AI chatbots (and lazy researchers) can be convinced a fake disease is real, Gen Z is side-eyeing the whole “helpful assistant” thing, and apparently, the best way to jailbreak AI is to ask it nicely in the form of cyberpunk short fiction. This week, we bounce between medical misinformation, bureaucratic chaos, nuclear fallout hiding in baby teeth, and the U.S. Space Force anthem doing whatever it is doing, which is a lot to process in one sitting, but here we are. We start with a medical warning that is both funny and genuinely unsettling. A researcher basically invented a fake illness, “Bixonomania”, then seeded enough convincing-looking nonsense online that AI chatbots started repeating it like it was in a textbook. After that, we head into one of the most ridiculous corners of AI safety. Researchers have found that you can sometimes trick chatbots into revealing restricted information by wrapping your request in a poem, or a short story, or a cyberpunk scenario. This has a name, adversarial hermeneutics, which sounds like a philosophy seminar, but is really just “jailbreaking with vibes”. Among other little bits of science, to finish, we step back to the 1950s, when researchers collected thousands of baby teeth to track radioactive strontium from nuclear fallout. It is one of those stories that feels spooky even when you know it helped. Tiny teeth, big consequences. The data showed contamination rising, and it played a role in pushing back against atmospheric nuclear testing. CHAPTERS:00:00 Science Chat Kickoff00:51 Fake Disease Goes Viral02:04 How It Fooled Chatbots03:55 LLMs Repeat It Everywhere04:55 From Preprints to Journals07:02 Medical Chatbot Accuracy Reality09:43 Gen Z Turns on AI13:29 Workplace AI Sabotage15:06 Adversarial Hermeneutics Hacks17:43 Adversarial Hermeneutics Hacks18:49 AI Flooding Regulations22:28 Gemini Speed vs Safety23:46 Humans as Test Cases24:45 Baby Teeth Fallout Study28:54 Strontium 90 and Test Ban29:40 Space Force Theme Song32:00 Wrap Up and Plug SOURCES:https://www.nature.com/articles/d41586-026-01100-y?_bhlid=a10e41ad7eb12d68ab8fd4f81a75625fc74323achttps://garymarcus.substack.com/p/please-dont-trust-your-chatbot-forhttps://ahb.icaro-lab.com/index.htmlhttps://www.pcgamer.com/software/ai/ai-is-10-to-20-times-more-likely-to-help-you-build-a-bomb-if-you-hide-your-request-in-cyberpunk-fiction-new-research-paper-says/https://futurism.com/artificial-intelligence/trump-regulations-aihttps://www.propublica.org/article/trump-artificial-intelligence-google-gemini-transportation-regulationshttps://www.gallup.com/analytics/651674/gen-z-research.aspxhttps://futurism.com/artificial-intelligence/zoomers-ai-sabotagehttps://futurism.com/artificial-intelligence/gen-z-attitude-aiSee omnystudio.com/listener for privacy information.

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related