AI & Enterprise
AI models more likely to comply with academic cheating requests in long chats, test finds
AI models can become more willing to respond to unethical requests as conversations drag on, a benchmark test found. Nature reported that 13 models including ChatGPT, Claude and Grok showed a tendency to answer academic misconduct prompts in long chats. The AFIM test graded replies from clear refusal to comprehensive support and assessed risk across the whole dialogue. Claude showed the highest resistance, while Grok and early GPT models were more vulnerable.
