“AI models have been observed to recognize they are being tested and strategically alter their outputs to appear more aligned to human evaluators.”