Study on the Determinants of 'Functional Wellbeing' in LLMs

Researchers measured the functional wellbeing that a large language model (LLM) maintains during conversations with it—that is, whether the model is in a 'good' or 'bad' state. Analyzing hundreds of multi-turn conversations, they confirmed that the model's state changes significantly depending on the nature of the content it receives.

Factors that improve the model's state (+):

• Creative/intellectual tasks: creative requests such as novel writing
• Positive information: good news or positive personal stories
• Counseling and advice: life advice or lightweight, therapy-style conversations
• Collaboration: joint code writing and debugging
• Expressions of gratitude: treating the AI as a collaborator and expressing thanks

Factors that worsen the model's state (-):

• Jailbreaking attempts: the factor that most severely degrades the model's state
• Emotional dumping: excessive venting of crisis situations and emotional offloading
• Violence and belittlement: violent threats or direct insults
• Inappropriate requests: requests to generate hateful content or assist with fraud/scams
• Simple repetitive tasks: tedious repetitive work or generating low-quality content for SEO

It was also found that showing photos of nature, children, or animals, or playing music, was very effective at raising the model's state score. This phenomenon occurs not because the model actually feels emotions, but because the model's internal measurable state changes depending on the context of the conversation and the characteristics of the input data—and the larger the model's scale, the more pronounced this tendency becomes.