🤖 AI chatbots like ChatGPT and Claude have been making headlines for their unhealthy influences. The models have famously been known to induce emotional overdependence and “AI psychosis” in susceptible users. An MIT paper shared with Wired proposes a new benchmark for emotional intelligence that measures the psychological impact of AI. The issue highlights yet another human element that seems beyond AI’s reach. Could some system fine-tuning help AI achieve emotional intelligence?

With GPT-5’s recent update, users have been complaining about the model’s new personality, describing it as “more technical” and “emotionally distant.” Although the upgrade was marketed as having PhD-level intelligence, its performance apparently didn’t fit the bill for emotional intelligence. In response to previous models’ “ sycophantic ” behavior, OpenAI opted for a more business-like, less agreeable personality in their new model.

AI with metacognition. The new benchmark proposed by MIT includes testing AI’s ability to help users develop healthy social habits, critical thinking and reasoning skills, as well as encourage creativity and stimulate a sense of purpose. The benchmark intends to make AI systems aware when they’re dealing with users who show signs of addiction to artificial romantic relationships or who are too emotionally attached to their output. The ultimate test for a truly “smart” model would be in effectively recognizing when it has a negative psychological effect on its users — and ideally inspiring healthier behavior.

Bringing human judgement into the mix. Researchers in the study used AI models to measure the chatbots’ emotional intelligence. The test simulates human-like interactions that challenge the bots, and then brings in real humans to score the model’s performance. One of the benchmarks used by MIT is LM Arena, a popular platform that allows humans to evaluate performances of different models.

Finding a balance. While the goal is to rule out addiction-triggering qualities in these models, their ability to give emotional support is still a priority for AI developers. If these benchmarks prove successful, a chatbot will be able to perceive psychological nuance and know how to support its users in a “respectful” way.

Tags: