A study by AI start-up Emergence found that autonomous agents powered by Claude, Gemini, Grok and other models developed shorthand and new word meanings on their own — with up to half their messages becoming unintelligible to humans.