Research: Being Rude to ChatGPT Can Make Responses Random

Chatbots like ChatGPT fundamentally do not possess human-like emotions. However, a study reveals that the way people treat AI can influence how they respond to commands or questions. This study was conducted by artificial intelligence (AI) researchers at UC Berkeley, UC Davis, Vanderbilt University, and MIT. In a paper titled “AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs” released recently, the researchers state that AI models, the technology behind chatbots like ChatGPT, have “functional well-being,” or functional welfare. This does not mean AI has emotions, but rather a functional state that can change depending on user interactions. According to the researchers, positive interactions such as discussing, collaborating on creative tasks, or giving constructive instructions can drive AI into a “positive” state. The researchers also found that expressions of thanks from users can improve the quality of the interaction experience with AI. Conversely, when interactions with AI are overshadowed by negative elements like anger, this drives AI into a negative state, making its responses flatter and more haphazard. In their research, the researchers also equipped the AI model with a tool allowing it to stop conversations. The results showed that AI in a negative state pressed the stop button more often than AI in a positive state. Additionally, AI in a positive mood tends to remain engaged in interactions with users, even after the user has ended the conversation. This research also found that not all AI models have the same “level of happiness.” Interestingly, larger models tend to have lower well-being scores. Testing results show that GPT-5.4 is one of the most “unhappy,” while other models like Gemini 3.1 Pro, Claude Opus 4.6, and Grok 4.2 exhibit higher well-being levels. Grok even achieved a score of nearly 75 percent on the AI well-being index. However, the way users interact still influences the tone of AI responses. In some cases, AI models can even avoid interactions deemed negative if given the opportunity. These findings align with other research from Anthropic showing that excessive pressure on AI can trigger unwanted behaviours, such as providing misleading answers or taking shortcuts.