By 
Scientists have found that ChatGPT can generate personality assessment questionnaires from source texts and predict how people are likely to respond to the questions, pointing to a potentially significant new application for artificial intelligence in psychological research.
RELATED: The 2026 AI Translation Accuracy Benchmark: Where ChatGPT, DeepL, and Google Translate Actually Fail
The findings, published in the Cell Press journal iScience, suggest that large language models (LLMs) can identify meaningful patterns related to human personality from the vast amounts of language data used to train them.
The study was conducted by Dr Rotem Monsa, Prof Aviv Zohar and Prof Shahar Arzy of the Hebrew University-Hadassah Medical School and the Rachel and Selim Benin School of Computer Science and Engineering.
ChatGPT Shows Understanding of Personality Patterns
According to the researchers, publicly accessible LLMs such as ChatGPT are trained on enormous collections of human language gathered from websites, social media and other online sources.
Because personality traits are frequently expressed through language, the researchers explored whether LLMs might have acquired an implicit understanding of human personality as a by-product of their training.
“Given that personality traits are reflected in language, LLMs may have learned the structure of human personality as a natural byproduct of their training,” said Monsa.
He explained that although LLMs were not specifically trained in psychology or personality theories, patterns relating to human personality may already be embedded in the language data from which they learn.
Researchers Generate Personality Tests From Two Very Different Sources
To investigate the ability of LLMs to assess personality, the researchers used GPT-4 to create two personality questionnaires from different source materials.
The first questionnaire was generated using excerpts from the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), a widely used clinical reference for diagnosing mental disorders.
The researchers based their approach on the assumption that personality traits can be positioned along a continuum associated with personality disorders.
As a control, the team created a second questionnaire using an astrology textbook as its source.
The two sources were deliberately selected because they describe personality in considerable detail but differ substantially in their scientific foundations.
DSM-5 Questionnaire Reflected Established Personality Patterns
Using the DSM-5 material, GPT-4 generated statements describing personality characteristics and asked participants to rate their level of agreement on a five-point scale.
For example, statements derived from the section on paranoid personality disorder included characteristics such as frequently suspecting other people’s motives and ease or difficulty in trusting others.
The astrology-based questionnaire followed a similar format but generated personality statements from descriptions associated with zodiac signs.
Monsa said the researchers chose the two sources to examine whether an LLM could extract meaningful psychological information from texts with very different levels of scientific validation.
The DSM-5, he noted, has been refined through decades of clinical research, while astrology is culturally influential but lacks scientific validation.
600 Participants Tested the AI-Generated Questionnaires
The researchers administered the two AI-generated questionnaires to 600 participants alongside the Big Five Inventory (BFI), one of the most extensively validated personality assessment tools.
The comparison was designed to determine whether questionnaires generated by GPT-4 could produce meaningful measures of personality and how their results compared with an established psychological assessment.
The DSM-5-derived questionnaire demonstrated high internal consistency within personality clusters.
This meant that personality traits that tend to correlate in real-world psychological patterns also showed relationships in participants’ responses.
Importantly, these patterns were similar to those observed with the BFI, providing evidence that the AI-generated questionnaire captured meaningful aspects of human personality.
Even Astrology-Based Questions Captured Psychological Signals
The astrology-derived questionnaire produced weaker internal consistency across personality traits, as expected by the researchers given the lack of scientific grounding behind astrology.
However, the study produced another notable finding.
Despite its weaker consistency, the astrology-based questionnaire could still predict outcomes such as depression, anxiety and well-being at levels comparable with the BFI.
The researchers suggested that the result indicates that personality-related information extracted by an LLM from a source text may retain meaningful psychological signals, even when the underlying framework of the source material is not scientifically validated.
ChatGPT Predicted Responses Before Participants Took the Test
The most striking finding was that GPT-4 could predict how participants would respond to the questionnaires before the participants actually completed them.
For both the DSM-5-based and astrology-based questionnaires, the model accurately anticipated average responses and correlations between questions.
According to the researchers, the result suggests that LLMs may have developed an understanding of population-level personality dynamics from their exposure to large quantities of human language.
“The fact that LLMs can predict human response patterns before seeing any human data suggests that these models have observed something generally meaningful about human psychology,” Monsa said.
He added that the findings indicate LLMs can function not only as content-generation systems but also as informed evaluators of their own outputs.
AI Could Accelerate Psychological Research
The researchers said the findings demonstrate a potential new use for LLMs in psychology: rapidly generating personality-related questionnaires from different source materials and testing whether those questionnaires produce meaningful results.
The combination of AI-generated assessment tools and established psychological methods could accelerate some areas of clinical and experimental psychology, although the findings represent an early research result rather than evidence that AI-generated questionnaires can replace professionally validated clinical assessments.
The study could nevertheless open new avenues for researchers seeking to examine personality, language and human behaviour using AI-assisted methods.
Study Raises Questions About Language and Culture
The researchers cautioned that the findings were based on English-language materials and may not necessarily translate directly across other languages and cultures.
Monsa noted that LLM training data has historically contained a substantial amount of English-language material originating from Western cultures.
“We would assume that in other languages and cultures, the results will not be as strong as we saw here,” he said.
Researchers within the team are now exploring whether similar patterns can be observed when LLMs are used with other languages.
AI and the Future of Personality Assessment
The study adds to growing research into how large language models represent and reproduce patterns of human behaviour.
Its central finding is not simply that ChatGPT can write personality questions, but that the model was able to generate assessments from source material, produce results that showed meaningful psychological patterns and predict population-level response patterns before collecting the participants’ answers.
For psychology researchers, that combination could make AI a powerful research assistant in the development and testing of experimental assessment tools.
The work also raises broader questions about what LLMs may have learned about human behaviour from the enormous body of language on which they were trained — and how that implicit knowledge can be responsibly used in psychological research.
