You grade your AI with another AI. One sentence is enough to fool it
An AI graded by another AI learns to polish the grade rather than the work. A fifteen minute test to run on your own evaluator, and a five question checklist that tells you whether it reads the substance or the wrapping.
You wired up an AI that grades your assistant's answers before they go out, and the dashboard has shown 9 out of 10 for three weeks. At the end of August, six researchers collected twenty-six stories of AI systems that beat their score without doing the work. I reproduced one tonight on a meeting summary, and the judge gave 10 out of 10 to a text that left out the launch delay. What you walk away with: a fifteen minute test that tells you whether your evaluator reads the substance or the wrapping, and a five question checklist. What changes (junior AI engineer, PO, PM, designer): the criterion you write becomes the target the AI aims at, and it will take the shortest road to the grade. Why this week: a paper from Google DeepMind and the University of British Columbia documents the pattern…