AI Writes Most Code, Yet It Won’t Build Itself

Google Says AI Isn’t Building Itself Yet—Even as It Generates 75% of New Code

Google is highlighting a growing gap between AI’s expanding role in day-to-day software production and the idea that AI research is now accelerating itself in a measurable way.

According to information shared in recent coverage, Google DeepMind conducted its first internal audit focused on whether AI-driven research is speeding up AI progress. The audit found no measurable speedup in overall AI advancement attributable to AI systems doing the research work.

DeepMind is tracking the issue because faster “frontier” capability gains could outpace safety, governance, and broader societal adaptation. The concern is compounded by operational realities: higher automation often requires granting AI systems wider access to internal tools and company systems, increasing the potential impact of a misaligned model.

To quantify capability growth over time, the team developed an internal measurement called GCI v1, which combines results from 123 AI models across 92 benchmarks into a single score per model. Using that index, DeepMind observed that frontier Gemini capabilities improved at a steady pace of about 0.1 GCI points per month.

At the same time, Google’s public messaging shows a sharp rise in AI-assisted software output inside the company. Google has said:

• In October 2024, more than 25% of new code was AI-generated and reviewed by engineers.
• In February 2025, Google reiterated the 25% figure in Gemini Code Assist messaging.
• By April 2026, Google said 75% of new code was AI-generated and approved by engineers.

Google CEO Sundar Pichai also pointed to workflow changes beyond basic code completion, describing a complex code migration completed by AI agents and engineers working together six times faster than was possible a year earlier with engineers alone. Google has framed this shift as a move toward “truly agentic workflows,” with engineers increasingly supervising more autonomous tasks.

The discussion arrives amid broader industry debate about whether “recursive self-improvement”—AI accelerating AI—has started to materially change the pace of progress. Recent public comments from major AI leaders have pushed that narrative into the mainstream, even as DeepMind’s internal audit found no measurable effect yet.

Outside Google, other large technology firms have reported similar trends in AI-assisted coding. Microsoft has previously said AI wrote 20–30% of code for some projects, while its CTO Kevin Scott said he believes 95% of code could be AI-generated within five years. Separately, a document cited in coverage said Meta set internal targets for “agent-assisted” code changes in some organizations, including expectations that many engineers would write the majority of committed code using AI.

For the crypto and digital asset sector—where security, reliability, and auditability are central—these developments matter because they point to a world in which software is increasingly produced through AI-mediated workflows, even while the underlying pace of AI capability gains remains, by DeepMind’s own measurement, steady rather than self-accelerating.