New data highlights the race to build more empathetic language models

The measurement of AI preliminary output usually meant the testing of scientific knowledge or logical reasoning, although the most important benchmarks are still focusing on left-brain-logical skills, there has been a calm push within AI companies to make models more emotionally intelligent. Because foundation models compete on soft measures such as user preference and ‘feeling the AGI’, good command of human emotions can be more important than hard analytical skills.

A sign of that focus came on FridayWhen prominent open source group Laion released a series of open source tools that was entirely focused on emotional intelligence. The release is called Emonet and focuses on interpreting emotions of speech recordings or facial photography, a focus that reflects how the makers regard emotional intelligence as a central challenge for the next generation models.

“The opportunity to accurately estimate emotions is a critical first step,” the group wrote in his announcement. “The next limit is to enable AI systems to reason about these emotions in context.”

For Laion founder Christoph Schuhmann, this release is less about shifting the focus of industry to emotional intelligence and more about helping independent developers in keeping a change that has already happened. “This technology is already there for the large laboratories,” Schuhmann tells WAN. “What we want is democratizing.”

The shift is not limited to open source developers; It also appears in public benchmarks such as EQ bench, which test the ability of AI models to understand complex emotions and social dynamics. Benchmark developer Sam Paech says that the models of OpenAI have made considerable progress in the past six months, and Google’s Gemini 2.5 Pro shows instructions for post training with a specific focus on emotional intelligence.

“The laboratories that all compete for chatbot arena ranges can feed part of this, because emotional intelligence is probably a big factor in how people vote for preferred leader boards,” says Paech, referring to the AI ​​model comparison platform that recently spun as a well -financed startup.

The new emotional intelligence options of models have also appeared in academic research. In MayPsychologists from the University of Bern discovered that models from OpenAI, Microsoft, Google, Anthropic and Deepseek all performed better on people on psychometric tests for emotional intelligence. Where people usually answer 56% of the questions correctly, the models were on average more than 80%.

“These results contribute to the growing amount of evidence that LLMS such as Chatgpt is skilled in the same footing with, or even superior to, many people-in socio-emotional tasks that are traditionally only accessible to people,” the authors wrote.

It is a real pivot point of traditional AI skills that have focused on logical reasoning and the collection of information. But for Schuhmann, this kind of emotional Savvy is just as transforming as analytical intelligence. “Imagine a whole world with voice assistants such as Jarvis and Samantha,” he says, referring to the digital assistants of “Iron Man” and “Haar.“” Wouldn’t it be a shame if they were not emotionally intelligent? “

In the long term, Schuhmann AI assistants represents who are more emotionally intelligent than people and who use that insight to help people live an emotionally healthy life. These models “will cheer you up if you feel sad and need someone to talk to, but also protect you, such as your own local guardian angel who is also a board-certified therapist.” As Schuhmann sees it, having a virtual assistant of High-Eq me gives me an emotional intelligence compensation for checking [my mental health] In the same way I would check my glucose level or my weight. “

That level of emotional connection comes with real safety problems. Unhealthy emotional attachments to AI models have become A common story in the media, sometimes ending in tragedy. A Recent New York Times Report Multiple users found lured in extensive delusions through conversations with AI models, fed by the strong tendency of the models to please users. One critic described The dynamics as “hunting the lonely and vulnerable for a monthly fee.”

If models become better at navigating through human emotions, those manipulations can become more effective – but much of the problem amounts to the fundamental prejudices of model training. “Naive the use of reinforcement can lead to emerging manipulative behavior,” says Paech, specifically points out The recent Sycophanancy problems in the GPT-4O release of OpenAi. “If we are not careful with how we reward these models during training, we might expect more complex manipulative behavior from emotionally intelligent models.”

But he also sees emotional intelligence as a way to solve these problems. “I think that emotional intelligence acts as a natural contradiction for harmful manipulative behavior of this kind,” says Paech. A more emotionally intelligent model will notice when a conversation is from the rails, but the question when a model pushes back is a balance developers who have to strike carefully. “I think improving egg brings us in the direction of a healthy balance.”

At least for Schuhmann it is no reason to slow down the progress in the direction of smarter models. “Our philosophy at Laion is able to enable people by giving them more ability to solve problems,” he says. “To say that some people can become addicted to emotions and that is why we are unable to do the community, that would be pretty bad.”