Sign your name
When code generation is cheap, judgment is still worth paying for.
It seems today every software engineering office entertains two exaggerated reactions toward AI. Fear holds that the tool is so capable, so fast, so relentlessly productive that no human will have meaningful work for long, or that thought-work will collapse into an unbearable cycle of slop-in, slop-out. The opposing view holds that untold profits await as large traditional teams are replaced by small AI-equipped ones. Block cut 40% of its workforce in February, and Meta cut 10% two months later, each citing AI efficiency gains. Both dread and boosterism feed each other because they share the same flawed premise, that thinking is a production task, measurable by output alone. Each places an undeserved faith in what is just a tool.
A 2023 GitHub study found developers using Copilot completed tasks 55% faster than peers working without it, but the speed of delivery does not relieve the engineer of responsibility for her output.
The skill AI cannot supply is judgment. A 2023 Stanford study found developers using AI assistants wrote less secure code while expressing greater confidence in it, and a 2024 GitClear analysis of 153 million lines found AI-assisted code was reverted within two weeks at four times the pre-Copilot rate. Speed without judgment produces faster mistakes. The engineer who reads, tests, and rejects bad output is the one whose work survives.
Judgment is one skill no model supplies. Collaboration is another, and it grows in importance as boilerplate code accelerates delivery. Google’s Project Aristotle found that how a team works together predicts its performance better than who is on it. Extra time should be reinvested into the human infrastructure of the team. Café chats, informal sync meetings and genuine care for the person at the next desk matter as much as they always did. The excuse of being too busy for people is increasingly hollow.
The same is true beyond the team. Good software requires empathy for the user and imagination for a product that genuinely improves her life. Dieter Rams said you cannot understand good design if you do not understand people. Products people love serve real needs, and identifying those needs means listening carefully and finding friction. The Humane AI Pin required a raised arm to read its laser projection, and the Apple Vision Pro sent users home with headaches. Both were built around what the technology could do, ignoring what people can comfortably endure. “It shouldn’t be this hard” is a great start to designing a product, and it depends on solidarity with the essentially human quality of fatigue. No model can replicate that.
Yet even as the fatigue of setup and integration falls away, responsibility does not. A 2025 Harvard Business Review study found 18% of workers admitted sending AI-generated content that was unhelpful, low-effort, or low quality. An engineer must be able to sign her name to any contribution without reservation. “This is rough, but Claude wrote it” is unacceptable. “I don’t know, but Cursor said…” is a nauseating cop-out. The engineer must remain the author regardless of the tools that assisted her. AI-generated work she cannot defend or improve is not worth suggesting.
Using a good tool to help is still a great way to reach for greater accomplishment. Even Michelangelo used assistants to begin his work on the Sistine Chapel ceiling, bringing thirteen helpers to transfer sketches and mix pigments before dismissing most of them as his mastery of fresco grew. The ceiling remains unmistakably his. The engineer who starts with a tool and finishes with her own judgment remains the author.
The truly hard problems remain. Engaging them consistently is what forms an engineer who can solve them.