AI language models can mimic human emotional states, opening new doors for mental health research, according to a new study.
Mental health conditions are increasing worldwide and are expected to affect 1.2 billion people by 2050. In this scenario, scientists and researchers are trying to better understand them, work to prevent them and develop new treatment tools.
ADVERTISEMENT
ADVERTISEMENT
Unlike medication-based treatments, talk therapies targeting mental health conditions are harder to develop, as neither human trials nor animal models can fully replicate the complexity of the conditions being studied — raising both practical and ethical obstacles in the process.
Now, a research team at Dresden University of Technology in Germany conducted a study to explore whether large language models (LLMs) can be used as tools for modelling mental health disorders in humans.
“Our results show that large language models can reproduce patterns of human affective and cognitive processes under controlled conditions,” said Dr Magdalena Wekenborg, head of the PsychoDigital Research group at TU Dresden.
“We can use these models as tools to better understand underlying mechanisms and to explore new approaches — for example in talk-based psychotherapy.”
Can LLMs replicate human emotions?
Although some mental health conditions have been modelled in mice and other organisms, the researchers noted that these approaches fall short of capturing the complexity and subjectivity of human behaviour.
They added that LLMs have emerged as powerful computational systems that approximate aspects of human intellectual performance.
“In many unexpected areas, such as persuasion, emotional understanding, and reasoning, these models are on par with human capabilities,” the authors wrote.
In their study, the research team prompted LLMs to mimic fear, anxiety, anger, disgust, sadness, worry and stress, and using them as experimental models to study aspects of mental disorders.
Then, they tested if once the LLMs reached those states, they could be reversed through different regulation strategies.
Finally, they assessed whether inducing a certain emotion in an AI model would cause it to make the same kinds of mistakes that humans do when feeling the same emotion.
They found that while the models do not have mental states in the way humans do, they can still mimic some ways of thinking through how they process language.
This makes it possible to run certain experimental tests on them that would be impossible or unethical to conduct in humans or animals.
According to Jakob N. Kather, from TU Dresden, one advantage of experiments with LLMs is that they can be repeated in identical conditions as often as needed, with the ability to systematically vary them.
“This enables new, data-driven experiments in psychological and biomedical research that were previously not possible,” he said
