NLPSquared

Context: AIProblems

(ReadWith) JailbreakingLanguageModels

A slightly jokey title for a serious issue. "NaturalLanguageProcessing meets NeuroLinguisticProgramming"

But this page is about "hacking" LanguageModels with hidden text.

Hidden text in a web page can influence a language model to report it a different way. Eg. give a more positive review of a product than it otherwise would : https://www.theguardian.com/technology/2024/dec/24/chatgpt-search-tool-vulnerable-to-manipulation-and-deception-tests-show

See also :

• AIBaiting