Your tests were written by the thing they're testing

Your tests were written by the thing they are testing. This is rarely said about AI-generated tests.
You let a model write the code. Then you let the same model write the tests for it, with the same training and the same blind spots. When the code is quietly wrong, the test is wrong in the same way, and it approves the bug. The test becomes a mirror, and a mirror always agrees with you.
Coverage reaches 100% and the pipeline is green. Then production does the one thing your tests could not: it disagrees.
A test should record what you meant. A test generated from the code only records what the code did.
In one survey, respondents said that 61% of their company's code is now written by AI (The Register, May 2026). Many of the tests guarding that code came from the same place.
More coverage will not fix this. What will is a person writing down what "correct" means before the model has a say.
So who writes the assertion: you, or the thing you are testing?
The data behind this post
- The Register, 20 May 2026, reporting CloudBees' survey: respondents said 61% of their organizations' code was generated by AI or with AI assistance: theregister.com