AI Code Generation Is Shifting the Bottleneck to Review

For years, the bottleneck in software was writing code faster.

AI has dramatically improved that. Code generation is now easier, faster, and more accessible than ever.

But optimizing one part of a system doesn’t automatically make the whole system more efficient. It creates an imbalance.

AI has over-optimized code generation, and in doing so, it has exposed everything else that wasn’t designed to scale.

Output has increased significantly

Recent data shows a consistent pattern: AI increases the volume of code produced.

According to Faros AI, teams using AI saw:

+98% pull requests
+154% PR size
+21% tasks completed

Even in simple cases, the effect is clear. A basic endpoint implemented with AI generated 6.4× more code than a manual version, according to LogRocket.

This is real productivity, but mostly at the level of generation.

The constraint moved downstream

That code still needs to be read, understood and validated.

And that part didn’t get proportionally faster.

Code review time increased by 91%. Stack Overflow also reported that 45% of developers say debugging AI-generated code takes longer than writing it.

In practice, a manual PR may take around three minutes to review, while an AI-generated PR can take between eight and twelve minutes.

The system didn’t become inefficient. The constraint simply moved.

Trust is not keeping up

In most technologies, usage builds trust. Here, adoption is outpacing confidence.

The dominant failure mode is not that AI-generated code is obviously wrong. It is that it is almost correct, but not quite.

That creates a subtle but important effect: more verification, more cognitive load and more time spent confirming correctness.

AI reduces the cost of producing code. But it increases the cost of trusting it.

The cost is redistributed, not removed

This is not just a technical shift. It is also an economic one.

The work doesn’t disappear. It moves.

As Daniel Stenberg, creator of cURL, puts it: “AI doesn’t increase the capabilities of the humans in the loop. It shifts the cost to review.”

This is already visible in real workflows. In the case of cURL, only around 5% of bug bounty submissions were valid, while around 20% were AI-generated noise.

The system didn’t get simpler. It got rebalanced.

The system is under new pressure

At the individual level, developers are faster.

At the system level, pressure builds elsewhere: more time reviewing, more time debugging and more coordination overhead.

DORA signals from 2025 and 2026 point in the same direction: AI can increase local productivity, but it also changes where friction appears across the software delivery system.

The key insight is simple: we didn’t remove constraints. We moved them.

Closing

AI is not making software engineering inefficient.

It is making one part extremely efficient and exposing everything else that isn’t.

Which leads to a deeper question:

What happens when the system we rely on to control quality can’t keep up with the speed of generation?

AI Code Generation FAQs

1. Does AI code generation improve developer productivity?

Yes, AI code generation can increase individual output, but the impact depends on whether review, testing and deployment workflows can scale with that output.

2. Why is AI-generated code harder to review?

Because it can be longer, more defensive, more verbose or almost correct but not fully aligned with the system’s architecture, standards or security requirements.

3. What is the main risk of AI-generated code?

The main risk is not only incorrect code, but code that looks correct enough to pass quickly while still increasing debugging, validation or maintenance costs.

4. How should engineering teams adapt to AI coding tools?

Teams should track review time, PR size, bug rates and delivery stability, while defining clear guidelines for when and how AI-generated code should be reviewed.