Google CEO Sundar Pichai said on April 22, 2026 that artificial intelligence now writes 75% of new code at the company, up from 50% last fall and 25% in October 2024. The claim, made at Cloud Next 26 in Las Vegas, came with no rejection rate, no security audit, and a quieter number Pichai himself once called more important: a 10% lift in engineering velocity.
The headline lapped Microsoft, Meta and Salesforce by a wide margin. It also arrived on the same week independent researchers reported that 45% of AI-generated code still ships with known security flaws, and that developer trust in those tools sits at a multi-year low.
What Pichai Actually Said At Cloud Next 26
Pichai’s blog post landed the same morning the conference opened to roughly 32,000 attendees in Las Vegas. He framed Google as customer zero of its own coding stack and put the milestone in one line: 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% in autumn 2025.
He paired it with two operational examples. A complex internal code migration, run by a mixed team of agents and engineers, finished six times faster than a year ago with humans alone. Marketing teams used Gemini to spin out thousands of creative variations for the Chrome launch, posting a 70% faster turnaround and a 20% lift in conversions.
The same post disclosed that Google’s first-party models now serve 16 billion tokens per minute through direct API calls, up from 10 billion last quarter, and that Gemini Enterprise paid monthly active users rose 40% quarter-over-quarter in Q1.
From 25% To 75% In Eighteen Months
Pichai first volunteered the AI-coding number on Alphabet’s Q3 2024 earnings call, when it was 25%. The figure crossed 30% in spring 2025. It hit 50% by autumn. Now 75%. At that pace, the question is no longer whether AI writes code at Google. It’s whether the metric still means anything.
The Number Pichai Buried On The Lex Fridman Podcast
Asked on Lex Fridman’s podcast last year how Google measures AI’s real impact, Pichai gave a different answer. “The most important metric, and we carefully measure it, is how much has our engineering velocity increased as a company due to AI?” he said in the full transcript of episode #471. The number he attached to it was 10%.
That gap, between a glossy 75% in a keynote blog and a 10% productivity gain in a podcast interview, is the part the Las Vegas stage skipped. Lines of AI-generated code is a top-of-funnel measure. Velocity is the bottom line.
A 75% share doesn’t account for code AI wrote that engineers heavily rewrote, suggestions accepted then deleted minutes later, autocomplete tab-presses lumped in with full agent commits, or the line between scaffolding and load-bearing logic. Without a rejection rate, the headline reads closer to a marketing measure than an engineering one.
How Google’s 75% Stacks Up Against Microsoft, Meta And Salesforce
Pichai’s claim, taken at face value, doubles the share publicly disclosed by his closest peers.
| Company | AI-Generated Code Share | Source / Date |
|---|---|---|
| 75% | Pichai blog post, April 22, 2026 | |
| Microsoft | 20% to 30% | Satya Nadella, LlamaCon, April 2025 |
| Salesforce | ~30%, with a 50% internal target | Benioff and Agentforce disclosures, 2025 |
| Meta | ~50% of development targeted for 2026 | Mark Zuckerberg, Joe Rogan podcast, January 2025 |
Each company counts the metric differently. None publishes a rejection rate. Microsoft CTO Kevin Scott has said publicly that he expects 95% of all code to be AI-written within five years. The race is now less about adoption and more about what the share actually represents inside the codebase.
The Security Hole The Keynote Didn’t Mention
Veracode’s Spring 2026 GenAI Code Security Update, released this month, flunked the industry on the half of the story Pichai skipped. Across hundreds of generation tasks, only 55% of AI outputs were secure. In 45% of cases the model introduced a known vulnerability.
- 72% security failure rate in Java, the worst-performing language tested.
- 38% to 45% failure rates in Python, C# and JavaScript.
- 86% of cases failed to handle cross-site scripting (CWE-80).
- 88% of cases failed to block log injection (CWE-117).
Real-world filings track the lab data. Independent CVE counters logged 35 disclosures in March 2026 traced to AI-generated code, up from 6 in January and 15 in February. Google itself patched a remote-code-execution prompt injection in its Antigravity coding tool in February. The company also shipped CodeMender, a Gemini agent that hunts and fixes critical flaws inside its own source tree, an admission that the underlying problem is live.
Why Most Developers Don’t Trust The Tools They Use Daily
The trust gap inside engineering teams is widening, not closing. Stack Overflow’s 2025 Developer Survey, published in late December, captured the contradiction in numbers: 84% of respondents use or plan to use AI coding tools. Only 29% said they trust the output, down 11 percentage points from 2024.
Among experienced developers, just 2.6% reported high trust. 20% reported high distrust. The single biggest frustration, named by 66% of developers, was suggestions that were almost right but not quite. Debugging AI output, the survey found, often took longer than writing the code from scratch.
Pichai’s 75% sits on top of that landscape. Engineers approve the code. They don’t necessarily believe it.
The Real Pitch At Cloud Next: Managed Agency
The 75% line was the marketing wrapper. The product story underneath was Google selling enterprises a way to manage AI agents the way they manage employees.
Pichai unveiled the Gemini Enterprise Agent Platform, a rebrand and consolidation of Vertex AI, with Agent Studio for low-code building, Agent Designer for no-code triggers, Agent Runtime, Agent Identity, an Agent Registry and Agent Observability. He paired it with two new chips: TPU 8t for training, scaling to 9,600 TPUs and 2 petabytes of shared memory, and TPU 8i for inference, with 1,152 TPUs per pod and three times more on-chip SRAM than the prior generation.
Google Cloud also committed $750 million to its 120,000-member partner ecosystem to seed agentic deployments, in a release timed to the same morning Pichai posted his blog.
“The experimental phase of simple copilots is over. Tab completion, context-unaware chatbots, and ‘AI, please start this for me’ is no longer sufficient. We’re in the era of making AI and agents complete relevant work, steered by human operators,” said Richard Seroter, chief evangelist at Google Cloud, in an interview with Fast Company.
The phrase that matters is the last one. “Steered by human operators” is the soft landing for the 75% claim. Engineers aren’t writing less code. They’re approving more of it.
The Quiet Question Nobody Put On A Slide
Where do new programmers learn the craft when boilerplate is gone? The traditional path through a software career, debug a tutorial, write a parser, ship a CRUD endpoint, get reviewed, is the work agents now eat first.
Hiring data hints at the shift. Software engineer job listings rose 30% in the first quarter of 2026 even as the broader tech sector cut roughly 80,000 roles in the same window. The postings disproportionately ask for LLM fine-tuning, agent orchestration, MLOps and retrieval-augmented generation. Pure implementation work is thinning at the entry level.
Google has not announced engineering-specific layoffs tied to its coding gains. Pichai’s public line is that AI raises engineer ambition, not the axe. The pressure point is the apprenticeship that produces the next senior engineer.
Frequently Asked Questions
What does “AI-generated code” actually mean at Google?
Google counts a line as AI-generated when a Gemini-based tool proposes it and an engineer accepts the suggestion, including through tab-completion or full agent output. The figure does not separate one-character autocompletes from full feature implementations, and Google does not publish a rejection or rewrite rate. Pichai says every AI output still passes human review and the standard test pipeline before merge.
Will AI replace software engineers at Google?
Google has not announced engineering layoffs tied to its 75% milestone. Pichai’s stated position is that AI raises engineer ambition rather than reducing headcount. The sharper pressure is on entry-level work, where boilerplate and routine implementation are the tasks agents handle best, narrowing the on-ramp for new graduates.
Is AI-generated code safe to ship to production?
Independent research says not without review. Veracode’s April 2026 study found 45% of AI generations contained known security flaws, with Java failing 72% of tasks. Stack Overflow’s 2025 survey showed 46% of developers actively distrust AI accuracy. Google routes all AI code through human approval, automated tests, and Gemini-based scanners such as CodeMender, but the underlying flaw rate remains high across the industry.
At Cloud Next, Pichai told the room Google’s next milestone isn’t a higher percentage. It’s managed agency, the day agents file pull requests inside governed pipelines and humans only intervene when something breaks. By that standard the 75% number isn’t the destination. It’s the speedometer reading on a road Google is still paving.
Leave a Comment