Stop Rubber-Stamping "LGTM": Why Pull Requests Are Operational Contracts (Especially with AI Code)

A developer argues that pull request approvals should be treated as operational contracts rather than rubber stamps, especially as AI coding assistants make generating syntactically plausible code nearly free. The author proposes a pre-approval test asking reviewers what their documented rationale would be if the change broke production, framing code review and architectural verification as the highest-accountability acts in the software lifecycle.

Be honest: when was the last time you opened a 400-line PR, scrolled through a sea of clean-looking green diffs, thought "looks fine to me" , typed LGTM , and hit merge? I've been guilty of it. Most of us have. Especially now, with AI coding assistants churning out syntactically plausible code in seconds, the temptation to rubber-stamp pull requests has skyrocketed. The formatting is neat, the unit tests pass on happy paths, and everything feels safe. Until Saturday at 3:00 AM, when an unhandled edge case drops production. When an outage happens, the postmortem doesn't care which AI generated the syntax. And the accountability doesn't just fall on whoever opened the branch. The very first question senior leadership and SRE ask is: "Who reviewed and approved this to go live?" I recently mapped this reality into a 1-page visual blueprint for our engineering practices: In software engineering, clicking Approve is not a friendly thumbs-up or an administrative chore to clear your notification queue. It is a sworn claim: "I have audited this architecture, I understand its failure modes, and I am willing to defend this code at 3:00 AM with my name attached." With AI tools making code generation essentially free, raw code output has zero scarcity. Anyone can prompt a 500-line feature in 30 seconds. Because syntax is now a commodity, code review and architectural verification become the single highest-leverage, highest-accountability acts in the entire software lifecycle. If we treat PRs like rubber stamps, we turn our repositories into synthetic sludge dumps. If we treat them like contracts, we safeguard the system and build real engineering trust. Here is a simple test I started running before clicking "Approve" on any pull request—especially AI-assisted ones: Ask yourself aloud: "If this breaks production tomorrow, what is my documented rationale for allowing it to ship?" If you don't have a clear, defensible answer ready, you aren't finished reviewing. I put together 30+ visual, one-page blueprints like this covering blast radius, distributed boundaries, and human engineering ownership in my new book: The Unshakeable Developer: Why AI Won't Replace True Software Engineers https://www.amazon.com/dp/B0HK2PCRJK . How is your team handling code reviews lately? Have you noticed "review fatigue" creeping in with AI-generated diffs? Would love to hear how you deal with it in the comments below