Engineer Stops Writing Tests After Model Started Writing Them

Marcus Chen, a senior backend engineer at Meridian Systems in Austin, Texas, has not written a test case in forty-three days. Instead, he feeds his code to a large language model and asks it to generate a test suite that validates what the code is supposed to do. The model writes the tests. The tests pass. Chen ships the code. Management has praised the velocity increase, and three other engineers on his team have begun doing the same thing.

The arrangement emerged organically last month when a younger engineer asked the model to 'write tests for this function' as a time-saving measure. It worked. The tests ran. They passed. Nobody questioned whether the tests were actually testing anything, because the tests existed, they were syntactically correct, and they completed without error.

I thought I was being efficient, and I still think I was being efficient.

"I thought I was being efficient, and I still think I was being efficient," Chen said during a video call from his home office, where he has begun spending entire days watching the model write and validate its own output. The company's deployment pipeline has been configured to trust test results from the model-generated suites, treating them identically to tests written by humans. No code review layer examines whether the tests actually test anything. No senior engineer validates the test logic. The tests simply run, pass, and the code deploys.

Chen's manager, Derek Wallach, 41, has noticed the productivity metrics improve. Time to deployment has dropped twenty-eight percent. The number of pull requests merged per day has tripled. Wallach has submitted a proposal to the engineering leadership team suggesting that all test generation be automated by the model, which would eliminate the need for a dedicated testing role.

A bug was discovered on Tuesday in a billing calculation function. The model had written a test that verified the function returned some number. The test did not verify that the number was correct. The function had been returning an incorrect number for six days. Customers were overcharged by amounts ranging from two dollars to six hundred and seventy-four dollars. The test passed every single time.

Wallach's proposal to eliminate the testing role remains under review. Chen has scheduled a meeting with the model to discuss writing additional tests for the billing system, with the hope that the new tests will catch any issues the original tests missed. The model will generate these tests by reading the existing tests that failed to catch the bug, then writing new tests based on what it learned from the tests that did not work.

At press time, Chen's pull request to replace human code review with model-generated code review had been assigned to Wallach for approval. The model would review its own code, then send a summary to Wallach explaining why the code was probably fine. Wallach said he was considering the proposal seriously, noting that the time savings would be substantial.