Dlaczego musisz sprawdzać kod AI

Vibe coding lets you build fast by describing what you want. But here's the catch: AI writes code that looks right and sounds right, even when it's quietly wrong.

The model isn't lying on purpose. It predicts plausible code. Plausible isn't the same as correct, safe, or what you actually meant.

This lesson is your reality check: you must review AI code before you trust it. It's the single habit that separates people who ship working apps from people who ship broken ones.

Ask AI for a function to calculate a discount, and it might hand you something clean and readable that does the wrong math. No warning. No hesitation.

Run this snippet. It claims to apply a 20% discount, but read the logic carefully before you trust the output.

That code printed 99.8, not 80. It subtracted 20/100 = 0.2 from the price instead of 20% of the price.

The code ran without errors. No red text, no crash. That's the danger: AI bugs often hide as working-looking code that produces silently wrong numbers.

If you'd shipped this to a real store, every customer would get almost no discount and you'd never see an error message telling you why.

Understanding why helps you review smarter. AI gets things wrong because it:

• Guesses your intent — it fills gaps in vague prompts with assumptions.
• Mixes patterns — it blends snippets from millions of examples, some outdated or buggy.
• Hallucinates — it can invent functions, libraries, or APIs that don't exist.
• Loses context — it may forget a rule you set ten messages ago.

None of this means AI is bad. It means AI is a fast junior teammate whose work you sign off on.

Reframe your role. You're not just typing prompts — you're the reviewer and decision-maker. Tools like Cursor, Claude Code, and Copilot generate; you approve.

In Cursor and Claude Code, AI proposes changes as a diff (what it wants to add or remove) and waits for you to accept. That pause is your superpower. Use it.

Pro move: never click 'Accept All' without skimming what changed first.

For any chunk of AI code, ask yourself three quick questions:

• Does it do what I asked? Match it against your actual goal, not what looks impressive.
• Could it break or be unsafe? Empty inputs, missing data, exposed keys.
• Do I understand it enough to fix it later? If it's magic to you now, it'll be a nightmare at 2am.

These three questions take 30 seconds and save hours.

Here's a trick: AI is often great at finding problems when you point it at code — even its own. After it generates something, ask it to critique its work.

This isn't a guarantee, but it surfaces obvious issues fast. Paste a prompt like the one below into Cursor or Claude Code.

One of the most common AI mistakes: it imports a package that sounds real but doesn't exist, or uses a function the library never had.

The example below references a made-up helper. If you blindly trusted it, your app would crash on the first run with a 'is not a function' error.

Beginners worry reviewing slows them down. The opposite is true. A 30-second review now prevents:

• Hours hunting a silent bug that ships to users.
• Embarrassing demos where the app does the wrong thing.
• Security leaks like exposed API keys or open data.

Fast building plus light review is the real vibe coding speed. Building blind just front-loads the pain to later.

Build this loop into every session:

• Generate a small piece (not a giant blob).
• Skim the diff before accepting.
• Run or test it to see real behavior.
• Ask AI to explain anything you don't understand.

Small chunks + quick checks = catchable mistakes. Huge accepted blobs = bugs buried where you'll never find them.

The mindset of a great AI builder: trust the AI to move fast, verify before you ship.

You don't need to write every line yourself. You do need to be the human who checks that the line is correct, safe, and what you wanted. The AI is the engine; you're the driver with eyes on the road.

Next up, we'll learn how to actually read code you didn't write — even if you're new to it.

You ask AI for a discount function. It returns clean code that runs with no errors, but the math is wrong and customers get tiny discounts. What does this teach about reviewing AI code?

What you locked in this lesson:

• AI is confident but not always correct — it predicts plausible code, not guaranteed-right code.
• The scariest bugs run without errors and produce silently wrong results.
• You are the senior reviewer: generate fast, but skim every diff before accepting.
• Ask the three questions: does it do what I asked, could it break, do I understand it?
• Watch for invented libraries and let AI critique its own code.

Reviewing is the habit that makes vibe coding reliable. Next: how to read code you didn't write.