Code Faster with AI, Proofread Longer: The Invisible Debt

Coding assistants are speeding up deliveries and inflating merge metrics. However, the bottleneck is shifting towards specification and auditing, with review times rising sharply and trust remaining fragile. Precise practices are proposed to prevent delegation from turning into a night queue.

Review time increases by about 90% with AI in large teams

In teams that are heavily adopting AI tools, the time spent reviewing pull requests has increased by approximately 90% based on data from over ten thousand developers. This increase is interpreted as a volume effect surpassing human review capacity. A significant portion of production code is now co-written by systems, while about three-quarters of developers report frequently observing hallucinations and not fully trusting the outputs. The gap between delegation and trust fuels a "verification debt": the time saved in writing returns in the form of poorly understood code audits that appear out of nowhere at 2 AM. These dynamics are particularly pronounced as AI is fast with small and new tasks but less reliable with large and old ones.

Being five times faster can lead to losing track just as quickly

Acceleration can come with a loss of focus. Developers are rereading pull request descriptions written by AI to remember their own intentions. Seemingly clean changes—net diffs, consistent naming, passing tests—can mask intention errors such as overly broad permissions introduced due to vague instructions. A colleague pointing out a wildcard before merging avoided a regression. The takeaway is clear: passing tests do not equate to fulfilling the intention, especially if the tests verify the existence of a permission rather than its safety.

The bottleneck shifts towards precise articulation and proof

Generating a first draft is almost free with AI. The main cost becomes formulating unambiguously what is expected and providing proof that the result aligns with those expectations. Experienced open-source developers have measured about a 19% slowdown with AI on their own large codebases, while they thought they were 20% faster; these results were later revised for sampling bias and are considered a signal. Senior engineers report that, on a large system already mastered, AI often does not save time: it shifts the effort from writing to verification, which may seem less but is not. The contrast persists between small and new tasks, where AI excels, and large and old sets, where reliability decreases.

Why the temptation of the "fourth terminal" persists

Three drivers fuel the race for parallelism: applause after accelerated deliveries, focus on a measurable backlog rather than intangible understanding, and fear. Research among over three thousand developers indicates that between 43% and 45% feel real stress about their professional identity in the face of AI tools. Running more agents gives the illusion of keeping up. Teams with a learning culture feel this stress much less than those operating under pressure and competition. In this context, the temptation to multiply sessions—up to a fourth—sets in. This has been practiced for months by practitioners who see it as the critical error of the assisted coding era, not for parallelism itself but for the misdirected use of the gained capacity.

Practices to reduce verification debt

Several concrete levers are proposed. Consult the entire reference documents, including edge cases, while an agent is in action, to guide prompts using a structured representation rather than mere impressions. Write a one-page specification detailing expectations, exclusions, definitions of what is complete, and points requiring human review: in one mentioned example, a single sentence prohibiting wildcards would have sufficed; spending fifteen minutes drafting the specification could have avoided a week of drift. Position oneself as an architect who develops the plan, delegates execution, and retains final validation; consider the code as a possible manifestation of intention. Ask a peer for ten minutes for a review: pairing on the review surpasses a fourth terminal, with the main failure being the lack of review, and a second human is the least costly review. Publish work notes and decisions to encourage feedback: two discussions led to changes in the structuring of environments. The underlying principle is to invest the capacity freed by AI into non-automatable tasks rather than more sessions. The first rule stated: specify first, always; if the intention cannot be clearly articulated, the request to AI remains premature, as the tool will fill in the ambiguity without asking questions. These practices do not deny the real gains observed—work completed before lunch, migrations sketched in an afternoon, managerial recognition, experimental results showing speed gains and an increase of about a quarter in merges—but aim to frame these gains to limit the resulting debt.