in a review, get asked why line 47 uses a map instead of a slice, and give a real answer. Not a vibe. A reason.
If you can do that, it does not matter where the first draft came from. If you cannot do that, it does not matter either, because you have already failed. You shipped code you do not understand. The AI is a sideshow. The actual problem is the gap between your name on the commit and your brain on the logic.
What good teams shifted to in 2026
The teams I have worked with this year, the ones that feel calm and fast, made a quiet cultural swap. The brag used to be "I wrote this in an afternoon." Now the brag is "I deleted three services and we ship the same thing." Or "I caught this in design before anyone wrote a line."
The valuable skill is not output volume. It has not been for a while, but AI made it really obvious. Anyone can produce a thousand lines of plausible Go in twenty minutes. The hard part, the part that pays, is deciding which thousand lines should exist in the first place. Taste. Judgment. Knowing when to push back on a feature request because it duplicates something three teams already built.
Models are bad at that. Models will happily build whatever you ask for, with great enthusiasm, even if the right answer was "we should not build this." That judgment call is still yours. Lean into it.
Two things to actually do
First, own your output. Fully. AI assisted or not, the moment you hit commit it is yours. Your name is on it. If it breaks production at 3am, the on-call page rings for you, not for a foundation model in San Francisco. Treat every line you ship like you wrote it from scratch, because in every way that matters, you did.
Second, be honest in PR descriptions about what was AI assisted. Not as a confession. As context for the reviewer. "The retry loop is mine. The exponential backoff math came from a model and I checked it against the Stripe docs." That is useful. It tells the reviewer where to look harder. It tells them what you verified. It is exactly the same energy as "I copied this regex from Stack Overflow and tested it against these cases." Nobody clutched their pearls about that one.
The dishonest version is pretending it was all you when half of it was not. The honest version is matter of fact about which parts came from where and how you checked them. Reviewers love the honest version. It saves them time.
The actual embarrassing thing
Using AI is not embarrassing. Using a debugger is not embarrassing. Using autocomplete is not embarrassing. The embarrassing thing is shipping code you do not understand and cannot defend.
If you can read what the model wrote, push back when it is wrong, rewrite the parts that are awkward, and explain every decision in review, then congratulations, you are doing the job. Stop apologizing in Slack. Nobody apologizes for spell check.