Key Takeaways
- Google reported in April 2026 that 75% of its new production code is AI-generated and engineer-approved, up from 50% in autumn 2025.
- Veracode’s 2025 GenAI Code Security Report found AI-generated code passed security checks only 55% of the time, making manual security review the critical remaining skill.
- Architectural judgment and low-level system knowledge are now the bottleneck, not code volume, the developer’s core workload has shifted from writing to reviewing and rearchitecting AI output. Google’s April 2026 disclosure that AI now generates 75% of its new production code would suggest human coding skills are becoming optional. Veracode’s 2025 GenAI Code Security Report cuts against that: AI-generated code passes security checks only 55% of the time, and the gap between what the model produces and what ships safely still requires an engineer who knows what to look for.
From Writing Code to Catching Its Mistakes
The history of programming is a story of increasing abstraction. Early programmers worked in machine code and assembly, wrestling with hardware at the lowest level. High-level languages, starting with FORTRAN in the 1950s and followed by COBOL, C and Python, shifted focus from hardware specifics to logic. IDEs added auto-completion, syntax highlighting and integrated debuggers. Each layer raised productivity and expanded what developers could build, but none removed the need to understand what sits underneath.
AI coding assistants are the next step in that progression. Tools like GitHub Copilot and Google Gemini’s coding features interpret natural language prompts and generate semantically meaningful code blocks, going beyond syntax assistance to attempt a read of developer intent. A January 2026 survey published by Sonar found that developers using AI tools daily report roughly 42% of their contributed code is AI-generated or heavily assisted. A separate January 2026 report from Netcorp put AI-generated code at 41% of all output, with developers saving between 30% and 60% of their time on coding, test generation and documentation. The speed advantage is real, particularly for well-defined tasks like CRUD endpoints, standard UI components and data transformations.
The debugging picture is less flattering. A January 2026 survey found that 95% of developers spend at least some effort reviewing, testing and correcting AI output, with 59% rating that effort as moderate or substantial. More telling: 38% of developers report that reviewing AI-generated code takes more effort than reviewing code written by human colleagues. AI-generated code is often syntactically correct while containing outdated API patterns, subtle logic errors or edge-case failures that are not immediately visible. Without a solid grasp of the underlying system, those convincingly wrong suggestions can sit undetected through review and surface later as hard-to-trace bugs.
The problem is structural, not incidental. AI layers tend to treat system telemetry as text to be searched rather than as semantically connected signals, missing the relationships that an engineer with deep system knowledge would recognise immediately. In distributed, cloud-native environments where context is fragmented across services, that gap widens, a challenge well-documented in teams working at scale on AI-assisted development.
Where Architectural Judgment Fills the Gap
Accelerated code generation has raised the value of the person who can verify what the model produced, not replace it. AI-generated code frequently introduces security vulnerabilities, including common OWASP Top 10 issues, even in controlled testing. Code duplication has increased while legacy refactoring has slowed: AI generates new code faster than teams can consolidate or clean up what already exists, pushing the developer’s workload from creation toward review and remediation.
Manual coding fluency has not become less relevant; its application has shifted. Developers who understand algorithms and system architecture at a low level are better positioned to catch the 45% of AI-generated code that fails security review, identify duplication that degrades maintainability and make architectural calls that no current model handles well. The pay premium now attached to those human skills reflects exactly this dynamic: the demand is not for people who write more code, but for people who can judge it.
Originally published at https://autonainews.com/ai-writes-75-of-google-code-but-only-55-passes-security-checks/
Top comments (0)