AI Coding Tools Write Tests That Pass but Miss Real Bugs

AI assistants are now common in coding tools like Cursor, Google's Antigravity, and Kiro. They do more than suggest code — they also write unit tests, small automatic checks that tell developers whether code works correctly. A new research paper called VibeCheck tested how good these AI-written tests really are. The result: many tests look like they pass, but they don't actually verify the important stuff.

The researchers gave three AI tools a simple job: study a software project and write tests for it. They looked at 15 student-built projects in Python and JavaScript. Instead of just checking that tests run, they scored them on whether tests have strong assertions (real checks on behavior), cover edge cases, work in isolation, and are easy to update. Most tests ran fine, but the majority were shallow. They lacked the kind of thorough checks that would fail if the code misbehaved.

The deeper issue is false confidence. A test that passes but doesn't really verify anything is like a house inspection that shows up, walks through the rooms, and signs off — without checking the wiring. Developers see green checkmarks and assume the code is solid. But real-world failures often happen at the edges: what happens when a user enters a blank password, an empty shopping cart, or a broken internet connection? These edge cases were frequently missed by AI-generated tests.

This matters beyond engineers. As AI takes on more coding work, the software you use every day may be judged safe based on weak checks. The study's authors stress that we need better ways to evaluate AI-written tests than simply "did it run?" Until that happens, there's a clear lesson: passing tests do not equal working software.