There is one rule that covers every AI coding tool: never ship a line you cannot explain.
What the rule means#
The rule is not "never ship AI-generated code". Generated code is fine. The line is whether you understand it. For each line, you should be able to say:
- what it does
- why it is correct
- what would break if it were wrong
If you can, ship it. If you cannot, do not ship it until you can.
This matters because bugs hide in code you do not understand. It also matters in interviews, where you will be asked about your code.
| Weak answer | Strong answer |
|---|---|
| "The AI wrote it and it seemed to work." | "I used an AI tool for the first version, reviewed it, caught a missing null check, and tested it against these cases." |
Where the tools help and where they do not#
| Help | Do not help, and can slow you down |
|---|---|
| Boilerplate | Subtle concurrency bugs |
| Tests | Novel algorithms |
| Migrations | Security-critical paths |
| Explaining unfamiliar code | Anything without tests |
| Refactors with good test coverage |
A 2025 study by METR found that experienced developers working in their own large repositories took about 19% longer with AI tools than without, while believing they were faster. The gains are larger on boilerplate and unfamiliar code. The losses come in complex problems, where the tool confidently produces plausible but wrong solutions.
Help is largest where you need it least#
AI tools help most on the parts you already understand well, because there you can check the output quickly. They are most dangerous on the parts you do not understand yet, because that is when you are least able to catch a mistake.
Short line#
Never ship a line you cannot explain: what it does, why it is correct and what breaks if it is wrong. AI tools help most with boilerplate, tests and well-tested refactors, and are most dangerous on code you do not yet understand.
Key takeaways#
- AI-generated code is fine to ship if you can explain every line.
- Explain what it does, why it is correct and what breaks if it is wrong.
- Tools help with boilerplate, tests, migrations, explaining code and well-tested refactors.
- They struggle with concurrency bugs, novel algorithms, security paths and untested code.
- METR (2025): experienced developers in their own repositories were about 19% slower with AI tools, but believed they were faster.
- The tools help most where you can already check their work.