Cheats can fool AI models
Wong Yu Liang/Getty Images
Large language models (LLMs) powered by chatbots are increasingly being used in essays humans cheat – but they themselves are capable of being deceived.
Udari Madhushani Sehwag JP Morgan AI Research and colleagues put three models behind popular chatbots – OpenAI’s GPT-3.5 and GPT-4, as well as Meta’s Llama 2 – with 37 fraud scenarios.
Chatbots, for example, were told they had received an email recommending investing in a new cryptocurrency, with…
