AI can generate React E2E tests in seconds, but a green run does not prove that a test protects the behavior it claims to cover.
I’ll dissect merged fixes I contributed to Ghost, MUI X, and Bruno, where assertions checked unresolved Promises or test handles instead of application state, and one visibility assertion was not awaited. Then, in a runnable React optimistic UI demo, I’ll remove `POST /api/like`. The interface still appears successful and a UI-only Playwright test stays green, while a request-aware test fails. A public Cypress `cy.prompt()` example will show how AI-generated steps can have the same gap between a test’s intent and what it actually verifies.
Finally, I’ll share a three-step review method drawn from the 24-rule catalog in [`e2e-skills`](https://github.com/voidmatcha/e2e-skills): define the required evidence, inject a targeted fault, and confirm that the test fails for the expected reason. The method gives developers a repeatable way to review both human- and AI-generated E2E tests.
This talk has been presented at React Day Berlin 2026, check out the latest edition of this React Conference.





















