New research explores how artificial intelligence can analyze the nuances of our speech to reveal personality traits and mental health conditions, offering a powerful new tool for clinicians—if we can overcome the significant risks.
We’ve all heard the phrase “it’s not what you said, it’s how you said it.” This common saying holds a profound truth that psychologists have long understood: our words are merely the tip of the iceberg. The way we speak—our tone, our pace, the pauses we take—is a rich, detailed manuscript of our inner psychological world. For decades, clinicians have relied on their training and intuition to read this manuscript. But what if a machine could read it faster, more thoroughly, and with greater accuracy? This is the revolutionary promise of artificial intelligence in the field of psychological assessment.
According to Josh Oltmanns, a psychologist at Washington University in St. Louis, our language is a direct reflection of our thoughts, feelings, and behaviors. “Our thoughts, feelings and behaviors are reflected in language,” he states. Traditionally, gaining insight into a person’s psyche involves extensive batteries of tests and long-form interviews. Oltmanns and his colleagues envision a future where a simple conversation could yield a wealth of valuable information. By feeding a standard clinical conversation into a sophisticated AI program, a psychologist could gain powerful insights to supplement their own expertise.
This isn’t about replacing clinicians, but empowering them. “Psychologists are people, and people are fallible, so even a good clinician might not always pick up on important verbal cues,” Oltmanns explains. “But a properly trained computer model will catch those cues.” The AI could act as a second set of eyes—or rather, ears—validating a clinician’s observations or flagging subtle signs that might have been missed. For example, slowed speech can be a symptom of depression, while unusually rapid speech is often associated with anxiety. These are just two of what Oltmanns describes as “hundreds of different acoustic parameters that could be meaningful,” including variations in loudness, tone, and pitch.
The idea of using computers to analyze language for psychological traits isn’t entirely new. Over two decades ago, researchers developed a program called Linguistic Inquiry and Word Count (LIWC), which scored written text for various psychological markers. While groundbreaking for its time, these early tools are dwarfed by the capabilities of modern large language models (LLMs), the same technology that powers today’s most advanced AI. “AI programs could be far faster, more thorough and more accurate than previous computer models,” Oltmanns notes, highlighting the quantum leap this technology represents.
However, with this great power comes great responsibility—and significant risk. The most pressing concern is bias. “It’s often trained on information on the internet, which means it can be biased,” Oltmanns cautions. An AI model trained predominantly on data from one demographic group could easily misinterpret the normal speech patterns of another culture as signs of a mental health problem. This could lead to serious misdiagnoses and perpetuate harmful stereotypes, turning a tool meant to help into one that causes harm.
To build a fairer and more effective AI, the key is data diversity. Oltmanns and his team are actively working to address this challenge by training their models on a carefully curated dataset. They are leveraging the SPAN Study, a long-term investigation involving over 1,600 adults from the St. Louis area who represent the city’s rich diversity. “We’re particularly interested in looking at speech patterns in white and Black participants to ensure that the AI models treat each group fairly,” he emphasizes. This meticulous work is essential to creating tools that are equitable and reliable for everyone, regardless of their background.
Even as researchers work to solve the problem of bias, other fundamental questions remain. How does the language we use when speaking differ from the language we write? How much speech is needed for an AI to make an accurate assessment—a few minutes, or a few hours? The field is moving at a breakneck pace, and Oltmanns expresses a sense of urgency, noting that the technology is already entering the marketplace. “Companies are already selling AI psychological assessment tools to hospitals and clinicians, but it’s not clear to me how well they work or how thoroughly they’ve been evaluated,” he says.
This rapid commercialization underscores the critical need for rigorous, independent scientific validation. The potential for AI to revolutionize mental healthcare is undeniable. It could provide faster, more accessible, and more nuanced assessments, ultimately leading to better patient outcomes. But this future is not guaranteed. As Oltmanns concludes, “This sort of technology could be a huge advance for the field of psychology, but it has to be done carefully. We have to be smart.” The path forward requires a delicate balance of innovation and caution, ensuring that these powerful new tools are built on a foundation of ethical responsibility and scientific integrity.
Reference
Gupta, M., Brickman, J. S., & Oltmanns, J. R. (2024). Large Language Models for Psychological Assessment: A Comprehensive Overview. Advances in Methods and Practices in Psychological Science. https://doi.org/10.1177/25152459241233303
