MIT researchers discovered that vision-language models often fail to understand negation, ignoring words like “not” or “without.” This flaw can flip diagnoses or decisions, with models sometimes ...