Google CEO Sundar Pichai drops Google's AI-code percentage and says self-improving AI is far off

Sundar Pichai. Photo: Maurizio Pesce / CC BY 2.0, via Wikimedia Commons — Wikimedia Commons By Adrian Kessler 25 October 2026, 9:42 am · 4 min read Add us on Google Google’s chief executive has stopped counting. Asked how much of the company’s code is now written by AI, Sundar Pichai said the question no longer makes sense, because AI has a hand in all of it. In the same conversation he played down the moment much of the industry is bracing for: AI that builds better versions of itself with people pushed out of the loop. Both claims may hold up. The trouble is who checks them. The figure Pichai set aside was the one number outsiders could line up against other companies. What replaces it, and the evidence behind his calmer view of self-improving AI, comes from Google measuring Google. What Sundar Pichai said about AI and Google’s code Pichai was talking to Alex Kantrowitz on the Big Technology Podcast, recorded at Google’s Mountain View headquarters and also published by Kantrowitz as a written Q&A. When Kantrowitz brought up the old claim that a quarter of Google’s code came from AI and asked for today’s number, Pichai called it “meaningless”, because every part of the code Google now touches “is influenced by AI deeply”. Recommended Reading Meituan’s LongCat 2.0 runs on chips the US can’t export and it just beat GPT-5.5 In its place he offered numbers from Gemini 4 Argon, Google’s newest model, which the company has put in the hands of its own engineers. In the four weeks since launch, he said, Google saw a 30% increase in CLs, its internal name for code submissions, and a one-third decrease in rollbacks, the changes that have to be undone after they ship. “We are tracking it with newer metrics now,” he said. The old metric did not fade on its own. In a blog post in April, Google put the AI-generated share of its new code at 75%, up from 25% in 2024 and 50% the previous autumn, according to Semafor and ETV Bharat. Semafor noted at the time that some of those claims were close to impossible to fact-check, and set them beside Snap, at 65%, and Meta, aiming for similar levels. Why Pichai says self-improving AI is still far off On recursive self-improvement, or RSI, Pichai framed his answer as a clarification of his own co-founder. Sergey Brin had called RSI a critical contributor to Argon, he said, but “we are pretty far from the end state of RSI.” Then he described where Google actually is: “The right way to describe it is, you have RSI loops along with AI researchers in various aspects of model development. I would say you almost have agents helping AI researchers, in some automated way, work together to build these models.” Brin had gone further days earlier. In a roundtable on Argon that Google shared, he said the company was “starting to get onto that RSI treadmill” and that “now our researchers and the model work hand in hand to make the next version of the model,” as Business Insider reported. A Gladstone AI report published this month said most researchers it interviewed at Anthropic , OpenAI and Google DeepMind expect RSI to begin within three to 12 months. Google’s safety report backs Pichai, on Google’s own tests On paper, the record sides with Pichai. Google DeepMind’s safety and security report for Argon finds “no significant end-to-end acceleration of AI R&D, AI-driven or otherwise” and rates that risk as “currently low”, TNW reported. AI agents, it says, “author a majority of production code and bug fixes, though code changes remain subject to human review before submission.” Fully delegated pull requests rose from 7.5% in April to 11.9% in September; about 88% still need repeated human steering or code written by a person. Those findings come from Google’s own instruments. The threshold for AI automating AI research is checked against an internal benchmark of 74 tasks that nobody outside the company can run. An analysis posted on LessWrong points to a line in the report itself: of the original tasks Google audited, “about a quarter of them contained bugs that prevent any submission from passing.” Because many runs timed out, the report gives no point estimate for Argon, which solved roughly half of the runs that finished. Google says its verdict holds even if every unfinished run had succeeded. Human review is the gate the safety case leans on. Thirty percent more code submissions means thirty percent more changes for the people holding that gate, and the metrics Pichai chose count output and reversals, not how closely each change was read. Pichai’s own fix: independent auditors Pichai did not brush the worry aside. Alignment research still has “clear gaps”, he said, and “it’s reasonable to be worried.” The next step, in his view, is cross-industry standards “including the ability to evaluate using independent evaluators”, modelled on accounting standards and outside auditors, with regulation that “will end up playing a role.” Measured against that standard, both of this week’s claims are still unaudited. Google counts its own changelists, grades its own rollbacks and scores its own research benchmark. The percentage he called meaningless was crude, but anyone could set it next to Snap’s. The numbers that replace it can only be read from inside Google. Tags: Google , Gemini , Sundar Pichai , AI coding , Sergey Brin Share This Post More Like This Dario Amodei, on the AI he builds: “a duty and an obligation to be honest about what is coming” ChatGPT adds a teen mode — parents can limit it, but not read it ChatGPT now shows you wearing any jacket or dress before you buy it SpaceX IPO filing puts Anthropic’s Colossus bill at $15 billion a year Apple Bets the $899 Mac mini M6 Can Replace Cloud AI for Most Users ChatGPT will start selling ad space next to the images you ask it to make Add us on Google