Summary: Psychologists are increasingly using artificial intelligence to detect subtle psychological signals in speech — from the words people choose to the tone, pace, and loudness of their voice. These cues can reveal personality traits and early signs of mental health conditions that a clinician might miss. AI offers the potential for faster, more consistent analysis, but researchers warn models must be trained on diverse data to avoid bias. With careful development and validation, AI could become a powerful tool to support clinical assessment.
Artificial intelligence applied to language and speech could transform psychological assessment by uncovering patterns that are time-consuming or difficult for humans to spot. When developed responsibly, these tools can augment clinical judgment, highlight overlooked cues, and provide scalable behavioral insights across larger populations.
Key Facts
- Language as data: Speech patterns and word choice contain measurable clues about personality and mental health.
- AI advantage: Machine models can process audio and text at scale, detecting subtle acoustic and linguistic signals more quickly than manual methods.
- Bias challenge: Fair, representative training data and careful validation are essential to prevent cultural or demographic misclassification.
Source: WUSTL
Words are windows into the mind. The words we choose — and the way we say them — reflect our thoughts, emotions, and behaviors, says WashU psychologist Josh Oltmanns. “Language captures aspects of personality and psychological functioning,” he explains.
Rather than relying solely on lengthy batteries of standardized tests, clinicians could extract meaningful information from even brief language samples. But because many of these signals are subtle, high-tech methods may be needed to detect them consistently.
AI tools trained to recognize linguistic and acoustic markers could streamline assessment and reveal patterns hidden within everyday speech, Oltmanns says.
“Psychologists are human and can miss things,” says Oltmanns, assistant professor of psychological and brain sciences at Washington University in St. Louis. “A well-trained computational model can identify cues that a clinician might overlook.”
In a typical intake, a clinician asks a client to describe their life and concerns. Alongside clinical judgment, that recorded conversation could be analyzed by software designed to detect personality traits and signals of mental health risk. The AI output could validate clinicians’ impressions or flag areas warranting closer attention.
Oltmanns and collaborators, including WashU PhD students Tu Do, Tong Li, and Tongyao Ran, are developing AI methods to surface these hidden cues in language. He recently outlined the promise and pitfalls of such approaches in the journal Advances in Methods and Practices in Psychological Science, with co-authors Mehak Gupta and Jocelyn Brickman.
Language conveys psychological information in multiple ways. Word choice matters across settings — from intimate conversations to social media posts. Earlier work by Oltmanns found links between social media language and the Big Five personality dimensions: openness, neuroticism, agreeableness, conscientiousness, and extraversion.
Beyond vocabulary, how someone speaks offers important signals. “Speech rate, pitch, loudness, and tone can all be diagnostic,” Oltmanns notes. Slower speech can indicate depression, while rapid or pressured speech is often associated with anxiety. Hundreds of acoustic parameters in spoken language may carry clinically relevant information.
Researchers have long used computational tools to quantify language. Early programs like Linguistic Inquiry and Word Count analyzed written text for psychological content. Modern advances in natural language processing and large language models (LLMs) expand those possibilities, enabling richer analyses of both text and speech.
“Contemporary AI can be faster, more comprehensive, and more precise than earlier methods,” Oltmanns says. But he emphasizes significant caveats: AI systems are only as reliable as their training data and evaluation, and models learned from internet-sourced text can reflect societal biases.
If models are not trained and validated across diverse populations, cultural and dialectal differences in speech could be misinterpreted as symptoms. To reduce this risk, Oltmanns is leveraging extensive interview data from the SPAN Study, an ongoing project that includes more than 1,600 adults from St. Louis and aims to represent the city’s racial and cultural diversity.
“We are examining speech patterns among White and Black participants to ensure the models perform equitably across groups,” he says. Careful sampling, transparent evaluation metrics, and ongoing monitoring are key to trustworthy deployment.
Important research questions remain. How do written and spoken language differ as sources of psychological information? What is the minimum amount of speech or text needed for reliable assessment? Which acoustic and linguistic features are most robust across contexts? Oltmanns acknowledges that these are open areas requiring systematic study.
The rapid pace of AI innovation adds urgency. “Companies are already offering AI-based psychological assessment tools to clinicians and health systems, but their performance and validation are not always transparent,” Oltmanns warns. If such tools are to benefit patients and providers, they must be developed carefully, validated rigorously, and integrated with clinical expertise.
About this AI and speech research news
Author: Leah Shaffer
Source: WUSTL
Contact: Leah Shaffer – WUSTL
Image: The image is credited to Neuroscience News
Original Research: Open access. “Large Language Models for Psychological Assessment: A Comprehensive Overview” by Mehak Gupta et al., published in Advances in Methods and Practices in Psychological Science. DOI: 10.1177/25152459251343582
Abstract
Large Language Models for Psychological Assessment: A Comprehensive Overview
Large language models (LLMs) offer powerful tools that can enhance understanding of psychological traits and clinical states. They present an opportunity to complement traditional self-report measures with scalable, behavior-based assessments derived from language and speech.
However, these models also introduce risks and methodological challenges. The article provides a practical overview and guide for psychological scientists evaluating LLMs for assessment. It reviews the evolution of transformer-based LLMs and their advances in natural language processing, then outlines experimental design considerations including language data collection, audio processing and transcription, text preprocessing, model selection, and evaluation.
The authors discuss analytic choices such as model output interpretation, evaluation metrics, hyperparameter tuning, visualization techniques, and topic modeling, providing examples from diverse areas of psychology. Finally, the paper addresses broader ethical and implementation issues and offers directions for future research.
Readers will gain foundational knowledge and practical guidance to navigate the process of applying LLMs to psychological assessment, while remaining attentive to fairness, validity, and clinical applicability.