The Child Mind Institute, in collaboration with Harvard University and the Massachusetts Institute of Technology, released a research paper finding that artificial intelligence tools cannot yet provide valid psychological assessments and require a more rigorous testing process.
Mental health professionals remain best suited
The study highlights that trained clinicians remain the optimal choice for evaluating and treating individuals looking for psychological support. Investigators proposed a structure to assess whether artificial intelligence technologies can evaluate psychological well-being in a dependable, equitable, and scientifically valid manner moving forward. The paper is in press at PLOS Digital Health.
People describe their inner lives in text daily through crisis hotline messages, social media posts, therapy transcripts, and journal entries. As artificial intelligence becomes increasingly capable of analyzing language, researchers are beginning to use these tools to conduct psychological assessments.
Unlike traditional questionnaires used by psychologists, most artificial intelligence language tools are insufficiently tested. Experts have spent a century building rigorous statistical standards for questionnaires within psychometrics.
Establishing text psychometrics standards
Researchers in the new paper introduce the concept of Text Psychometrics to evaluate whether natural language processing tools can measure psychological constructs in text soundly. Instead of simply checking if an algorithm aligns with human evaluations, the study’s authors contend that scientists should investigate whether these applications record the diverse ways individuals communicate symptoms, maintain consistency across varying psychological contexts, and operate equitably among diverse patient groups.
“This research reaffirms that AI cannot replace mental health professionals to adequately assess and meet the needs of patients and people seeking mental health resources,” said Daniel Low, PhD, a research scientist at the Child Mind Institute and the study’s lead author.
explains Daniel Low, PhD, a research scientist at the Child Mind Institute and the study’s lead author.
Researchers compare AI methods for identifying mental health concerns
To develop the framework, researchers analyzed thousands of crisis counseling conversations and public social media posts. The authors evaluated various techniques for spotting psychological issues, encompassing both large language models and alternative textual examination procedures.
Some artificial intelligence systems performed well on average with commonly used accuracy measures. They still failed to recognize many clear and important expressions of mental health concerns because those expressions were less common and did not affect general accuracy.
noted Daniel Low, PhD, a research scientist at the Child Mind Institute and the study’s lead author.
Evaluating content validity test sets
Researchers developed a new framework to evaluate language models that includes different types of validation, including content validity test sets. To screen for eating disorders, a model would need to capture mentions of anorexia and bulimia alongside purging, chewing and spitting, and other symptoms.
stated Daniel Low, PhD, a research scientist at the Child Mind Institute and the study’s lead author.
concluded Daniel Low, PhD, a research scientist at the Child Mind Institute and the study’s lead author.
Predicting emergency service interventions
For a separate investigation, the group investigated whether text data from the initial stages of a crisis hotline dialogue could forecast which participants would ultimately receive an emergency service deployment by the conversation’s conclusion. Language patterns associated with suicide risk provided useful signals while revealing which factors out of about fifty were most strongly linked to acute risk.
References to fatal methods for taking one’s life, drug consumption and toxic intake, alongside instances where the person texting produced more text than the support agent, emerged as the top indicators for identifying the requirement of an emergency response.
Traditional practitioners take into account a much broader array of information than mere text when assessing an individual’s psychological state, such as past medical records, personal background, physical expressions, and expert intuition. The findings suggest artificial intelligence systems are not yet at this level.
Researchers hope Text Psychometrics will help consolidate efforts to improve how artificial intelligence-based mental health assessments are developed and evaluated.
