Software engineers have spent two years being told their jobs are the most automatable in tech. A report published by IEEE Spectrum this week documents what actually happened instead: the work did not disappear — it moved. AI-generated code has flooded into production codebases, and the scarce skill is no longer writing software. It is judging it.
🔍 THE BOTTOM LINE: AI has shifted the bottleneck in software from generating code to reviewing it — and the teams adapting best are not cutting junior engineers but redefining the role around exactly the skills AI lacks: specifying intent, questioning output, and owning the result.
The numbers behind the shift come from a survey of more than 1,100 developers by Sonar, an AI code verification startup, cited in the IEEE report. Respondents estimated AI contributed 42 percent of the code they added to shared codebases. Yet the same developers reported 96 percent do not fully trust AI output to work correctly. In the survey, 38 percent said reviewing AI-generated code takes more effort than reviewing a colleague’s work, and 61 percent said AI frequently produces code that looks correct but turns out to be unreliable.
What the Flood Looks Like Inside Real Teams
The scale is easiest to see at companies that went all-in early. Synthesia, an AI video platform, moved its 118 engineers onto AI coding tools in November 2025. Chief technology officer Peter Hill says pull requests were up 120 percent year-over-year as of August, with 95 percent containing AI-generated code. One recurring failure mode: AI tools write duplicate versions of functions because they lack context — Synthesia engineers found as many as 10 versions of the same function, and had to retrain the agent after cleanup. “I don’t know if we ever get to the point where you can truly trust the agentic generation of code,” Hill told IEEE Spectrum.
Other teams are attacking the problem before a line of code exists. McLaren Stanley, a senior principal engineer at Amazon Stores, told the publication his team now spends more time writing specifications — detailed plans for what the AI should build — after one missing instruction caused an agent to generate 25,000 lines in the wrong version of the programming language Swift. The fix took fifteen minutes once the spec was corrected. At Amazon Web Services, senior principal engineer David Yanacek says agents themselves now run the first round of checks — testing, plan-comparison, security scans — before a human reviewer steps in. The volume pressure is real everywhere: at Bonterra, a nonprofit software provider with about 290 engineers, CTO Tanuja Korlepra says proposed changes tripled within three months of adopting AI, code entering review rose tenfold, and review times tripled.
The money is following the bottleneck. In August, AI code review startup CodeRabbit raised US$143 million at a US$1.5 billion valuation, reporting more than two million reviews a week for 17,000 customers — a signal that “machine checking the machines” is becoming an industry of its own, with its own hiring demand.
The Junior Engineer Question — and the Companies Rethinking It
The sharpest career question in the report is whether juniors can still learn the craft when AI writes much of the code. If entry-level engineers spend less time writing software, how do they develop the judgment to review it? The report’s most quotable warning comes from Korlepra: “If the industry stops hiring juniors, the industry stops producing seniors.”
The counter-examples are as interesting as the warning. IBM told IEEE Spectrum it is using AI to give new engineers harder assignments sooner — recent graduates now work on product features once reserved for senior engineers, with AI handling implementation and testing, while juniors diagnose and fix failures before senior developers give final approval. Neel Sundaresan, IBM’s general manager of automation and AI, estimates AI helps junior engineers perform 70 to 80 percent of some tasks that once required a senior. At Making Sense, a software consultancy, juniors are deliberately kept involved in deciding why a customer needs a feature, not just checking AI output. Bonterra assigns its juniors outcomes alongside experienced colleagues: directing agents, questioning their output, staying responsible for the result.
For New Zealand readers, the pattern reads as a career-advice story more than a layoff story. The skill bundle these companies describe — writing precise specifications, reviewing critically, verifying outputs, owning outcomes — is the same one NZ employers have been demanding all winter, with SEEK data showing AI-skilled job ads more than doubling and local coverage of the country’s training gap finding only 24 percent of New Zealanders have had any AI training. A 2026 graduate who can spec a problem clearly and interrogate what an agent produced is, on this evidence, more valuable than one who can simply type code fast — a skill the agents now have.
There is a caveat worth keeping in view. This is one survey and one magazine report covering a fast-moving corner of the industry; other studies, including the Stanford research this site has covered, show AI-exposed entry-level hiring falling sharply. Both can be true: fewer junior seats in some tasks, bigger mandates for juniors at companies that restructure the role. The differentiator in the data is not age — it is whether the junior is treated as a code typist or an outcome owner.
FAQ
What percentage of code is written by AI now? Developers in a Sonar survey of 1,100+ engineers estimated AI contributed 42 percent of the code they added to shared codebases, per IEEE Spectrum.
Is AI replacing junior software engineers? The picture is mixed. Some research shows AI-exposed entry-level hiring falling, but IBM, Bonterra and Making Sense told IEEE Spectrum they are expanding junior responsibilities — specifying work, directing agents and owning outcomes — rather than cutting the roles.
What skills matter most for developers in an AI-first world? Reviewing and verifying AI output, writing precise specifications, and system judgment. In the Sonar survey, 96 percent of developers said they do not fully trust AI code to work correctly — the gap between generation and trust is where the careers are.
Related reading: LinkedIn Counts 1.3 Million New AI Jobs. Most Don’t Need a Degree, Stanford’s AI Employment Gap for Young Workers Widens to 19 Percent, and NZ Employers Want AI Skills More Than Australian Ones Do.
❓ FAQ
How is code review changing with AI? Teams are reviewing plans before AI codes, using agent-based first-pass checks, and routing risky changes to senior humans — with the review stage, not the writing stage, now the bottleneck.
🔍 THE BOTTOM LINE
The AI-reshaping-jobs story has mostly been told as subtraction: tasks removed, roles cut. The IEEE report documents the other half — addition. Companies are building whole functions around reviewing, verifying and specifying AI code, funding a new generation of tooling, and redesigning the junior engineer role around judgment rather than typing. For careers, the practical read: the entry point into software is changing, not closing. Learn to direct and interrogate the machines, because the industry still needs seniors — and seniors only come from juniors.
