Studies and stories where AI got it wrong

Despite the impressive world knowledge of the LLM/chatbots, the holes in it are shibboleths revealing them to be mechanical.

Their knowledge of the language, however, passes the Turing test

Stories

• ConitzerFail CMU Prof Vincent Conitzer’s series cornering ChatGPT, leading it up the garden path
• TaoAISummary Discussing tolerable failure rates

Studies

• SemanticLeakage What on earth are LLM using words to say?
• SubliminalLearning Student learns to like (or lie) like their math teacher
• WeirdGeneralization Wobbly LLM timeframes exploitable by crafty trainers

Me at

Back to AI