The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Google’s 75% figure is real but narrowly defined: it refers to new code generated by AI and approved by engineers. That is very different from an AI system independently choosing Google’s projects, designing experiments, deploying software, or improving the next generation of AI without people.
What Google actually announced
At Google Cloud Next on April 22, 2026, Alphabet and Google CEO Sundar Pichai said: “Today, 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall.”
As an Amazon Associate I earn from qualifying purchases.
The wording contains two important limits. The statistic covers new code, not every line in Google’s existing codebase, every commit, or all engineering hours. It also counts code after an engineer has approved it. Google has not presented the number as an independently audited industry benchmark; it is a company disclosure.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Google’s announcement also describes a move toward “truly agentic workflows.” Engineers are orchestrating what the company calls fully autonomous digital task forces. Google said one particularly complex migration completed by agents and engineers took six times less time than the company’s estimate for completing it with engineers alone a year earlier. That is a report about one migration, not a productivity multiplier that applies to every team or project.
Why 75% of generated code does not mean AI is building Google
Code output is only one stage of software development. An AI model can produce a large share of implementation code inside a process where people still decide what to build, define constraints, supply context, evaluate behavior and accept the change.
Building the next generation of AI involves a much broader loop: selecting research problems, designing experiments, obtaining and preparing data, interpreting results, deciding whether an apparent improvement is reliable, and feeding validated discoveries into model development. A high share of machine-written code does not show that an AI system is running that entire loop or creating a self-accelerating cycle.
The oversight numbers behind the headline
A reported Google DeepMind internal assessment provides a more granular view of how much autonomy the systems currently have. About 88% of code changes in the reported process still required multi-turn human steering or human code authorship. Every change was reviewed by a person before submission.
| Measure | Reported result | How to interpret it |
|---|---|---|
| AI-generated new code approved by engineers | 75% in April 2026, up from 50% the previous fall | Share of new code output that passed an engineer-approval step; a Google-reported figure |
| Changes needing multi-turn human steering or human authorship | About 88% | Most reported changes remained human-steered; this is internal Google DeepMind data, not an industry rate |
| Fully delegated pull requests | 7.5% rising to 11.9% | Autonomous completion is increasing but still represents a minority of pull requests in the reported period |
| Human review before submission | Every reported change | Approval remains a control point even when an agent performs most of the implementation |
The increase in fully delegated pull requests is meaningful: agents are handling more bundled changes with little intervention. It does not overturn the larger pattern that people still steer or author most changes and review all of them before they enter the normal development process.
Generation, delegation and approval are different things
Generation
Generation is when an AI system writes code or proposes a modification. It may create a function, translate code between languages, write tests, or suggest a fix. A generated suggestion can still require substantial human direction and checking.
Delegation
Delegation means assigning a bundled task to an agent that carries it through with little or no intervention while it works. A delegated migration or refactor is a stronger autonomy claim than an autocomplete suggestion, because the system must plan and execute several related steps.
Approval
Approval is the human decision to accept the result into the normal software-development process. Google’s 75% statistic combines AI generation with this approval gate. It therefore should not be rewritten as “AI completes 75% of Google’s engineering work.”
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What Google engineers do in this workflow
Engineers remain responsible for turning a product or research goal into an actionable task, giving agents the relevant codebase context, setting boundaries, and deciding how much work to delegate. They inspect proposed changes, run or interpret tests, handle ambiguous failures, and approve or reject submissions.
For large migrations, several agents can work on related subtasks while engineers coordinate the sequence and resolve issues that cross system boundaries. This changes the engineer’s mix of work from typing every line toward planning, review, debugging, architecture and risk management. It does not remove the need for technical judgment.
Google Research has previously described this progression as an expanding set of internal software-engineering tools, from inline completion to assistance that understands a codebase and development environment. The 2026 announcement places those tools in a more agentic phase, with people coordinating multiple agents rather than treating AI as a single code-suggestion box.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Is Google’s AI improving itself now?
Not according to the reported internal assessment. The assessment found no measurable acceleration in the pace of AI research attributable to AI itself. The capability trend was described as steady rather than showing a clear takeoff.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →The Decoder’s account of the same assessment also says task success declines as internal agent assignments become more complex. That matters because routine code generation and open-ended research are different tests. An agent may perform well on a narrowly specified migration while struggling to choose a worthwhile experiment, diagnose an unexpected result or decide whether a result is genuinely new.
Those findings do not imply that agent performance is static forever. They establish only that the reported period did not show a measurable, self-driven acceleration of AI progress.
Does the figure mean programmers are being replaced?
The available figures cannot support a prediction about layoffs or a date when programmers will disappear. They show that Google is shifting more implementation work to models while retaining human direction and review.
The practical effect depends on the task. Repetitive transformations, boilerplate and well-specified migrations are easier to delegate. Requirements discovery, system design, security decisions, production incidents and research choices demand context and accountability that are not captured by a code-output percentage.
For readers evaluating similar claims from other companies, ask five questions:
- Does the percentage cover new code, all code, commits, pull requests or engineering time?
- Was the code generated without intervention, or did people steer the model through multiple turns?
- Is human review required before deployment or submission?
- Which task types were measured, and how does performance change as complexity rises?
- Is the number company-reported or independently audited?
On those tests, Google’s statement describes rapid adoption of AI-assisted and agentic coding inside a human-controlled process—not an AI system that is independently building and improving itself.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
