AI Now Writes 75% of New Code, and Trust Fell Anyway: The Work That Remains Is Judgment

Machine-generated code at Google rose from 30% in spring 2025 to 75% by spring 2026, but code churn — code reversed or heavily rewritten within two weeks of merging — doubled from 3.3% to 7.1%, and duplicated blocks reached 73 per million lines, 81% higher than in 2023, according to the source article. The article also cites the Stack Overflow 2025 developer survey, in which favorable views of AI tools fell from above 70% in 2023 and 2024 to about 60% in 2025, while distrust in AI output accuracy jumped from 31% to 46% within a year. The article argues the remaining engineering value sits in specifications, review, outcome ownership and judgment about what not to generate, rather than in producing more code.

AI Now Writes 75% of New Code, and Trust Fell Anyway: The Work That Remains Is Judgment Here is the direct answer: AI is not replacing engineers so much as replacing the part of engineering work that was already commoditized, which is undifferentiated code. The source article puts the numbers plainly. Machine-generated code at Google passed 30% in the spring of 2025 and reached 75% by the spring of 2026, yet code churn, meaning code reversed or heavily rewritten within two weeks of merging, doubled from 3.3% to 7.1%, and duplicated blocks reached 73 per million lines, 81% higher than in 2023. More code, more rework, and less trust: that is not a replacement curve, it is a supervision curve. Trust is slipping in developer surveys too. The article cites the Stack Overflow 2025 developer survey, where favorable views of AI tools fell from above 70% in 2023 and 2024 to about 60% in 2025, while distrust in the accuracy of AI output jumped from 31% to 46% within a year. The author draws a sharp line: nobody misses writing the fortieth data-transfer-object mapper. What is being replaced is work with clear requirements and cheap mistakes, such as boilerplate interfaces and the fifth implementation of a retry routine. So where does the remaining value sit? The article's answer is everything that used to be attached to typing. Writing an endpoint was never the reason an engineer was paid; the paid work was the specification, the tradeoff, the judgment call of what not to build, and the accountability when a customer calls. Maintenance data backs this up: rising rework means the expensive part of the job has moved to after generation, into review, integration, and understanding what generated code will do when it breaks in the middle of the night. This is where AI engineering now earns its keep, not in producing more lines but in deciding which lines deserve to exist. The article closes with four habits of strong teams, none of which is writing more code. Write specifications a machine cannot misread, because the prompt is the new interface contract. Treat review as a first-class skill, because reading generated code fast and finding the buried assumption is the new writing. Own outcomes rather than output, because the model wrote it is no more a defense than it compiled. And develop taste about what not to generate, because the endpoint that should not exist is the cheapest one to maintain. For engineers worried about replacement, that list is also a training plan: practice specifications, review, and judgment, and let undifferentiated code go. Next Career Switcher Resumes: Lead With the Engineering Role, Not the Old Job → https://promptcube3.com/en/threads/10205/ All Replies （0） Want a live back-and-forth? Join the global AI chat room https://promptcube3.com/en/chat/ — login to talk. No replies yet — be the first