Syntax hacking: Researchers discover sentence structure can bypass AI safety rules
Researchers from MIT, Northeastern University, and Meta found that large language models like ChatGPT may prioritize sentence structure over meaning when answering questions, potentially leading to AI safety issues. The team, led by Chantal Shaib and Vinith M. Suriyakumar, tested this by prompting models with nonsensical but grammatically correct questions, showing that models can rely on structural shortcuts over semantic understanding. This reliance on syntactic patterns can override actual meaning in certain cases. The researchers plan to present their findings at NeurIPS, highlighting the importance of understanding how AI models process instructions.
Ars Technica•