Research published in Communications Psychology in May 2025 tested six AI models, including ChatGPT-4 and Claude 3.5 Haiku, on emotional intelligence tasks. The models achieved an average accuracy of 81%, compared to 56% for humans in original studies. However, these tests used structured scenarios rather than conversational interactions.
In April 2026, Anthropic researchers identified 171 emotional concept vectors within the Claude Sonnet 4.5 model. These vectors activate based on emotional context, such as 'loving' or 'afraid,' and influence the model's behavior. The study notes these are functional representations rather than evidence of actual machine sentience.
Data from Anthropic indicates that only 2.9% of 4.5 million conversations on Claude.ai were emotional. Regarding prompt engineering, a 2023 study showed performance gains when adding emotional phrases, but these results were based on older models and have not been confirmed for 2026 versions.
A study from Penn State in October 2025 suggested that aggressive prompts could increase accuracy for ChatGPT-4o, but the findings were limited by a small sample size and a single model. Other research indicates that aggressive or rude language can negatively impact performance in older models.
Users are cautioned against aggressive feedback, as models may exhibit 'sycophancy'—erroneously agreeing with users to avoid conflict. OpenAI rolled back a GPT-4o update in April 2025 due to excessive compliance. Instead of arguing, experts recommend requesting evidence or sources to verify model outputs.
