The Cost of AI Isn't Generation. It's Verification.

SHOW ME THE CODE, THEN SHOW ME THE IDEA

In 2000, Linus Torvalds wrote the line that became software engineering's most quoted sentence: “Talk is cheap. Show me the code.” Code was the scarce thing then, expensive to write, the only proof that an idea actually worked.

That changed gradually, then suddenly. GitHub Copilot arrived in 2021 and was met mostly with developer skepticism. Then, on February 2, 2025, Andrej Karpathy tweeted a name for what was happening: vibe coding, “you fully give in to the vibes... and forget that the code even exists.” It got 4.5 million views and became Collins Dictionary's Word of the Year. Eighteen months later, Karpathy himself called the term passé, replacing it with “agentic engineering,” done “with more oversight and scrutiny.”

Show me the code became show me the idea. Now it's becoming show me that someone checked.

THE SPEED IS REAL

Roughly 84 to 92% of professional developers now use AI coding tools regularly. For prototyping, median task completion time drops 20 to 45%, a number that holds up across independent studies, not just vendor marketing. Gartner projects 40% of new enterprise production software will use these techniques by 2028. In Y Combinator's Winter 2025 batch, a quarter of startups had codebases that were 95% or more AI-generated.

BUT SPEED AND SKILL DON'T ALWAYS MATCH

Notice what happens when you actually measure the people using these tools, rather than just counting how many of them there are.

A controlled METR study followed 16 experienced developers on mature million-line codebases. AI didn't save time. It added overhead. The developers predicted 20% faster. Most still believed they'd been faster after the study ended, even though the data said otherwise.

A separate 2026 Science study found early-career developers adopted fastest and gained nothing measurable, while senior developers adopted more cautiously and gained real productivity. The tool wasn't the variable. Judgment was.

WHERE THIS GOES WRONG

A pattern engineering teams now call the 90-day reckoning: a small team ships fast for a few months, everything works, then a minor bug takes a week to fix, not because it's complex, but because nobody, including whoever generated the code, can trace why it works the way it does. A function is 300 lines long, handles three unrelated things, and the commit says “add user flow.”

That story has numbers behind it. An Uplevel study of roughly 800 developers found speed gains neutralized by a 41% increase in bug rates. An analysis of 8.1 million pull requests confirmed technical debt rises 30 to 41% after adoption. GitClear found code duplication at four times the previous rate, refactoring down from 25% of changed lines to under 10%.

Put the two Gartner numbers next to each other and they stop looking contradictory: 40% enterprise adoption by 2028, and a 2,500% increase in defects without governance, forecast for the same year. That's not two predictions. It's one shift, described from two angles.

WHAT ACTUALLY SEPARATES THE TWO OUTCOMES

The 2025 DORA report said it more precisely than most commentary around it: AI acts as an amplifier. It strengthens high-performing teams and worsens the dysfunction of teams that already struggle.

That single finding explains almost everything above it. The failure pattern isn't “AI writes bad code.” It's prompt, generate, glance, commit, with nobody reviewing what shipped.

A researcher who audited 50 AI-built apps in early 2026 found 88% had database-level security entirely disabled. Not misconfigured. Disabled. Nobody had checked.

WHAT I ACTUALLY LOOK FOR

I don't write code. I evaluate companies that run on it.

The question I'm surfacing during diligence is never whether a company uses AI tools. Almost everyone does now. It's whether anyone reviews what shipped, or whether the codebase is a pile of things that worked on the first try and nobody looked at twice.

I ask for commit history, not the pitch deck. A few things tell me almost everything. Do commit messages describe what changed and why, or do they all say some version of “update”? When something breaks, does the team revert immediately, or do they debug forward under pressure and leave three half-fixes in the code? Is there a pattern of checking a claim or a number before it reaches a customer, or does confidence substitute for verification? Can any engineer explain why a piece of business logic exists, or only that it currently works?

A team that reverts fast, checks before it ships, and can still explain its own code eighteen months later is buying itself years. A team that can't is buying itself a rescue-engineering bill it doesn't know about yet.

None of that shows up in a demo. It shows up later, in exactly the kind of conversation nobody enjoys having.

WHAT 2027 PROBABLY LOOKS LIKE

The market isn't slowing down. Projections put it at $12.3 billion by 2027, up from roughly $4.7 billion in 2026. But the term itself may not survive that growth. Analysts expect “vibe coding” to be largely retired from professional vocabulary by 2027, replaced by the language Karpathy reached for: agentic engineering, governed AI development, spec-driven delivery. The activity keeps growing. The name for doing it carelessly is going out of fashion, because doing it carelessly is what's producing the bills.

At least one Fortune 500 company is expected to have a public, AI-traceable incident causing more than a billion dollars in damages by 2027, a forecast analysts consider close to inevitable given the incident rate already logged in 2025 and 2026.

Meanwhile, AI code review and verification is projected to become a $10 billion category by 2028, growing faster than code generation tooling itself. The market already senses where the real bottleneck sits. It was never generation speed. It's whether anything gets checked before it reaches a customer.

THIS ISN'T AN ARGUMENT AGAINST VIBE CODING

It's an argument for knowing which version you're doing. Fast and unreviewed is one risk profile. Fast and verified is a different one entirely, even though both got called the same name for the first eighteen months of this technology's existence.

AI has reduced the cost of writing software. It hasn't reduced the cost of understanding it.

Code generation is becoming a commodity. Judgment is not.