Unit tests very roughly double the amount of effort required to make any meaningful change to your codebase. They are also require maintenance same as ordinary code - but the customer does not care in the slightest whether or not they pass. On the other side, they can only really tell you about low level bugs that you already expected, they cannot surface system level bugs - the actual hard bugs that cause problems for you and your customers.

Then there is the danger of thinking that green=all good, an example of 'automation bias' where we learn to trust the automation even as things go wrong.

As makers, it is also tempting to believe that [all] problems can be solved by making something (i.e. code), but actually many problems are not of that nature, and cannot be solved in that way.

Thank you, that makes sense. What I meant was that today all unit tests are basically written by an AI so the "cost" is almost zero. Am I wrong?

Sorry yes. If an LLM written unit test fails, then it has to be determined whether the test was wrong or the code was wrong. This is an expense in human oversight, unless of course we believe that LLMs will get it right at a high enough rate that they can be left to code everything themselves completely automatically.