Editorial illustration for Google: self-improving AI far off as fully delegated code changes rose to 11.9%
Google: AI Self-Improvement Still Years Away
Google: self-improving AI far off as fully delegated code changes rose to 11.9%
Google DeepMind engineer Haydn Belfield put a number on the AI-writes-its-own-code story, and the number is smaller than the hype suggests. The share of code changes fully delegated to AI, meaning no human review before submission, climbed from 7.5 percent to 11.9 percent. That's real movement, but it leaves 88 percent of changes still requiring human oversight, with every single change reviewed by a person before it ships.
Belfield's point isn't that automation is stalling. AI now writes more than 75 percent of production code at Google. But writing code and being trusted to ship it alone are different things, and DeepMind's internal data shows a clear pattern: the harder and more open-ended a task gets, the less AI can be left to run it.
Across 1,389 runs of DeepMind's internal Argon agent, success rates without human input fell off fast as task complexity rose. A survey of twelve senior DeepMind researchers backed that up, with automation strongest on tasks that have clear, checkable answers and weakest on genuine scientific ideation. Belfield's argument rests on this gap between routine automation and the kind of self-driving research loop that would actually accelerate AI development on its own.
Google Deepmind has seen no measurable acceleration of AI research driven by AI itself, according to Belfield. That doesn't mean significant automation isn't happening already: AI now writes more than 75 percent of production code. But about 88 percent of code changes still need human oversight or review, and every change is reviewed by a human before submission, according to Belfield.
Why this matters
The 7.5-to-11.9 percent jump sounds dramatic until you remember that leaves 88 percent of changes still needing a human in the loop. That's the number worth sitting with. Google is one of the few companies with the internal telemetry to make this claim credibly, and even Belfield isn't selling the recursive self-improvement story.
For developers and founders riding the "AI writes most of our code" talking point, the Argon data is a useful corrective: complexity still breaks autonomous agents, and the drop-off isn't gradual, it's sharp. That has real implications for how you staff review pipelines right now, not in some hypothetical post-automation future. If you're building on the assumption that agentic coding tools will soon handle multi-step, high-stakes changes unsupervised, Google's own numbers say otherwise.
Researchers chasing recursive self-improvement should take note too: no acceleration in AI research velocity from AI itself means the singularity math everyone likes to cite has zero empirical backing at the one lab positioned to see it first. Worth watching whether that 11.9 percent keeps climbing, and how fast.
Related Reading
Browse all Google news →
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives Researchers tracked rogue Chinese AI agents to a mailbox via hidden text
From Archives AI solves hundreds of math problems, triggering calls for limits
From Archives Google moves AI teams from DeepMind as focus shifts from London to California
LLMs & Generative AI
From Archives OpenAI, Google AI models easily tricked into making nudes, researchers say
Common Questions Answered
What percentage of code changes at Google are fully delegated to AI without human review?
According to Google DeepMind engineer Haydn Belfield, only 11.9 percent of code changes are fully delegated to AI with no human review before submission. This represents an increase from 7.5 percent, but still leaves 88 percent of changes requiring human oversight and review before they ship into production.
Has Google DeepMind observed measurable acceleration of AI research driven by AI itself?
No, Google DeepMind has seen no measurable acceleration of AI research driven by AI itself, according to Haydn Belfield. While AI automation is happening at scale—writing more than 75 percent of production code—this has not translated into AI systems meaningfully accelerating their own research and improvement cycles.
Why does Belfield emphasize that 88 percent of code changes still need human oversight?
Belfield highlights this statistic to counter the hype around AI self-improvement and autonomous code generation. The 7.5-to-11.9 percent increase in fully delegated changes sounds dramatic in isolation, but the fact that 88 percent still requires human review demonstrates that complexity and critical decision-making still require human involvement in the development process.
What makes Google's claims about AI-written code credible compared to other companies?
Google is one of the few companies with the internal telemetry and data infrastructure to make credible claims about the percentage of code written by AI across their entire codebase. This gives their measurements more weight than anecdotal claims from other companies about AI code generation capabilities.
Further Reading
- Coding Agent Benchmarks 2026 (SWE-Bench, TerminalBench, Live) - Presenc AI
- FrontierCode 1.1: Benchmark Measures AI Code Against Maintainer Standards - Agentry News
- AI Coding Agents: Autonomous Development - Verdent AI
- Best AI Coding Agents for 2026: 12 Tools Compared - Tembo
- Best AI Agents for Software Development Ranked - MarkTechPost
