[DigitalToday reporter Yoonseo Lee] Google DeepMind said it has not yet confirmed a phenomenon in which artificial intelligence creates more advanced AI on its own.

On Oct. 15, blockchain media outlet Decrypt reported that Google DeepMind said in the Gemini 4 Argon safety report it found no evidence that the pace of its internal AI research and development has meaningfully accelerated.

More than 75 percent of DeepMind's production code is now written by AI. Even so, DeepMind said there has been no statistically significant figure in the pace of AI performance progress since April 1, 2026. It said its internal capability index rose by 0.1 point a month.

The assessment runs counter to the "recursive self-improvement scenario" that the industry has recently worried about. It refers to a situation in which AI makes a smarter AI, which then makes an even more advanced model, outpacing the speed of control, verification and regulation. The report defined such a phase as one in which existing safeguards are severely insufficient.

Hayden Belfield (헤이든 벨필드), a Google DeepMind researcher, said there is no meaningful end-to-end acceleration across machine learning (ML) R&D, regardless of whether it is AI-led. He said he views the risks of AI-led ML R&D as currently low.

DeepMind used an internal metric, GCI v1, combining 123 AI models and 92 benchmarks to check this. Under the metric, performance gains for the Gemini family of models were close to linear. In late 2024, Gemini 1.5 Pro was around 0.5 points. In September 2026, Gemini 3.8 Flash exceeded 2.9 points. Gemini 4 Argon was about 3.3 points and sat above the trend line, but DeepMind said it viewed this not as a result of AI raising development speed on its own but as the impact of increased computing resources and greater organisation-wide investment.

Automation also remained far from full autonomy. At DeepMind, the share of pull requests (PRs) rose to 11.9 percent in September from 7.5 percent in April. But about 88 percent still required separate instructions, and humans had to write code directly.

Practical task-handling capabilities also appeared limited. The model was able to solve on its own only about half of internal research engineering tasks. That is far below the 90 percent level that DeepMind views as a warning signal for automation risk.

There was also a divide between areas where automation was relatively easy and areas where it was difficult. A survey of 12 senior executives and researchers at DeepMind found the level of automation was highest for "clearly defined and verifiable execution tasks." By contrast, it was lowest for "open-ended scientific idea discovery, new experiment design, diagnosing learning instability and setting long-term research directions." That means AI is strong at carrying out instructions, but its ability to decide which ideas to pursue is limited.

The results also contrast with strong warnings from some in the industry. Anthropic Chief Executive Dario Amodei has argued that recursive self-improvement is already accelerating, and OpenAI CEO Sam Altman and Elon Musk have agreed. Days ago, it also mentioned that Jacob Kokxon, a former Anthropic engineer, left the industry after claiming AI could kill everyone within 10 years.

DeepMind did not fully rule out the possibility of risk. The figures are limited to data inside its own lab. The same report specified it cannot rule out the possibility that an unrestricted Gemini 4 Argon could exceed risk thresholds in areas related to cyberattacks or biological weapons information. As a result, Google's assessment is meaningful as material showing internal automation levels at the current point in time, but it is hard to see it as a conclusion that ends the safety controversy surrounding high-performance models.