June 2026 Summaries

In the context of AI-generated code, testing and verification serve distinct roles, with testing focused on ensuring code behaves correctly and verification ensuring it aligns with pre-established agreements. Traditionally, humans have carried the intent behind code, allowing for informal verification through code reviews. However, AI agents lack intent and rely solely on prompts, which can lead to code that passes tests but does not meet agreed-upon specifications. Aviator's experiment demonstrated that AI can efficiently verify code against a detailed specification, identifying intent mismatches that traditional CI pipelines might miss. This highlights the need for explicit, human-approved specifications and verification as distinct stages in the development process, with testing continuing to address bugs and regressions. Verification has thus become a critical layer in the pipeline, ensuring that AI-generated code aligns with human intentions, while reducing the reliance on human code reviews for extensive, line-by-line checks.

Code review processes face significant challenges with the rise of AI-generated code, as these tools produce code faster than humans can effectively review it, creating a bottleneck that traditional methods cannot address. Research has shown that as changes grow larger, the quality of review comments declines, and AI tools have exacerbated this issue by increasing the volume of pull requests while prolonging review times. Relying on AI to review AI-generated code introduces problems such as non-determinism, intent misinterpretation, and blind spots due to the same model both generating and reviewing the code. To overcome these challenges, teams are advised to codify review feedback into deterministic checks, shifting human involvement to earlier stages where intent is defined, and utilizing AI where deterministic methods fall short. By moving the focus from code review to intent verification, teams can ensure that the final product aligns with the original purpose and constraints, allowing AI to assist in enforcing standards and verifying specifications without compromising quality.

AI-generated code now constitutes 42% of committed code, but contrary to expectations, review times have increased rather than decreased, as AI accelerates code production but not the review process. Faros AI's analysis of 10,000 developers showed that while teams with high AI adoption merge significantly more pull requests, the review time has increased by 91%, suggesting that the bottleneck lies in the human review process. The problem is not merely a matter of scaling by adding more reviewers, as this does not address the context required to catch intent issues with AI-generated code. Instead, a shift towards intent-driven verification is recommended, where code is checked against a pre-approved specification before implementation begins. This approach can be programmatically executed and has been demonstrated to be effective in experiments where AI-generated code was efficiently verified against human-specified acceptance criteria. Moving human judgment upstream and allowing tools to handle the verification process can help manage the increased load without compromising code quality, emphasizing the need for a structured approach to specification and verification in AI-assisted development.