AI
Human Accountability Must Drive AI Code Development, Not Tool Blame
As AI writing code becomes mainstream, shifting role boundaries demand engineers ask better questions—but only if organizations build cultures where people dare to take responsibility.
InfoQ
In a wide-ranging conversation on the InfoQ Engineering Culture Podcast, Hugo Venturini, a software engineer and current CTO of an AI-focused startup, argued that the real challenge AI poses isn't technological—it's organizational. The danger, he warns, lies not in what AI can do, but in how humans abdicate responsibility when it fails.
Who Owns the Code AI Produces?#
Despite the rapid adoption of AI code generation, the fundamental principle must not shift: humans remain accountable for decisions and outcomes, even when machines assist in creation. Venturini emphasizes that AI is simply a powerful tool. When it produces faulty code, the engineer or team deploying it bears responsibility, not the algorithm. This principle seems obvious in theory but breaks down in practice when organizations allow developers to blame the tooling rather than examining their own judgment in accepting and merging the AI's output.
The problem compounds when leaders themselves sidestep accountability by framing AI as a black box. Instead, Venturini advocates for applying proper feedback loops—tooling that signals to the AI how to improve and return better results—ensuring that by the time humans review generated code, much of the refinement has already been automated away.
Pull Requests and Code Review at Scale#
One visible symptom of AI adoption is exploding pull request sizes. Teams report PRs reaching 2,000 lines or more, with reviewers drowning in backlogs of hundreds of changes. Venturini argues this is not a new problem: large pull requests and review bottlenecks predated AI by years. The mistake many teams make is treating every line as review-worthy.
Before AI, organizations learned to distinguish between boilerplate—code that compiles, passes tests, and behaves as intended—and the genuinely novel or risky portions of a product that warrant scrutiny. That same principle applies today. Some engineers, particularly senior ones with strong track records in their teams, can merge code without review, accepting personal accountability if something breaks. The real question is not whether to review everything, but what actually needs human eyes. AI hasn't changed that calculus; it's only made the scale more visible.
Venturini acknowledges the outstanding challenge: how do teams manage the sheer volume of PRs that AI enables? This remains an open question the industry has yet to solve convincingly.

