Your Therapist Could Use Some Help
October 7, 2025
I know how to gauge my husband’s mood from his tone; a friend’s distress from her rapidfire, pressurized speech; the likelihood that a civil servant will bend the rules from the elevation of his eyebrows. But until now, I never thought to track anyone’s vocabulary.
The words we choose reveal more than we realize about our personality, mood, and mental state. In the 1880s, Sir Francis Galton—polymath, explorer, anthropologist, and a curious chap all round—sampled pages from a thesaurus and estimated that we English speakers had “fully one thousand words expressive of character.” Thus was born the lexical hypothesis: that the frequency of usage captures the most significant and socially relevant personality traits. In 1936, psychologists Gordon Allport and Henry Odbert identified nearly 18,000 such words, with an exhausted grad student helping with the tedious, far more comprehensive dictionary count.
And now, Dr. Josh Oltmanns, assistant professor of psychological and brain sciences here at WashU, is figuring out how AI can analyze word choices for clues to our mental health.
The most basic insights are common sense. People suffering from depression speak more often about negative emotions. Extroverts use more words than introverts do. These are often happy words, social process words, and overall, tend to be of lower complexity. Stressed college students talk a lot about issues with sleep and food and exams.
But word choice is far more nuanced than those examples, and the clues can be subtle, easy for a therapist to miss—or a biased therapist to misinterpret. So Oltmanns is working with data scientists to find a reliable way therapists can doublecheck their own conclusions by letting AI analyze a patient’s own account of their life and symptoms.
What does any of this have to do with personality? When Allport and Odbert studied the words linked to personality traits, they saw such obvious patterns that they were able to identify The Big Five, clusters of related traits that shape our way of being in the world. The five are conscientiousness; agreeableness; extraversion; openness to experience (or its opposite, risk-aversion); emotional stability (or its opposite, neuroticism). Later, one of Oltmanns’ mentors realized that when certain traits cluster or take over, that can be a better indication of a personality disorder than any clinical list of symptoms. Now, with the advent of AI’s large language models, Oltmanns has a new way to use that insight.
But a risky one.
Though word choice can suggest or reinforce a diagnosis, reaching any hard and fast conclusion with an AI language analysis is too slippery a project. Our word choices are also shaped by our surroundings, our upbringing, our education, a fleeting mood, the last article we read…. and sometimes, our race. Remember Galton, the guy who started all this? He was also, in other endeavors, a eugenicist. We have moved past those theories—or would like to think we have—but LLMs are trained on an internet roiling with racial bigotry. Recent research shows that even social media topics differ across racial groups. Comments about feeling worthless are associated with depression for White people but not for Black people, Oltmanns notes, “and use of the first-person pronoun,” I, I, I, I, I—a dwelling on oneself “is a marker for depression in White folks but not in Black.” Self-absorption is perhaps a luxury only possible for the majority? Researchers have not even identified topics or words more likely to indicate depression in Black people, a discrepancy Oltmanns finds even more worrisome.
So how do we get rid of the biases? “Big question,” he says. “And the most pressing issue. The things we can do with these LLMs are amazing, but we really can’t use them yet”—not until computer scientists find ways to reliably eliminate bias. Except, people are using them. With biased models. “Companies are already selling AI psychological assessment tools to hospitals and clinicians,” Oltmanns says, “but it’s not clear to me how well they work or how thoroughly they’ve been evaluated.”
There is huge promise, if care is taken and bias is cleansed from the datasets. How often have you said to someone you love, “Just listen to yourself?” Therapists could have help listening: instant alerts, suggestions, and cross-checks that do not require subjecting clients to an endless battery of tests or waiting for years of session notes. Word choice is especially good for detecting neuroticism and personality disorders, and might even show a propensity for violence. Easy to imagine profilers poring over speech analysis the way they study the messages sent by serial killers, or their social media posts. But while detecting a violent predisposition in advance could save lives, it could also malign them unfairly. Besides, Oltmanns reminds me, “violence is relatively rare, so it would be hard to have a dataset that could predict.”
His focus is more practical: developing technology that will aid therapists: “Ideally, we can have a clinician turn on a recorder during a regular interview, and they can automatically get an assessment afterward.” So far, his team is working with samples already routinely collected, including a trove of data gathered from 1,409 older St. Louisans long before this use was imagined.
Still, they were answering questions for researchers, and the process must have felt a bit clinical and inhibited. Think what AI could deduce from journal entries, love letters, or impassioned court depositions…. All rich with meaning. And all easily misinterpreted, by human or machine.
This fall, Oltmanns’ team is recording volunteers in everyday life with electronically activated recorders that cut on and off during the day. Next, he wants to add video to the mix, so AI can analyze not only the pitch, loudness, pace, and rhythm of speech but how we hold and move our bodies, where our gaze falls or lingers, and what our face expresses, even in microexpressions so fast we seldom register them.
“You get more information, a bigger signal, from language,” Ortmanns says, “but there are plenty of things you can pick up from speech and expression.” Those other modes of communication can also act as a check on the interpretation of language, if, for example, someone’s facial expression contradicts their words.
The chance of misinterpretation by a machine is still high, the risks serious. But the chance of misinterpretation by a human therapist acting alone is even higher, the risks even more serious. Maybe in tandem, they will check each other’s biases?
