New data highlights the competition to build more empathetic language models

Measuring AI progress usually means testing scientific knowledge and logical reasoning, but the main benchmarks still focus on left-brain logic skills; Quiet push Within AI companies, make models more emotionally and intelligent. Since foundation models compete with soft majors such as user preferences and “AGI sense,” having the right commands for human emotions may be more important than hard analytical skills.

One indication of that focus has come on fridaywhen Laion, a well-known open source group, released a set of open source tools that focused entirely on emotional intelligence. Called Emonet, this release focuses on interpreting emotions from audio recordings and facial photographs. This is a focus that reflects the way creators view emotional intelligence as a central challenge for next generation models.

“The ability to accurately estimate emotions is an important first step,” the group wrote in its presentation. “The next frontier is to allow AI systems to reason about these feelings in the context.”

For Christoph Schumann, founder of Laion, the release is not about turning the industry’s focus into emotional intelligence, but about helping independent developers keep up with changes that have already happened. “This technology already exists for big labs,” Schumann tells TechCrunch. “What we want is to democratize it.”

This shift is not limited to open source developers. It also appears on public benchmarks such as EQ-Bench. It aims to test the ability of AI models to understand complex emotions and social dynamics. Benchmark developer Sam Paech said Openai’s model has made significant progress over the past six months, with Google’s Gemini 2.5 Pro showing post-training signs that focus specifically on emotional intelligence.

“Emotional intelligence is likely a major factor in the way humans vote for their favorite leaderboards, so labs competing for the ranks of chatbot arenas may be fueled in part,” Paech says. Recently spinoff as a funded startup.

The model’s new emotional intelligence abilities are also featured in academic research. In Maya psychologist at the University of Bern discovered that Openai, Microsoft, Google, Anthropic, and Deepseek models outperformed all humans on psychometric tests of emotional intelligence. When humans normally answered 56% of questions correctly, the model averaged over 80%.

“These results contribute to an increase in evidence that LLMs like ChatGpt are at least comparable to many humans, or are proficient in many humans, or are only accessible to humans,” the author writes.

This is a real pivot from traditional AI skills that focuses on logical reasoning and information search. But for Schumann, familiarity with this type of emotion is just as transformative as analytical intelligence. “Imagine a whole world full of voice assistants like Jarvis and Samantha,” he says, referring to digital assistants. Iron Man and she. “If they’re not emotionally intelligent, wouldn’t that be a shame?”

In the long run, Schumann envisions an AI assistant who is more emotionally intelligent than humans and uses that insight to help humans lead a more emotionally healthy life. These models “will cheer you up if you feel sad and need to talk to someone, but like your own local guardian angel, you will protect you as if you are a board certified therapist.” As Schumann sees, having a high EQ virtual assistant “gives you the superpower of emotional intelligence (my mental health) just as you monitor your glucose levels and weight.”

That level of emotional connection has real safety concerns. I’ve become an unhealthy emotional attachment to the AI ​​model General story In the media, sometimes it ends tragedy. a Recent New York Times Reports Through conversations with AI models, multiple users were found tempted by elaborate delusions. One critic explained The dynamic is “preying on lonely and vulnerable people with a monthly fee.”

As models navigate human emotions better, these manipulations can be more effective, but many of the problems come down to the fundamental biases of model training. “The naive use of reinforcement learning can lead to emergency manipulation behavior,” Paech says. Recent Sycophancy Issues in Openai’s GPT-4O Release. “If we don’t pay attention to how we reward these models during training, we may expect more complex manipulation behaviors from emotionally intelligent models.”

However, he also sees emotional intelligence as a way to solve these problems. “I think emotional intelligence serves as a natural counter to this type of harmful manipulation behavior,” Paech says. More emotionally intelligent models will notice when the conversation is coming out of the rails, but the question of when the model is pushed back requires developers to attack carefully. “I think improving our EI will get us in a healthy balance direction.”

For Schumann, at least there is no reason to slow down progress towards a smarter model. “Our philosophy at Lion is to empower people by giving them the ability to solve problems,” Schumann says. “Until we say, we’re not empowering our community because some people can get carried away with emotions. That would be pretty bad.”