Researchers introduce PsychAdapter to tune AI text by personality and age

The open-source method adds psychological and demographic controls to language models, with proposed uses in therapist training, personalized education, and social science research.

Researchers have developed PsychAdapter to generate AI text shaped by selected personality, age, and mental health characteristics

Researchers from Stony Brook University, Stanford University, Vanderbilt University, and other institutions have developed PsychAdapter, a method for generating AI text that reflects selected personality traits, ages, and mental health characteristics.

The study, published in npj Artificial Intelligence, modifies open-weight language models by feeding numerical psychological scores directly into their processing layers. The approach is designed to provide more precise control than prompting a model to imitate a broad personality type.

PsychAdapter can use combinations of variables including age, gender, the Big Five personality traits, depression, and life satisfaction. The researchers report that expert raters matched generated text with the intended profile with 87 percent accuracy for personality and 97 percent for mental health characteristics.

The team identifies potential applications in simulated patient conversations for therapist and crisis-line training, educational materials adapted to a reader's age or reading level, and digital cohorts used to test social science ideas before involving real participants.

The researchers have released the source code and trained models on GitHub for research use.

Psychological scores shape language throughout the model

Large language models are trained on writing produced by many different people, but the researchers argue that their outputs often converge around a relatively uniform tone.

PsychAdapter is intended to restore some of the variation found in human language by controlling several traits at the same time.

Instead of asking a model to write like an introvert or an extrovert, a researcher can enter scores across multiple dimensions. The model then generates text reflecting that combination throughout its neural network.

The researchers say this allows more specific profiles to be created, including combinations that would be difficult to reproduce consistently through prompts alone.

H. Andrew Schwartz, Co-Senior Author and Associate Professor of Computer Science and Psychology at Vanderbilt University's College of Connected Computing, says: "Language models are now ubiquitous behind programs like ChatGPT and Claude, but they fundamentally, statistically, miss the idea that each person talks differently. PsychAdapter addresses this, helping to generate language across the many psychological dimensions that distinguish people."

Different prompts can make the selected traits more visible. The researchers found, for example, that highly extraverted profiles were more likely to refer to parties and friends, while introverted profiles more often mentioned reading or gaming.

Lead Author Huy Vu says: "Having this simultaneous control also meant we could model more complex psychological traits. For example, we could model dominant people who are socially dominant by combining high extraversion with low agreeableness."

PsychAdapter can be applied to open-weight models including Google's Gemma 3, Meta's Llama 3, and OpenAI's GPT-2.

Training combined social media language with survey-based models

PsychAdapter was trained using approximately 500,000 tweets and 700,000 publicly shared blog posts.

Each post was assigned scores for traits such as age, extraversion, and depression using existing machine learning models. Those earlier models had been developed by comparing survey responses with the language used by participants.

This created a training dataset in which pieces of writing were paired with psychological profiles.

The researchers say this approach grounds the generated language in observed patterns from human writing rather than relying on a language model's own interpretation of how a particular personality should sound.

The psychological scores are embedded within the model's processing layers and influence output throughout text generation.

The study reports expert-rater accuracy of 87 percent for personality and 97 percent for mental health profiles. The supplied material does not provide the number of raters, performance across each individual trait, or comparisons with alternative personalization methods.

The reported results also do not establish that a generated profile accurately represents an individual person or can be used as a clinical assessment.

Johannes C. Eichstaedt, Co-Senior Author and Faculty Fellow at the Stanford Institute for Human-Centered AI, says: "This is the next step in a tradition that runs from the earliest psychologists to the age of big data. For the last half century, psychology has built the tools to analyze language to understand personality — now we can generate language for personality as well."

Education and clinical uses bring safeguards into focus

The researchers propose using PsychAdapter to create simulated patients with defined symptom profiles for training therapists, clinicians, and crisis-line workers.

Such simulations could allow trainees to practice responding to different communication styles without involving real patients.

In education, the method could be used to adapt materials according to age, reading level, or selected learner characteristics. The announcement does not include evidence that this form of psychological personalization improves learning outcomes or student engagement.

Social scientists could also create simulated groups with different demographic and personality profiles to test research questions or interventions before conducting studies involving people.

The same controls could be used for targeted persuasion, behavioral profiling, or content designed to influence specific psychological groups.

Eichstaedt says: "It's absolutely critical that any content generated for specific audiences or psychological profiles be clearly marked as AI-generated."

He adds: "The same tool that could help train crisis counselors or personalize education can also be used for micro-targeting and influence operations."

The research team includes contributors from Stony Brook University, Stanford University, Vanderbilt University, New York University, the University of Melbourne, the University of Texas at Dallas, and the University of Pennsylvania.

The source code and trained models are now available on GitHub for research use.