AI-Generated Code: Who Reviews It? The Paradox of Increasing Developer Stress

Sam Newman, a software engineering expert, emphasized that AI coding agents excel at helping produce more code than before, raising the question of how much AI-generated code needs to be reviewed by humans. He explained that this question is emerging as a critical challenge in modern software development.

Newman questioned whether the exponential increase in AI-generated code necessarily leads to overall efficiency improvements in the entire development process. He argued that a deep consideration of the human role is needed to ensure the quality and reliability of AI-generated code.

Sam Newman cited several studies, pointing out that AI's ability to produce more code does not translate into improved throughput for the entire software development cycle. He explained that while developers write more code, this does not necessarily lead to more product releases.

This phenomenon suggests that the speed of code generation itself is not the bottleneck in the development process. Newman emphasized that beyond simply increasing code volume, the focus should be on improving the efficiency of the entire development workflow that leads to actual product releases.

Sam Newman stated that what the aforementioned studies suggest is that the ability to generate code is not the bottleneck in releasing good, working software. He emphasized that code generation itself is already fast enough and is not the primary cause of delays in the development process.

Newman posed an important question: if code generation is not slowing down development, then what is the true bottleneck in software development? This question serves as a crucial starting point for rethinking the software development workflow in the AI era.

Sam Newman identified code review as one part of the software development cycle that increasingly appears to be a major bottleneck for those using AI-written code. The explosive growth in AI code generation volume is overwhelming traditional code review processes.

He explained that while the introduction of AI has accelerated the speed of code generation itself, the exponentially increased volume of code that needs to be reviewed is causing delays in the overall development process at the code review stage. This is emerging as one of the most urgent challenges to address in software development in the AI era.

Sam Newman announced an in-depth discussion on code review in the AI era, presenting three main topics. First, he will analyze the fundamental reasons why AI makes code review more difficult. Second, he will re-examine the purpose of traditional code review processes and evaluate if they are still valid in the AI era.

Finally, Newman stated that he would discuss what changes need to be made to how software is created and how organizations worldwide are re-examining their relationship with code to make the most of AI agent coding. These three topics will comprehensively cover the problems and solutions faced by software developers in the AI era.

According to Feros' March acceleration report this year, the capabilities of AI coding agents clearly led to the generation of more Pull Requests (PRs). Specifically, teams with high adoption rates of AI tools included more code within the PRs themselves.

The report indicated that the average size of generated PRs increased by over 50% compared to the previous year, while developers were simultaneously handling more work threads. This numerically proves that the introduction of AI contributed to an explosive increase in code generation volume.

Sam Newman pointed out that despite the Feros report's findings that more code is being generated thanks to AI, actual software release volumes are decreasing. He explained that other reports, including Feros's own research, also show that the overall software development process is slowing down.

Newman emphasized that many developers are producing a smaller volume of products than before, which he described as a paradoxical outcome of AI adoption. This raises the important implication that an increase in code generation volume does not necessarily lead to an increase in final product release volume.

Sam Newman cited the concept of Lean Manufacturing, explaining that a system can become clogged when there is a lot of Work In Progress (WIP). Even if all developers are working diligently, the processing speed of tasks significantly decreases when there are many PRs waiting for review and developers frequently switch contexts.

He analyzed that while it may seem like a lot is happening, there is not much substantive progress. This implies that AI, by increasing code generation while simultaneously burdening the review process, ultimately reduces the overall system's throughput.

Sam Newman outlined the four main goals of traditional code review as follows: First, 'accuracy' is to ensure the code works as intended and to instill confidence from a functional and security perspective. Second, 'shared learning' is the process where the PR submitter learns through feedback, and reviewers also gain knowledge by examining others' code.

Third, 'strategic and practice alignment' is to ensure the code aligns with the organization's strategy and coding practices. Fourth, 'increased codebase awareness' is to enhance understanding of the overall codebase by reviewing others' work and fostering a sense of connection among team members. These four goals define the core role of code review in traditional software development.

Sam Newman stated that with the increase in PRs and code review volume, the human review aspect of the PR process is under considerable strain. He analyzed that this phenomenon is likely a major cause of the rising burnout reports among developers who heavily use AI tools.

Early evidence shows that AI usage is actually increasing, rather than decreasing, developer stress and pressure. This is a paradoxical outcome that contradicts the initial expectation that AI adoption would enhance developer work efficiency.

Sam Newman emphasized the paradox of AI adoption by citing a recent Harvard Business Review article, 'AI Doesn't Reduce Work, It Intensifies It.' The article mentioned that many employees are performing more tasks simultaneously and feeling more pressure than before using AI.

This suggests that, contrary to initial expectations that automation would save time and reduce such pressure, it is actually increasing work intensity. Unlike the anticipation that AI would reduce repetitive tasks, it is adding new kinds of burdens.

Sam Newman once again highlighted the paradoxical situation where AI is introduced to reduce pressure but actually leads to increased pressure. He cited a shocking figure from the 'Whiplash' report, which indicated a 225% increase in time spent on work in progress.

This figure shows that more tasks are being handled simultaneously, and developers are constantly switching between tasks. Newman analyzed that AI appears to be providing developers with more work, making them process more tasks, rather than helping them perform more efficiently.

Sam Newman stated that to cope with the problems arising from the increasing number and size of PRs, more and more companies are using AI tools to review code. Interestingly, this code is largely written by AI.

This means that a situation where AI reviews code written by AI is becoming a reality. Newman posed the fundamental question, 'What then is the human role?', emphasizing the need for a radical redefinition of the human developer's role.

Sam Newman emphasized the need to re-examine the purpose of code review in a world where AI writes and modifies code. He raised a fundamental question about whether the same benefits can still be derived from human-based code review processes.

Regarding the first goal of code review, accuracy verification, Newman argued that humans have never been able to fully verify code accuracy through manual review, and it will be even more impossible now with increasing code volumes. This suggests that even the core goals of traditional code review must be re-evaluated in the AI era.

Sam Newman explained that it is impossible to teach AI coding agents to learn through feedback like humans do, regarding the second goal of code review, 'shared learning.' He pointed out that agents do not learn like humans and cannot significantly influence the training process of next-generation LLMs (Large Language Models) during review.

This means that while knowledge sharing and growth among human developers were a crucial part of code review, this 'shared learning' purpose is not valid for AI-generated code. He therefore concluded that a redefinition of the learning targets and methods is necessary for AI code review.

Sam Newman emphasized that while AI agents can be nudged to create better code, this is fundamentally different from the relationship of giving feedback to a human. He explained that feedback to AI primarily serves to collect data for system improvement and is far removed from human developer growth or interaction.

These differences suggest that a new approach to the meaning and application of 'feedback' is needed in the context of AI code generation. It is difficult to apply human-centric feedback models directly to AI.

Sam Newman argued that AI-written editorial summaries would actually be far more useful for understanding what has changed than manually reviewing code. He explained that it is much more efficient for AI to summarize key changes than for humans to manually review vast amounts of AI-generated code.

This implies that human developers can quickly grasp overall changes through AI-provided summaries and focus on important design decisions, rather than reviewing the detailed parts of AI-generated code. It suggests the possibility of optimizing the code review process by leveraging AI's strengths.

Sam Newman reported that Chainway, a well-known supply chain security company, recently announced that it would cease human code review and instead dedicate more time to collaboration on design decisions and related documentation. This is an important example demonstrating the changing paradigm of code review in the AI era.

Chainway's decision shows a commitment to focus on high-level design and architectural discussions—what to build and why—rather than low-level review of how code works. This suggests that the human role is shifting from direct code review to more strategic areas.

Sam Newman explained that the idea of allowing AI to handle code generation more permissively is very similar to the small but growing concept of 'Dark Factories' in industrial automation. A Dark Factory refers to a factory operated solely by robots and automation systems without human intervention.

This analogy illustrates the possibility that AI coding could evolve towards generating and reviewing code without direct human involvement. This suggests that in the software development process, the human role might focus on overall system management and design, rather than direct participation in code generation and review.

Sam Newman stated that many companies he has spoken with, especially those more advanced in using agent AI coding, are actively adopting this Dark Factory model for code. This means that AI-driven software development methods are moving beyond the experimental stage and being applied in real industrial settings.

These companies are seeking changes by having AI handle code generation and some review processes, allowing human developers to focus on higher-value tasks. Newman suggested that this model could be a major direction for future software development, emphasizing the need for preparation.