Key Takeaways

  • Google CEO Sundar Pichai disclosed AI is writing approximately 30% of all new code at Google
  • This disclosure has been followed by similar disclosures from Microsoft and Meta indicating comparable figures
  • The AI-written code is reviewed by human engineers before deployment — maintaining human oversight
  • SWE-bench scores for top AI models climbed from 33% to nearly 81% of real GitHub issues in 18 months
  • ASEAN developer community is experiencing similar AI-assisted productivity improvements in Singapore and Vietnam

The Facts

Google CEO Sundar Pichai's disclosure that AI now writes approximately 30% of all new code at Google has sent ripples through the technology industry. While the specific percentage varies by team and project type, the directional signal is consistent with similar disclosures from Microsoft (citing comparable AI-assisted code proportions in internal development) and Meta (referencing AI's role in feature development velocity).

The specific qualifier in all these disclosures is important: AI-written code is reviewed by human engineers before deployment. The human review step means this is augmentation rather than replacement — AI handles the implementation while humans retain architecture, design, and review responsibilities. The productivity implications are still substantial: if AI handles 30% of implementation, developers can focus more time on design, architecture, and the complex problem-solving that AI cannot reliably perform.

SWE-bench — the benchmark measuring AI ability to resolve real GitHub issues — rose from 33% in August 2024 to nearly 81% by December 2025. This trajectory suggests AI coding capability is improving faster than most developers anticipated, and the 30% of code figure may represent a temporary plateau before further capability improvements push it higher.

Technical Deep-Dive

The quality assurance challenge for AI-written code is the critical gating factor on how far AI code generation can scale. The SWE-bench improvement to 81% means AI can resolve approximately 4 in 5 GitHub issues when tested in isolation. In production development, however, code must meet quality standards that SWE-bench doesn't fully capture: it must be maintainable (readable by other engineers), consistent with existing code style, secure (no SQL injection, no hardcoded credentials), and compatible with the broader system architecture.

Human code review serves all these functions simultaneously. An experienced engineer reviewing AI-generated code can identify architectural misalignments, security vulnerabilities, and style inconsistencies in minutes — tasks that require contextual understanding of the codebase, the team's standards, and the business domain that AI systems do not yet reliably possess.

The emerging practice of "AI pair programming" — where a developer describes intent and the AI generates implementation, with the developer reviewing, revising, and directing — is the most productive human-AI collaboration pattern for current AI capability levels.

The ASEAN Perspective

Singapore's developer community has broadly adopted AI coding assistants, with the government's push for digital transformation creating demand for software development that outpaces local human talent supply. In this context, AI's ability to handle 30%+ of implementation work is not an academic productivity metric — it is a genuine enabler of project delivery at a scale that would otherwise require significantly more developers.

Vietnam's rapidly growing developer community is an interesting case study in AI coding adoption in an emerging market context. Vietnamese developers working on international projects through outsourcing relationships have adopted AI coding tools at high rates, with the productivity improvements partially offsetting wage cost disadvantages relative to ASEAN competitors.

For ASEAN businesses commissioning software development — whether through internal teams or external agencies — the AI coding productivity improvement should be reflected in project estimates. A development agency that hasn't adopted AI coding tools will be priced at a disadvantage relative to one that has.

RECATOOLS Verdict

The 30% figure is likely to be remembered as a milestone rather than an endpoint. The trajectory of AI coding capability improvement — from 33% to 81% on SWE-bench in 18 months — suggests the proportion of AI-written code in enterprise development will continue to increase.

The implication for developers is not technological unemployment but role transformation: from implementation-heavy work toward architecture, design, review, and the complex judgment that AI cannot yet reliably provide. Developers who invest in design and architecture skills, and in the judgment required to effectively review AI-generated code, will be more valuable — not less — as AI coding capabilities improve.

Frequently Asked Questions