Three-quarters of Google's new code is now written by AI

TL;DR

• Sundar Pichai, confirmed by Google and reported by Business Insider, disclosed that 75% of new code at Google is now AI-generated — up from 50% six months ago, ~0% two years ago.
• This is the first first-party, audited, operational data from the largest software-engineering organisation on Earth. It is the first number in the "can AI really do knowledge work" debate that should settle anything.
• Caveats matter: generated ≠ shipped unedited; new code ≠ all code; Google has selection advantages a typical enterprise doesn't.
• The 25-percentage-point jump in six months is the fastest productivity-tooling shift in the history of software. Faster than version control, IDEs, or cloud migration.
• Coverage has been thin because there is no launch, no demo, no CEO on stage. That is the reason to pay attention.

The data anchor

75%.

That is the share of new code written inside Google that is AI-generated, according to Sundar Pichai, confirmed by the company, reported by Business Insider this week. Six months ago the figure was 50%. Two years ago it was effectively zero.

Every argument about whether generative AI can really replace knowledge work has, until now, relied on anecdotes, surveys, or consultancy projections. This is first-party, audited, operational data from the largest software-engineering organisation on Earth. It is the first number in this debate that should settle anything.

Caveats that matter

"AI-generated" is not "AI-finalised." A Google engineer accepting a Gemini or Codex-style suggestion is still reviewing, editing, and integrating. The human is in the loop. The number measures keystrokes, not judgement.

"New code" is not "all code." Google's production stack is largely long-standing human-written code. The 75% applies to net-new lines committed this year.

Google has a selection advantage. It builds the tooling, dogfoods it aggressively, and runs internal models fine-tuned on its own codebase. The Microsoft number is lower. At a typical enterprise, much lower — Datadog's State of AI Engineering 2026, published this week, puts the average closer to 25%.

With those acknowledged: a 25-percentage-point increase in six months at Google's scale is the fastest productivity-tooling shift in the history of software. Version control, IDEs, cloud migration — none moved this fast.

The argument it settles

For three years the "AI can't really code" camp has relied on two arguments: the code is buggy, and it can't handle real systems complexity.

Both were defensible in 2023. Both are now contradicted by a company whose primary asset is its code quality.

Google does not tolerate buggy production code. It has the most rigorous internal review culture in the industry. If 75% of new code passes that review with AI involvement, "AI can't really code" is reduced to a claim about which AI, on which codebase, with how much human oversight. Those are useful questions. They are not the same as the original claim.

The argument it opens

The harder question is what this does to software engineering as a profession. Three live positions.

Position A — productivity dividend. AI makes engineers more productive; Google ships more, faster; headcount stays flat or grows because there is now more valuable work to do. The position Pichai publicly holds.

Position B — labour substitution. Productivity gains get banked as headcount reduction. Google has reduced engineering headcount modestly over eighteen months while output has grown. The correlation is real; the causation is contested.

Position C — skill bifurcation. Senior engineers get more productive; junior engineers get squeezed because the tasks they would have used to build judgement are automated. Evidence is beginning to show in graduate hiring data across the industry, though not yet conclusively at Google.

The honest answer is all three are happening at once, in different proportions, in different teams. The dishonest answer is to pick the one that fits a pre-existing ideology and declare victory.

The hype to deconstruct

The loud version of this story is "AI replaces engineers." That is not what the number says and it is not what is happening. The number says a specific kind of keystroke — boilerplate, scaffolding, well-understood patterns — is now automated. Architecture, judgement, review, and integration are not. Anyone telling a CTO they can cut engineering headcount 75% this year is selling a vendor contract, not a finding.

The quiet version of this story is the one that matters: the ladder that produced senior engineers is quietly being pulled up behind the current cohort. That is the under-covered consequence.

Why coverage has been thin

No launch event. No demo. No CEO on a stage with a flashing screen. The number surfaced in a Business Insider story built around an internal Google memo and a quote Pichai gave a fortnight ago.

The AI press covers product announcements the way sports press covers matches: the score, the highlights, the quotes. A statistic from an internal memo is the equivalent of a quarterly attendance figure — important, boring, structurally revealing. It does not travel.

If you are making decisions about where AI actually changes your work, the 75% figure is more useful than every product launch of the last quarter combined.

The corroborating signal

The same week Google disclosed 75%, three other datapoints landed:

• Datadog's State of AI Engineering 2026 — 69% of surveyed enterprises now run three or more models in production. Multi-model is the norm.
• Cloudflare — 241 billion tokens processed on its internal AI engineering stack in a single week.
• GitHub — declined to publish a comparable Copilot number this week; rumoured to be preparing an announcement for its May DevOps event.

The direction of travel is not ambiguous.

Cross-layer implications

This is not a developer-tools story only. It reaches into university CS curricula (the junior-engineer task load is disappearing), into outsourcing economics (offshore engineering teams that sold "boilerplate at scale" lose their value proposition), into hiring and visa pipelines (the H-1B case for junior engineers softens), and into M&A math (engineering-cost assumptions in SaaS acquisition diligence need repricing).

What this means for you

• Senior software engineer: your unit of work is no longer lines written. It is architecture, judgement, review, integration. Compensation and hiring are already reshaping around this. The Google number tells you it is not a prediction any more.
• Junior software engineer or student: the task load that used to build your judgement — CRUD endpoints, boilerplate, bug fixes, test scaffolding — is disappearing. Build judgement by other routes: reading production codebases, pair-programming with seniors, taking on design work earlier than previous cohorts did. The easy on-ramp is closing. Don't wait for your employer to tell you.
• CTO or head of engineering: the question is no longer should we adopt AI coding tools. It is why is our number 25% when Google's is 75%? If the answer is "our codebase is older" — fine, but that's the ceiling, not the floor. If the answer is "our engineers resist the tooling" — that's a management problem.
• CFO or board member: AI-assisted engineering is a line you need to forecast. Ask what the current percentage is, what the twelve-month target is, and what the expected headcount implication is. Any engineering leader who can't answer all three has not done the work.
• Not in software: watch which profession publishes its version of this number next. Legal, financial analysis, marketing production, customer support are the live candidates. When one lands a 75% figure from a credible first-party source, that profession's work has structurally changed too.

Uncertainty ledger

• Breakdown by team and language. Google has not said whether the 75% concentrates in ads systems, Cloud, Search, or Android — or whether infrastructure code shows the same pattern as application code.
• Acceptance rate. 75% generated is not 75% shipped unchanged. Google has not disclosed edit rates.
• Whether the number plateaus, climbs, or reverses. If the next datapoint in October is 85%, the profession is reshaping faster than any of the positions above predict. If it's still 75%, the current tooling has hit its ceiling.
• Microsoft and Meta equivalents. Both companies are believed to track internally; neither has published.

Bottom Line

75% is the number that ends the "can it really" phase of the AI-and-knowledge-work argument and opens the "what is a senior engineer in 2028" phase. It is an internal memo, not a launch, and that is precisely why it is the most useful number of the quarter. Every CTO who cannot articulate their own percentage, where it is going, and what the organisational design implication is, will be asked by their board inside the next two quarters. Start writing the answer now.

Written in the tradition of — E.

Sources

• Tier 1: Business Insider — Sundar Pichai internal memo reporting (April 2026); Google — Q1 2026 earnings call commentary
• Tier 2: Datadog — State of AI Engineering 2026 report (April 2026); Cloudflare — Agents Week 2026 infrastructure disclosures