Skip to main content

AI Writes Code Faster Than Humans Can Review It

Core idea: Verification becomes the bottleneck.

The Verification Tax​

One of the four forces reshaping software teams, as identified in Steven Brovich's AWS leadership talk, is the verification tax:

"AI generates code 10 times faster, but it's three times harder to validate. The reviewing bottleneck eats the velocity if you're not careful."

This is the hidden cost of AI-assisted development. The generation speed is intoxicating. But every line of generated code needs to be reviewed, tested, and validated — and that work doesn't get faster just because the code was generated faster.

The Anthropic Hackathon Data Point​

At Anthropic's "Build with Claude" hackathon (February 2025), 13,000 applications were submitted. 500 were accepted. 277 shipped production code. 21 million lines of code were generated.

But the winners weren't the teams that generated the most code. The winners were domain experts who understood the problem deeply and could evaluate what the AI produced. The bottleneck wasn't generation — it was validation.

Why Reviewing AI Code Is Harder​

AI-generated code presents unique review challenges:

  1. It looks plausible. AI code is typically well-formatted and syntactically correct. The bugs are in the logic, not the syntax.
  2. It's unfamiliar. You're reviewing code written by a non-human mind. The patterns and approaches may be unexpected.
  3. It's voluminous. The sheer volume of code generated means reviewers are overwhelmed.
  4. It hides subtle bugs. Security vulnerabilities, race conditions, and edge cases are easy to miss in generated code.

The Deskilling Trap​

Brovich also highlighted the deskilling trap:

"Juniors using AI can ship about 17% more code, but they understand 17% less of what they've actually shipped. The people coming through your pipeline are getting faster and less grounded at the same time."

This is the paradox. AI makes everyone faster, but it also makes it easier to ship code you don't fully understand. The review process is the safety net — but only if it's taken seriously.

How to Survive the Verification Bottleneck​

  • Automate what you can. Unit tests, integration tests, linting, and security scanning should all run automatically on every PR.
  • Review in layers. Don't try to review everything at once. Separate architectural review, security review, and functional review.
  • Invest in testing. The more comprehensive your test suite, the less you need to rely on manual review for correctness.
  • Know when to say no. Not all generated code is worth keeping. If you can't validate it, don't deploy it.
  • Build small, review often. Smaller PRs are easier to review thoroughly than massive generated code dumps.

What This Means for You​

  • Generation speed is not delivery speed. The bottleneck has shifted from writing code to verifying it.
  • Review processes need to evolve. The same review workflows that worked for human-written code may not scale to AI-generated code.
  • Testing is your best friend. Automated tests are the most scalable way to validate AI output.
  • Don't let velocity compromise safety. The verification tax is real, but skipping it is how vulnerabilities get deployed to production.

Portions of this article are based on insights from Steven Brovich's talk "A leader's guide to advanced team structures in an agentic world" at AWS Events. You can watch the full talk here: A leader's guide to advanced team structures in an agentic world.

Last updated on