·Updated: How we work

AI-driven development, myth and reality: what "AI writes all the code" leaves out

AI has pushed the cost of generating code close to zero, but most of the cost of software was never in generation; it was in verification, integration and operations. Why the bottleneck moved from how fast code is written to how working code is proven.

AI developmentvibe codingverificationAI-nativeSI
Gaebari

Read in: 한국어·English·Español

"AI writes all the code now" is only half true. The true half is generation. AI can produce a function, a component or a full set of create/read/update/delete screens in seconds. The half it leaves out is everything else: whether that code actually does what was intended, whether it clashes with the existing system, whether it falls apart on edge cases, and whether it still holds up six months later.

The cost of software was never really in the typing. Checking that code is right, integrating it and fixing it when it breaks take far longer than writing it. AI made the cheapest part even cheaper and left the most expensive part almost untouched. So the bottleneck did not go away. It moved.

This post separates three myths about AI development from three realities. Not from a marketing angle, but from the operating angle of a team that has actually shipped work the AI-native way.

Myth 1: "AI takes care of everything"

A demo takes five minutes. Production is different.

Ask an AI to "build a shopping cart for an online store" and you get a screen that appears to work. The trouble starts after that. What if two people put the last unit of the same product in their carts at the same moment? What if the network drops during payment? Does the total come out right when a coupon and loyalty points are used together? AI does not raise these questions on its own. It will answer them if asked, but drawing up the list of questions is still a human job.

The illusion that it "takes care of everything" comes from confusing generation with specification. AI is good at turning a given specification into code. But deciding what to build, and what counts as "correct", is not generation. When that definition is missing, AI fills the gaps with plausible defaults, and the odds that those defaults match your business rules are not high.

Myth 2: "You can ship without verification"

The real risk with AI code is not that it can't write code. It is that it is plausibly wrong.

It compiles, the screen loads, the happy path works. It looks fine in code review too. Yet under certain conditions it allows negative stock, skips a permission check in one place, or handles decimals wrongly in a currency conversion. Bugs written by people usually catch the eye because they look awkward. Bugs written by AI match the tone of the surrounding code, so they hide better.

This is not just an impression. We ran internal measurements across different task sizes, using the same generate-and-verify pipeline Gaebari uses to build deliverables. The share of code generated by AI from a single prompt that passed the verify-gate (the tests) on the first try collapsed sharply as size grew: 93–100% for a single file of 300 lines of code (LOC) or less, an average of 48.7% for a single file of about 800 LOC, and 0% as soon as the work was split across two or more files. Even with the temperature fixed at 0, it collapsed deterministically at the same point every time.

"AI made it, so it's fast" and "AI made it, so I trust it" are different statements. Fast generation does not exempt anything from verification. In fact, because generation got cheaper, the volume of output that needs verifying went up, so if verification is not automated, overall speed actually drops. Having a person check everything by eye cannot keep up with the speed of generation.

Myth 3: "Development cost drops to zero"

The cost did not disappear. It moved.

Break down the total cost of a line of software and it falls roughly into writing, verification, integration, and operations and maintenance. Contrary to popular belief, writing is the smallest piece and maintenance is the largest. AI brought the smallest piece close to zero. What about the rest?

StageBefore AIAfter AI
Writing codePeople, slowAI, close to zero
Verification (proving it works)People, expensiveStill expensive, and more of it
Integration (joining the existing system)People, expensiveStill people
Operations and maintenancePeople, most expensiveStill most expensive

This is why projects that start out expecting "AI development is free" fail. Generation cost going to zero does not mean total cost went to zero, but if quotes and schedules are based on generation cost, the budget blows up at the verification, integration and operations stages. If you don't know where the cost went, you can't manage it either.

Reality 1: The real leverage is the verify-gate

Now that generation is cheap, what sets work apart is not how fast it is written but how you prove that it works.

At the core of how Gaebari ships deliverables is the verify-gate. Output made by AI is never shipped automatically. Only what passes verification defined in advance gets through. If it fails, it goes back into the generation loop. The condition for passing is not whether it looks plausible to a person, but whether it actually meets the defined criteria.

This gate is what stops the "plausibly wrong" problem described in Myth 2. It accepts as a premise that generation is cheap and imperfect, and makes the pass criteria strict instead. Make a lot, quickly, but ship only what is proven.

“Testing shows the presence, not the absence of bugs.”
Edsger W. Dijkstra — Computing Pioneer, —

A verify-gate is not a cure-all. As Dijkstra said, testing cannot prove the absence of bugs. But the value of a gate is consistency, not perfection. The strictness of verification does not swing with a person's mood or deadline pressure; the defined criteria are applied the same way every time. Since AI can repeat generation endlessly, verification has to be endlessly repeatable as well for the two to match.

Reality 2: The oracle-author loop, where people write the definition of correct

For a verify-gate to work, there has to be a definition of "what is correct". Writing that definition is called writing the oracle (the criteria that judge whether an answer is right).

In AI-native operations, this is where most of the human time goes. Instead of typing code, people define "what proves that this feature is correct": what output a given input should produce, which edge cases must be blocked, which invariants must never break. That oracle becomes the pass condition of the generation loop.

The loop is simple. A person writes the oracle → AI generates the output → the gate checks it against the oracle → if it fails, generate again; if it passes, ship. Generation is cheap, so it can run many times, and verification is automated through the oracle, so no one has to look at every attempt. Expensive human time is concentrated on setting the criteria.

This is why "AI replaces developers" is wrong. What got replaced is typing. What remains for people is the harder job: defining what is correct.

Reality 3: The human role didn't disappear, it moved

Before AI, most of a developer's time went into implementation. After AI, that time is redistributed to three places: specification (what to build), the oracle (what is correct), and judgment (catching what the gate missed and deciding trade-offs).

None of these three is generation. They all need context, domain knowledge and business judgment. Narrow it down to online retail and it becomes even clearer. "How should the settlement cycle be set?", "How does the returns policy interact with loyalty points?", "At what point is stock synced across sales channels?" AI does not make these decisions for you. It only implements, quickly, what has already been decided.

So AI-native development is less "people work less" and more "people do different work". More precisely, it keeps only the highest-leverage work, setting criteria and making judgment calls, and hands the rest to the machine.

Summary: myth and reality in one table

Myth (illusion)Reality
AI takes care of everythingSpecification and judgment are still a human job
Ship without verificationPlausible wrong answers have made verification more important
Development cost drops to zeroThe cost moved to verification, integration and operations

AI is a tool. A powerful one, but it does not decide for you what to build or what counts as correct. Using AI development well does not mean taking people out. It means building a system that pairs cheap generation with a strict verify-gate and ships only deliverables that are proven to work. In an era when generation is free, that is where the competitive edge is decided.

FAQ

자주 묻는 질문

Will AI replace developers?

Much of the work of typing code will be automated. But deciding what to build (the specification), what is correct (the oracle), and catching what the gate missed are still human jobs. The role does not disappear; it moves toward harder work.

Can I use AI-written code as is, without verification?

That is risky. The defining trait of AI code is not that it can't write code but that it is plausibly wrong. It compiles and the screen loads, but under certain conditions it can break your business rules. It needs to go through a verify-gate that proves it works against defined criteria.

Does developing with AI bring the cost down to zero?

The cost of generating code gets close to zero. But most of the cost of software was never in generation; it was in verification, integration, operations and maintenance, and that part stays. The cost did not disappear. It moved.


Gaebari ships its deliverables this way: pairing cheap generation with a verify-gate and shipping only what is proven to work. Contact: gaebari.com/contact

AI-driven development, myth and reality: what "AI writes all the code" leaves out | 개바리