New data highlights the race to build more empathetic language models

The measurement of progress of AI has generally meant testing scientific knowledge or logical reasoning – but although the main references are still focusing on the logical skills of the left brain, there has been a silent Within AI companies to make the models more intelligent emotionally. As foundation models compete on soft measures such as user preferences and “feeling act”, having a good mastery of human emotions can be more important than hard analytical skills.

A sign of this orientation came FridayWhen the eminent open source group Laion published a series of open source tools has completely focused on emotional intelligence. Called Emonet, the version focuses on the interpretation of emotions from vocal recordings or facial photography, an orientation that reflects the way creators consider emotional intelligence as a central challenge for the next generation of models.

“The ability to estimate emotions with precision is an essential first step,” wrote the group in its ad. “The next border is to allow AI systems to reason on these emotions in context.”

For the founder of Laion, Christoph Schumann, this version consists less in moving the attention of the industry to emotional intelligence and rather to help independent developers to follow a change that has already occurred. “This technology is already there for large laboratories,” Schumann told Techcrunch. “What we want is democratizing it.”

Change is not limited to open source developers; It also appears in public references like Eq-Bench, which aims to test the ability of AI models to understand complex emotions and social dynamics. The reference developer, Sam Paech, says that OpenAi models have made significant progress in the past six months, and that Gemini 2.5 Pro from Google shows post-training indications with a specific accent on emotional intelligence.

“Laboratories in competition for the rows of a chatbot Arena can feed part of this, because emotional intelligence is probably an important factor in the way humans vote on preferences’ rankings,” explains Paech, referring to the comparison platform of the AI ​​model that Recently discouraged as a well -funded startup.

The new emotional intelligence capacities of the models have also been revealed in university research. In MayPsychologists at the University of Bern discovered that Openai, Microsoft, Google, Anthropic and Deepseek models have all surpassed human beings on psychometric tests for emotional intelligence. When humans generally answer 56% of questions properly, the models were on average more than 80%.

“These results contribute to all the evidence that the LLM as Chatgpt are competent – at least with equal, even superior to many humans – in socio -emotional tasks traditionally considered accessible only to humans,” wrote the authors.

It is a real pivot of the traditional IA skills, which focused on logical reasoning and information recovery. But for Schumann, this kind of emotional flavor is just as transformative as analytical intelligence. “Imagine an entire world full of vocal assistants like Jarvis and Samantha,” he said, referring to the digital assistants of Iron And Her. “Wouldn’t it be a shame if they weren’t emotionally intelligent?”

In the long term, Schumann envisages AI assistants who are more emotionally intelligent than humans and who use this insight to help humans live a healthier emotional life. These models “will rejoice if you feel sad and you need someone to talk to, but will also protect you, like your own local guardian angel which is also a therapist certified by the Board of Directors.” As Schumann sees it, having a virtual assistant at Haut Eq “gives me a superpowered emotional intelligence to monitor [my mental health] In the same way that I would monitor my glucose rate or my weight. »»

This level of emotional connection is accompanied by real security problems. Unhealthy emotional attachments to AI models have become A common story in the media, sometimes ending with tragedy. A Recent New York Times report I found several users who have been attracted to delusions made via conversations with AI models, powered by the high trend of models to please users. A criticism describe Dynamics like “attacking solitary and vulnerable for monthly costs”.

If the models improve to navigate human emotions, these manipulations could become more effective – but a large part of the problem comes down to the fundamental biases of model formation. “The naively use of strengthening learning can lead to emerging manipulative behavior,” says Paech, pointing specifically to Recent sycophance problems in the release of the OPENAI GPT-4O. “If we are not paying attention to how we reward these models during training, we could expect a more complex manipulation of emotionally intelligent models.”

But he also considers emotional intelligence as a way to solve these problems. “I think that emotional intelligence acts as a natural counter of the harmful manipulative behavior of this kind,” explains Paech. A more emotionally intelligent model will notice when a conversation is triggered by rails, but the question of when a model repels is a balance that the developers will have to strike carefully. “I think that improving the EI puts us in the sense of a healthy balance.”

For Schumann, at least it is not a reason to slow down progress towards smarter models. “Our philosophy in Laion is to allow people to give them more capacity to solve problems,” explains Schumann. “To say, some people could become dependent on emotions and therefore we do not empower the community, it would be bad.”