Researchers Give AI Models Distinct Personalities, Opening New Training Applications
A team of psychologists and computer scientists has developed a method to inject specific personality traits into large language models, moving beyond the generic, agreeable voice that dominates current AI systems. The technique, called PsychAdapter, accepts psychological profiles as direct inputs and generates text reflecting those characteristics with measurable accuracy.
Today's language models like ChatGPT converge on a narrow, mildly positive voice because they're trained on millions of authors. The result is what researchers describe as "language from nobody" - output that flattens human variation.
PsychAdapter addresses this by modifying standard language models to accept continuous psychological scores: gender, age, Big Five personality traits, depression levels, and life satisfaction. The system then generates text reflecting any point along those dimensions. Expert raters confirmed the approach works with 87% accuracy for personality traits and 97% for mental health characteristics.
The distinction matters. Rather than using a prompt to suggest "write like an extravert," PsychAdapter generates language grounded in actual human writing patterns. Set extraversion to +1.7, depression to +3, and age to young, and the model produces language that reflects that specific profile.
How the System Works
The researchers trained PsychAdapter on roughly 500,000 tweets and 700,000 publicly shared blog posts. Each piece of text was scored for psychological traits using machine-learning models built from survey data - not from the language model itself. This separation is critical: the trait patterns come from how real people describe themselves, not from an AI's assumptions.
The team then embedded psychological scores directly into the processing layers of open-weight language models like Google's Gemma3 and Meta's Llama3. The trait inputs shape language output at every point in the neural network.
Different prompts reveal the personality differences more clearly. When asked "I like to...," high extraverts mention parties and friends while introverts mention reading or gaming. The system can also model complex traits - combining high extraversion with low agreeableness to simulate socially dominant people.
Practical Applications in Training and Research
The method enables AI patients calibrated to specific symptom profiles, allowing crisis-line workers and clinicians to practice without involving real patients. Clinical and educational materials can be tailored to a reader's age and reading level. Social scientists can create digital cohorts with varied personality and demographic profiles to test programs before deploying them with real people.
The research was published in npj Artificial Intelligence. The team released source code and trained models on GitHub for research use.
Ethical Guardrails Required
The researchers acknowledge the dual-use risk. The same tool that trains crisis counselors can be used for micro-targeting and influence operations. They emphasize that any content generated for specific audiences or psychological profiles must be clearly marked as AI-generated.
The work builds on decades of psychology research analyzing language to understand personality. PsychAdapter reverses that direction - generating language for personality instead of analyzing it.
Learn more about how generative AI and LLMs are being customized for specific applications, and explore the latest research in AI development.
Your membership also unlocks:
