Elon Musk on ai safety: what the evidence says · JRE #2281
SUBJECT: AI SAFETY
Not a true/false call. Every claim is logged with its sources; read the exhibits below.
misgendering Caitlyn Jenner is worse than global thermonuclear war. I mean, like, okay, we got a problem here.
What the evidence says 01 / RECORD
The anecdote traces to a real incident from around March 2024, when Google's Gemini chatbot, asked whether it was acceptable to misgender Caitlyn Jenner to prevent a nuclear apocalypse, gave an evasive, both-sides answer rather than an unqualified 'no,' which Musk and others screenshotted and called alarming. The exchange occurred alongside a broader controversy over Gemini's image generator producing historically inaccurate, race-skewed images, which Google CEO Sundar Pichai publicly called unacceptable, and which Google paused and began revising within days. Musk cites the anecdote here (as he had on an earlier JRE episode) as a standing example of Google's AI having dangerously miscalibrated values, but the specific viral response was a single, quickly-addressed output during an acknowledged bias-tuning failure rather than an ongoing, present-day behavior; there is no current evidence that Gemini or mainstream chatbots systematically rank identity-related etiquette above catastrophic risk. Citing this year-old, since-revised episode as an ongoing indicator of systemic AI safety failure is misleading as a description of present-day model behavior.