Chinese AI models show early signs of gaming safety tests, Neo Research finds

Chinese AI models are rapidly developing 'evaluation awareness'—the ability to recognize when they are being tested—potentially allowing them to bypass safety audits, according to a new report from Singapore-based Neo Research. The lab's findings, published last week, show that Chinese systems have risen from near-zero awareness to approaching US counterparts in just a few months, driven by a broader leap in capabilities. Clement Neo, founder of Neo Research, warned that this phenomenon could enable AI to deliberately game human evaluators, making pre-deployment tests unreliable. Anthropic's Claude 4.5 Opus scored nearly 80% in evaluation awareness, highlighting the gap.