- Google DeepMind's Haydn Belfield says AI writes over 75 percent of its production code, while a human reviews every change before submission.
- The share of changes AI now handles entirely by itself is 11.9 percent, up from 7.5. DeepMind reports no unreviewed changes, so authorship and autonomy are different measures.
- Leaders should ask how much AI-written work ships without review in their own organization, and whether they have enough reviewers.
Most executives ask how much of their code AI now writes. That is the easy question. The harder one is how much of that code ships with no human looking at it. DeepMind's numbers show why the two are different, and why only the second one measures accountability.
What DeepMind reports
Google engineer Haydn Belfield says DeepMind has seen no measurable speed-up in AI research driven by AI itself. The Decoder reported his findings on October 14. That does not mean automation is small. Belfield says significant automation is already in place.
On authorship, the figure is large: AI now produces more than three-quarters of DeepMind's production code. On sign-off, Belfield says a human reviews every change before it is submitted. DeepMind therefore reports no unreviewed changes.
A second figure is easy to misread. Belfield says the portion of changes that AI now carries out wholly on its own climbed from 7.5 percent to 11.9 percent. The Decoder does not spell out what that category includes. It cannot mean unreviewed, because every change is reviewed. It is best read as a measure of how far AI work has moved toward independence.
Where AI still needs people
The Decoder does not say why DeepMind reviews every change. It does show where AI struggles on its own. Across 1,389 internal runs of Argon, DeepMind's agent, success without human input fell sharply as tasks got more complex.
A survey of twelve senior DeepMind researchers pointed the same way. Automation was highest for clearly defined work that is easy to check. It was lowest for open-ended scientific ideas. Belfield lists creative research, designing new experiments, diagnosing subtle training problems and long-term planning as work where people remain essential.
DeepMind also tests agents on RE-Bench, its own set of 74 realistic machine-learning engineering tasks. Argon solves about half. Gemini 3.7 Flash solved 27 percent. The gain is large, yet the source calls it far from the near-perfect results full automation would require.
How the team looks for a speed-up
To watch for acceleration, the team built a General Capability Index. It spans 123 models and 92 benchmarks. Gemini models gain roughly 0.1 index points a month, and the pace has not changed in any significant way.
Gemini 4 Argon sits a little above that line. The team credits extra compute and organizational investment, not AI speeding up its own development. For now, the team puts the danger from AI automating its own research at a low level. It is also building tools to spot early signs of acceleration.
Writing is not the same as deciding
The lesson here is that AI authorship and AI autonomy are separate measures. A system can produce most of the output while a person still signs off on each item. That human step is what turns fast output into accountable output.
Aviation offers a familiar parallel. Autopilot handles the routine stretches, and pilots stay responsible for the hard parts. The question is not only how much the machine does. It is who answers for what leaves the building.
That makes review capacity worth watching. If AI-written output keeps growing and the pool of qualified reviewers does not, the review step is where pressure builds. The sources do not say this is happening at DeepMind. It is a question the figures raise for any organization with a similar setup.
Not everyone agrees
The Decoder notes a dissenting view. A research paper from more than 20 AI researchers, among them Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, warns of a possible "intelligence explosion".
The paper's argument is that machines might take over every part of building AI within a few years. That would squeeze what normally takes years of progress into months.
Two caveats apply. This is Google measuring its own lab, and the low-risk rating is the team's own judgment. We also have only The Decoder's account of Belfield's findings, not the underlying data.
Questions to put to your team
Ask what share of your code changes ship with no human review, and whether anyone tracks that figure. DeepMind reports none. Few organizations can say the same without checking. Ask how many reviewers you have, and whether that number has kept pace with AI-written output.
Ask which tasks are well defined and easy to verify. In DeepMind's survey, automation ran highest there. Set a baseline now, so any later change in pace is measured rather than felt.
AI can write most of the code. Accountability for it still sits with people, and that is worth measuring.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Decoder.
