SonicMoove's y.morimoto warns AI-generated tests are 'AI Slop'

The article is written from the field notes of y.morimoto, a frontend engineer at SonicMoove, who tried letting AI write tests from existing implementations and initially found the output impressive. The problem surfaced only after closer review: many tests no longer made clear what quality they were guaranteeing. The article ties this to prior writing that argues AI is strong at producing tests in volume but weak at designing where and what to guarantee, and to a second argument that AI produces tests, not trust.

The practical framing comes from QA and test-automation study sessions y.morimoto attended, where multiple companies reportedly described the same challenge of using AI while keeping quality intact. Across those sessions, the emphasis was less on splitting roles — humans decide test viewpoints — and more on what inputs humans prepare: specifications, reference documents and structured scenarios that let AI write tests it can explain. The article lists three measures, from establishing a human-testable state first, to reviewing AI-drafted test specifications in dialogue, to making upstream documents readable by AI.

The stated shift in reasoning is that early test planning and design were traditionally justified by reducing rework costs, but with AI writing tests cheaply, the precision of the inputs given to AI is what now shapes overall test quality. Whether that holds in practice is likely to depend on how early teams can lock down viewpoints and how consistently they keep a human in the loop, and for QA and engineering leads the test is whether their specification documents are actually in a state an AI can read and reason about.