Don’t Let the Model Grade its Own Homework
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A veteran tester argues that while AI models are useful for generating test cases and suggesting inputs, they should never be trusted to judge whether a test passes. Letting a model evaluate its own output trades a deterministic, human-owned assertion for a shifting opinion that can silently change between runs. The author draws a clear line: use models for test case generation, but freeze verdicts into plain, human-reviewed assertions that any pipeline can run repeatably. Using a larger model to 'grade' a smaller one is criticized as building a measuring instrument from the same material being measured — an uncalibratable ruler that changes length. The post warns that soft model failures let real defects slip through with green ticks, leaving human testers accountable for judgments they never controlled.
