Google CEO Sundar Pichai Says Measuring AI-Written Code No Longer Makes Sense

Artificial intelligence is changing the way software is built, and even the world’s largest technology companies are rethinking how they measure its impact. At Google, AI has become so deeply involved in software development that counting the percentage of code written by machines may no longer tell the whole story.

Google CEO Sundar Pichai recently explained that the company needs new ways to measure the effectiveness of AI coding tools. Rather than focusing on how much code AI generates, Google is paying greater attention to productivity, software quality, and the ability of engineers to tackle more complex projects.

KanxeY Tales

His comments offer a glimpse into a future in which AI is no longer simply a tool that suggests lines of code. Instead, it is becoming an active partner in the software development process.

Why Measuring AI-Generated Code Is Becoming Less Meaningful

For years, technology companies have used the percentage of AI-generated code as a convenient way to demonstrate how quickly artificial intelligence is advancing.

Google has previously shared figures showing a substantial increase in AI-assisted coding. In April 2026, Pichai reported that AI was generating approximately 75% of the company’s new code, compared with 50% the previous autumn.

However, during a recent conversation with Alex Kantrowitz on the Big Technology Podcast, Pichai suggested that such percentages are becoming less useful.

His reasoning is straightforward: AI now influences almost every part of Google’s coding process. Engineers may use AI to generate code, improve existing software, identify problems, or make changes across large and complicated systems.

In such an environment, separating code written by humans from code written by AI becomes increasingly difficult. A single piece of software may involve contributions from both throughout its development.

Instead of asking how much code AI produces, Google wants to understand what its engineers can accomplish with AI’s assistance.

This represents a broader shift in how businesses evaluate artificial intelligence. The important question is no longer simply how much work a machine performs, but whether the technology helps people deliver better results.

Google’s AI Coding Tools Are Helping Engineers Do More

One example highlighted by Pichai is Google’s internal AI coding tool, Gemini 4 Argon, which the company has been using to support its engineers.

According to Pichai, the first four weeks following its launch brought two notable changes in Google’s software development activity.

The company recorded a 30% increase in code submissions, while software rollbacks declined by approximately one-third.

Code submissions, known within Google as changelists, represent proposed changes to a software project’s code. Rollbacks occur when previously introduced changes must be reversed, often because of bugs, unexpected behaviour, or other problems.

Both measurements matter.

An increase in code submissions can indicate that engineers are completing more work. Meanwhile, fewer rollbacks may suggest that fewer changes need to be undone after implementation.

Taken together, these figures offer a more informative picture of productivity than the volume of AI-generated code alone.

However, the results should be interpreted carefully. The figures cover a relatively short period following the tool’s introduction, and they do not independently establish how much of the improvement was caused by AI.

Even so, they illustrate the type of outcome Google increasingly wants to measure: not merely the amount of code produced, but how effectively engineers can develop and maintain software.

From Writing Code to Solving Bigger Problems

The influence of AI on software engineering extends beyond automatically completing short sections of code.

Modern AI coding agents can assist with debugging, explain unfamiliar codebases, suggest architectural improvements, and help engineers manage changes involving thousands or even millions of lines of code.

These capabilities can allow developers to spend less time on repetitive tasks and more time thinking about the larger challenges behind a software project.

For example, an engineer working on a major system migration might use AI to analyse existing code, identify dependencies, propose modifications, and assist with testing. The engineer can then review the results, correct mistakes, and decide how the changes should be implemented.

This does not eliminate the need for technical expertise. Instead, it changes where that expertise is most valuable.

Developers still need to understand system architecture, security, performance, reliability, and the needs of the people using their software. They must also determine whether AI-generated suggestions are correct.

As these tools improve, the role of a software engineer may increasingly involve directing AI systems, reviewing their work, and solving problems that require broader technical judgement.

In this sense, AI is changing not only how code is written but also how software projects are approached.

Why Human Review Still Matters

Although AI can accelerate development, generating code quickly does not automatically make that code reliable.

AI systems can produce incorrect logic, introduce security vulnerabilities, misunderstand requirements, or suggest changes that appear reasonable but fail in real-world conditions.

For this reason, human oversight remains an important part of Google’s approach.

The company has reported that AI agents are increasingly involved in writing production code and fixing bugs, but their work remains subject to engineering review.

This distinction is important. An AI system may generate a technically plausible solution, yet a qualified engineer must still determine whether it is appropriate for the application, compatible with existing systems, and safe to deploy.

A productive development process must therefore balance speed with quality.

If AI allows engineers to submit more changes but creates additional errors, the apparent productivity gain could disappear through debugging and rework. Conversely, if developers can complete more work while reducing unnecessary reversals, the benefits become more meaningful.

That is why Google’s emphasis on code quality and rollbacks deserves attention alongside its claims about increased output.

Is Google Moving Towards AI That Can Improve Itself?

Pichai’s comments about AI coding also connect to a much larger question: how close are we to artificial intelligence systems that can help develop increasingly powerful versions of themselves?

This idea is known as recursive self-improvement (RSI).

In a fully developed recursive self-improvement process, an AI system would help create an improved version of itself. That newer version could then contribute to developing an even more capable successor, potentially accelerating progress over successive generations.

Such a process could have major implications for the future of artificial intelligence.

However, Pichai has cautioned against interpreting today’s developments as evidence that fully autonomous self-improving AI has already arrived.

He described Google’s current approach as a process in which AI agents assist human researchers with different aspects of model development. These agents can help researchers work more efficiently and coordinate parts of complicated tasks, but people remain involved in the process.

In other words, AI is increasingly helping to build AI, but that is not the same as a system independently improving itself without meaningful human involvement.

The distinction matters because assisting researchers with individual tasks is considerably different from reliably managing the entire process of designing, training, evaluating, and deploying a more advanced model.

What Google’s AI Safety Findings Suggest

Google DeepMind’s safety reporting provides additional context for understanding the company’s progress.

According to reporting on its assessment of Gemini 4 Argon, Google found no significant end-to-end acceleration of AI research and development attributable to AI, and assessed the associated risk as currently low.

The report also described growing use of AI agents in software development, while highlighting the continuing role of human oversight.

These findings suggest that AI is already contributing to important parts of technical work without necessarily being capable of automating the entire research process.

However, safety assessments have limitations. Their conclusions depend on the tasks being tested, the evaluation methods used, and the assumptions behind the experiments.

Independent researchers have also raised questions about the reliability of parts of Google’s internal benchmarking process, including problems affecting some benchmark tasks.

This does not automatically invalidate Google’s conclusions. It does, however, highlight why independent evaluation and transparent testing are valuable when companies make claims about the capabilities and risks of their own AI systems.

The most responsible interpretation is that AI-assisted research is advancing, while the extent to which AI can independently accelerate its own development remains an open question.

What This Means for Software Developers

Google’s changing approach to AI coding metrics reflects a wider transformation that could affect software developers, technology businesses, and people learning to code.

First, traditional measures of programming productivity may become less useful as AI takes on more routine work. Counting lines of code or measuring how quickly developers write them does not necessarily reveal whether a software project is successful.

Second, developers may need to develop new skills. Understanding how to work with AI coding agents, evaluate generated code, identify errors, and manage complex systems could become increasingly important.

Third, businesses may begin judging AI tools by their practical outcomes rather than their headline-grabbing capabilities. Faster development, fewer defects, lower maintenance costs, and more reliable software may matter more than the amount of code a model can generate.

Finally, human judgement will remain essential wherever software failures could affect security, finances, privacy, or critical services.

AI may change the work developers do every day, but effective software engineering will continue to require careful reasoning, testing, and responsibility.

The Bigger Picture: Measuring Results, Not Just Code

Sundar Pichai’s remarks point to an important change in the way Google views artificial intelligence.

In the early stages of AI-assisted programming, the percentage of code generated by machines offered an easy way to communicate progress. As AI becomes embedded throughout the development process, that single number becomes less informative.

A better assessment considers the complete picture: how much work engineers can accomplish, how difficult the problems are, how reliable the resulting software is, and how much human intervention is still required.

Google’s reported improvements in code submissions and rollbacks provide an example of this broader approach, although longer-term and independently verifiable evidence would help establish how much these gains can be sustained.

The same principle applies to AI research itself. Systems that assist human scientists in building more capable models represent meaningful progress, but they should not automatically be described as fully autonomous systems capable of improving themselves.

Artificial intelligence is moving beyond simple code suggestions towards more active participation in complex technical work. The next stage of progress will depend not only on what these systems can produce, but also on how reliably they perform, how their results are evaluated, and how effectively people remain involved.

For Google and the wider technology industry, the real measure of AI’s success may ultimately be less about who writes the code and more about what that code makes possible.