Writing tests is one of those tasks everyone knows they should do and almost nobody enjoys. AI can take a big chunk of that work off your plate in 2026… if you use it well. Because misused, it fills your project with useless tests that give false security. Here’s how to write good tests with AI, based on our experience.
Why AI is good at writing tests
Generating the repetitive cases, the “scaffolding” and the happy-path checks is exactly what’s most boring and what AI does in seconds. To kick off coverage for a module without starting from scratch, it works very well.
The danger: useless tests
AI tends to write tests that check the obvious or simply mirror the implementation (if the code is wrong, the test passes it anyway). A test that always passes protects nothing; it just inflates coverage and gives you a false sense of security.
How to get tests that actually help
- Ask for edge cases, not just the happy path: empty inputs, nulls, extreme values, errors. That’s where a test’s value is.
- Tell it the behaviour you expect, not “write tests for this”. A test should verify intent, not implementation.
- Review every test: would it fail if the code were wrong? If not, it’s useless.
- Make them readable: a test is also documentation of how the code should behave.
Our experience writing tests with AI
- What saves us most: the scaffolding and the repetitive cases. That alone removes a huge amount of drudgery.
- The mistake we made: accepting the batch of tests without looking. They all passed… because they tested nothing real. Now we explicitly ask for edge cases and check that each one can fail.
- The trick that works best: describing the expected behaviour with a couple of “this should happen” and “this should fail” examples. The quality jumps.
Our advice: use AI for the muscle (writing many tests fast) and bring the brain yourself (what actually needs testing). A test that can’t fail is worse than no test, because it lulls you. Always review.
Frequently asked questions
Does 100% coverage with AI mean the code is good?
No. Coverage measures executed lines, not whether the tests are good. You can have 100% with tests that check nothing useful. Quality over quantity.
Can I trust AI-generated tests?
Only after reviewing them. Check that they verify the expected behaviour and would fail if the code were wrong.
Conclusion
AI removes the drudgery of writing tests, but the judgement of what to test is yours. Ask for edge cases, verify each test can fail and treat them as documentation. To review code well, see how to review AI-generated code and how to debug with AI.
