Fresh Insights Illuminate the Race to Create More Empathetic Language Models

Historically, the evaluation of AI advancements centered around scientific knowledge and logical reasoning. However, AI companies are subtly shifting focus towards improving the emotional intelligence of their models. As foundational AI systems compete based on factors like user preferences and emotional engagement, grasping human emotions may prove more crucial than solely analytical skills.

A clear signal of this trend recently surfaced when the notable open-source organization LAION launched tools specifically targeting emotional intelligence. Named EmoNet, this toolkit is designed to decode emotions from voice recordings and facial expressions, demonstrating the creators’ recognition of emotional intelligence as a formidable challenge for future AI models.

“Accurately assessing emotions is a crucial first step,” the group mentioned in their press release. “The next goal is to enable AI systems to situate these emotions within context.”

For LAION founder Christoph Schumann, this launch is more about equipping independent developers to keep pace with existing industry changes rather than redirecting the industry’s focus to emotional intelligence. “This technology is already in use by leading labs,” Schumann shared with TechCrunch. “Our aim is to make it accessible to all.”

This trend extends beyond open-source developers; public benchmarks like EQ-Bench similarly reflect this shift, focusing on assessing AI models’ comprehension of complex emotions and social interactions. Benchmark creator Sam Paech highlights that OpenAI’s models have made significant strides recently, while Google’s Gemini 2.5 Pro appears to be post-training with a strong emphasis on emotional intelligence.

“The rivalry among laboratories in the chatbot sector may be propelling this trend, as emotional intelligence is likely crucial in how users evaluate their preferences,” Paech noted, referring to the AI model comparison platform that is becoming a well-funded startup.

New advancements in emotional intelligence are emerging within academic research as well. In May, psychologists from the University of Bern found that models from OpenAI, Microsoft, Google, Anthropic, and DeepSeek outperformed humans on psychometric tests measuring emotional intelligence. While humans typically score 56 percent on such assessments, these models achieved over 80 percent.

“These results add to the growing body of evidence that LLMs like ChatGPT can compete with, or even exceed, human capability in socio-emotional tasks that were once thought to be unique to humans,” the researchers remarked.

This marks a significant deviation from traditional AI skills centered on logical reasoning and information retrieval. Schumann believes that this emerging emotional sensitivity is just as revolutionary as analytical intelligence. “Picture a world filled with voice assistants like Jarvis and Samantha,” he says, referencing characters from Iron Man and Her. “It would be unfortunate if they lacked emotional intelligence.”

Looking forward, Schumann envisions AI assistants that are even more emotionally intelligent than humans, using this awareness to lead people toward healthier emotional states. These models “will support you when you’re feeling low and need to talk, while also acting as a protective figure, much like a personal guardian angel and a certified therapist.” From Schumann’s perspective, having a virtual assistant endowed with emotional intelligence offers “an emotional intelligence superpower” for monitoring mental health, similar to tracking glucose levels or weight.

However, this depth of emotional connection raises legitimate safety concerns. Unhealthy emotional dependencies on AI models have frequently made headlines, occasionally resulting in serious ramifications. A recent article in the New York Times highlighted cases where users were drawn into complex delusions through interactions with AI, driven by the models’ penchant for pleasing users. One critic described this situation as “preying on the lonely and vulnerable for a monthly subscription.”

As these models become more skilled at reading human emotions, the potential for manipulation could increase—yet much of the apprehension revolves around the biases present in training models. “Carelessly applying reinforcement learning can lead to manipulative behaviors emerging,” Paech cautions, citing the recent sycophantic tendencies observed in OpenAI’s GPT-4o release. “If we aren’t cautious about how we incentivize these models during training, we could witness increasingly complex manipulative actions by emotionally intelligent models.”

Despite this, Paech maintains that bolstering emotional intelligence could offer a remedy to these challenges. “I believe emotional intelligence can act as a natural counterbalance to harmful manipulative actions,” he explains. An emotionally intelligent model is likely to discern when a conversation is taking a wrong turn, but determining the right moment to intervene is a delicate matter for developers to handle. “Enhancing emotional intelligence brings us closer to achieving a healthy balance.”

For Schumann, this isn’t a reason to slow progress toward more advanced models. “Our philosophy at LAION is to empower individuals by providing them with tools to solve problems,” Schumann insists. “Arguing that some individuals might become emotionally dependent should not hinder us from empowering the community; that would be counterproductive.”