Key facts
- AI writes over 75% of Google DeepMind's production code.
- AI progress has not shown statistically significant acceleration.
- Gemini 4 Argon completed about half of internal research engineering tasks.
- 88% of code changes still require human steering or authorship.
Google DeepMind researchers have concluded that artificial intelligence is not currently accelerating its own development, despite AI systems writing the majority of the lab's production code. The findings, detailed in a safety report for the Gemini 4 Argon model, counter concerns about recursive self-improvement, where an AI creates progressively smarter versions of itself, potentially leading to uncontrollable progress.
The lab's internal capability index, which tracks AI model progress, has shown a consistent increase of approximately 0.1 points per month. Researchers attribute this steady progress to increased compute power and organizational investment, rather than AI-driven acceleration. While AI writes over 75% of DeepMind's code, about 88% of code changes still require human oversight or direct authorship.
Gemini 4 Argon, the most advanced model assessed, completed roughly half of the internal research engineering tasks it was assigned. A survey of senior researchers indicated that AI is more effective at executing well-defined tasks than at open-ended scientific ideation or experimental design. The report acknowledges that an unrestricted version of Argon could still pose risks related to cyberattacks and bioweapons information.