Why You Should Treat AI Coding Agents Like Junior Developers

A developer prompted an AI tool to add a CSV export button, got a working demo, and merged it without a thorough review. The diff revealed an unfamiliar dependency, duplicate date‑formatting logic, a memory‑hungry export routine, and missing authorization checks—issues that would likely have been flagged by a senior engineer. The episode mirrors a METR experiment where 16 seasoned open‑source contributors tackled 246 tasks with and without AI assistance. Researchers anticipated a 24 % speed gain, yet participants actually took 19 % longer, even though they still believed they had been 20 % faster. Large‑scale surveys reinforce the disconnect: the 2025 DORA report (≈5,000 respondents) found over 80 % claim AI boosted productivity, yet also linked AI use to less stable releases; Stack Overflow’s 2025 survey showed 84 % planning AI adoption, but only a third trust its output, with 66 % frustrated by “almost right” answers. GitClear’s analysis of 211 million code changes from 2020‑2024 documented a drop in refactoring from 25 % to under 10 % and a rise in copy‑pasted lines from 8.3 % to 12.3 % during the AI boom, hinting at a shift toward boilerplate‑heavy contributions. Security‑focused tests echo the same pattern: Veracode’s 2025 evaluation of more than 100 models on 80 coding tasks reported injected vulnerabilities in 45 % of cases and failed to prevent XSS in 86 % of relevant instances. A USENIX Security study generated 576 000 code samples and found invented package names in 5.2 % of commercial‑model outputs and 21.7 % of open‑source models, with nearly half of those names recurring across prompts—potentially giving attackers reusable namespaces.

These findings place AI coding assistants squarely in the “junior developer” category, but with a scalability twist: an AI can produce many more merge requests than a human newcomer, amplifying both its strengths and its blind spots. The key differentiator is learning: a human junior improves through mentorship and code reviews, while an AI only adapts if its feedback is codified in persistent documentation or tooling. Consequently, organizations that lack mature CI/CD pipelines, automated testing, or rigorous review processes risk magnifying instability when they lean heavily on AI‑generated code. Teams must treat AI output as a provisional artifact, subject to the same design briefs, incremental slicing, and review gates they would apply to a new hire, and must embed safeguards—static analysis, dependency vetting, and runtime monitoring—to catch the subtle regressions AI tends to overlook.