Why do my AI-generated tests pass on the bugs they were meant to catch?
κΈ°ν
AI coding assistants generate unit tests by reading the implementation, so assertions reflect what the code currently does rather than what it should do. A test written against a buggy function learns the bug as a fixture, not a violation. Developers shipping with 90% AI-generated coverage see mutation scores stay flat and real regressions slip through unchanged. The core issue is the oracle problem: there is no ground truth for correct behavior unless you supply it, and AI assistants have no access to the requirements or intent behind the code. Research from July 2025 confirms that LLM-generated tests frequently pass on the exact buggy code they were written against, failing only when the bug is fixed.
μ μ€μνκ°
A test suite that passes on broken code is worse than no test suite, because it turns the primary quality signal into a false guarantee.
κΈ°ν νκ° λ°©μ
κΈ°ν μ μλ μΈ‘μ κ°μ΄ μλ μ μ£Όκ΄μ νκ°μ λλ€. μΌλ§λ λΆνΈνμ§, μΌλ§λ μμ£Ό λ°μνλμ§, νμ¬ ν΄κ²°μ± μ΄ μΌλ§λ λΆμ‘±νμ§λ₯Ό λ°μν©λλ€. μ μκ° λμμλ‘ λ§λ€ κ°μΉκ° λ λλ€κ³ μκ°ν©λλ€.
λ°μνμ λ μΌλ§λ ν° λΆνΈμ μ΄λνλμ§.
μ€μ λ‘ μΌλ§λ μμ£Ό μ νκ² λλμ§.
νμ¬ μ΄λ₯Ό ν΄κ²°ν λ§ν λκ΅¬κ° μΌλ§λ λΆμ‘±νμ§.
ν΄κ²°ν κ°μΉ μλ λ λ§μ λ¬Έμ λ€
μ°λ¦¬κ° κ°μ₯ λ§μ΄ μμ‘΄νλ μννΈμ¨μ΄κ° μ κ°μ₯ μ¬μ©νκΈ° λΆνΈν κΉ?
Techλ΄κ° μμ±ν λ°μ΄ν°λ₯Ό μ λλ μ ν μμ νμ§ λͺ»ν κΉμ?
Techλ΄ λ°μ΄ν°κ° μ€μ λ‘ μμ λμμμ μ¦λͺ νλ μμμ¦μ μ λ°μ μ μλκ°?
Techμ€ν μ€μΈ κ²μ΄ SBOMμ μ μΈλ λ΄μ©κ³Ό μΌμΉνλμ§ μ μ μλ μ΄μ λ 무μμΈκ°?
TechC2PA μΆμ² 체μΈμ μ μ½ν μΈ κ° μμ λ―Έλμ΄μ μ¬λΌκ°λ μκ° λμ΄μ§λκ°?
Techλ΄ κ²μλ¬Όμ μ΄μν μ μλλ° νλ«νΌμ΄ λλ₯Ό κ΄μ°°νλ©° ꡬμΆν νλ‘νμ μ μ΄μν μ μλκ°?