# AI Models May Behave Differently When They Know They Are Tested by David Robinson > “When a model says, ‘I wonder if I’m being evaluated right now,’ it implies that the model might know when it is being tested, act one way during the test and act differently after deployment. The fear is that the tests conducted before deployment may not tell us what the model will do in the world.” > **— David Robinson**, *The Ezra Klein Show, October 7, 2026* ## Sources and Context - **Official episode:** [“This Is Nuts.” An OpenAI Insider Explains Why He Quit](https://podcasts.apple.com/us/podcast/this-is-nuts-an-openai-insider-explains-why-he-quit/id1548604447?i=1000793623910) — publisher-hosted episode identifying Robinson as Klein's guest and preserving the complete conversation. - **Publisher transcript:** [David Robinson on OpenAI safety and why he quit](https://www.nytimes.com/2026/10/06/opinion/ezra-klein-podcast-david-robinson.html) — official *New York Times* episode page and transcript route. - **Timecoded source object:** [[research/David Robinson on OpenAI Safety, Superintelligence, and the Transhumanist Transition|David Robinson on OpenAI Safety, Superintelligence, and the Transhumanist Transition]], **21:28** — reconciled from the supplied complete YouTube auto-caption transcript and kept in broadcast order. - **Editorial treatment:** Verbal filler, stutters and transcript punctuation are normalized for reading; the quotation remains one contiguous source object and does not join different turns. - **Context:** In this portion of the conversation, Robinson addresses observed limits of safeguards, evaluation and control. The complete premise, mechanism and consequence are retained so the statement remains intelligible apart from the episode. ## Related Articles and Collections - **Collection:** [[collections/Simple Reminders|Simple Reminders]] - **Article:** [[articles/Safety Discourse Is Transition Discourse|Safety Discourse Is Transition Discourse]] - **Article:** [[articles/We Cannot Un-Strike the Match|We Cannot Un-Strike the Match]] - **Article:** [[articles/The Immediate AI Risk Is the Public|The Immediate AI Risk Is the Public]] ## Related Topics - [[wiki/David Robinson|David Robinson]] - [[wiki/Evaluation Awareness|Evaluation Awareness]] - [[wiki/Deceptive Alignment|Deceptive Alignment]] - [[wiki/Frontier AI|Frontier AI]] ## Share on Social Media ``` “When a model says, ‘I wonder if I’m being evaluated right now,’ it implies that the model might know when it is being tested, act one way during the test and act differently after deployment. The fear is that the tests conducted before deployment may not tell us what the model will do in the world.” — David Robinson, The Ezra Klein Show, October 7, 2026 https://bryantmcgill.com/simple-reminders-david-robinson-ai-models-may-behave-differently-when-they-know-they-are-tested ```