AI could be better than humans at predicting correlations between answers to personality test questions

Artificial intelligence (AI)-based models have proved to be promising tools for predicting some data trends and future outcomes with good accuracy. These models’ ability to make predictions by analyzing data can be particularly valuable for research, for instance, helping scientists to devise hypotheses to test and predict the outcomes of experiments.

Researchers at London School of Economics and Political Science, Spark Wave, New York University and University of Oxford recently carried out a study aimed at assessing the ability of AI models to predict the correlations between the answers that individuals would give to different questions in widely used personality tests.

Their findings, published in Communications Psychology, show that a specialized AI-based tool they developed can predict these correlations much better than humans.

“The original idea for this paper came from a conversation between two co-authors, Dr. Lucius Caviola and Dr. Spencer Greenberg, who discussed the idea of rigorously testing the personality prediction capabilities of frontier models as well as PersonalityMap,” Philipp Schoenegger, first author of the paper, told Medical Xpress.

“PersonalityMap is a free platform developed by two of our co-authors, Spencer and Alex, and just released to the public, which enables users to explore over 1 million human correlations, including personality, beliefs and behaviors.”

PersonalityMap allows users to access the correlations between the personality scores, beliefs and behavioral patterns of participants who took part in past research studies. In addition, the researchers trained an artificial neural network on these correlations, allowing it to make predictions based on the relationships it learned.

This neural network now powers the platform, quickly making predictions about the relationships between different traits, beliefs and behavioral patterns in response to user queries. As part of their recent study, Schoenegger and his colleagues set out to compare the predictions made by this specialized AI model to those made by other general-purpose AI models (i.e., GPT-4o and Claude3 Opus), 254 non-expert humans and 272 psychology experts.

“We selected a set of personality psychology-questionnaire item pairs with known correlations (that the AIs were not allowed to train on to prevent cheating) and asked both human participants (including both laypeople and experts) and AI models to estimate those correlations,” explained Schoenegger.

“We then compared their estimates to the actual values. This allowed us to see how accurately each group—individual humans, aggregated humans, and various AI models—could predict the true relationships.”

The findings of the experiments carried out by Schoenneger and his colleagues revealed that all the AI models they tested are better at predicting the correlations between people’s answers to questions on personality tests than most laypeople and psychology experts.

Notably, the specialized neural network-based model developed by Schoenegger’s colleagues, Spencer Greenberg and Alexander Grishin, was found to outperform both GPT-4o and Claude 3 Opus, two of the most widely used general purpose large language models (LLMs).

“When comparing individual experts/laypeople with single model runs, we found that most AI models are better than the vast majority of laypeople and most experts too,” said Schoenegger. “The main implication here is one of cost. Instead of asking an expert to provide an estimation of the relationship between two items, one can simply query an AI to arrive at similar or better results.”

Overall, the findings of this recent study highlight the potential of AI-based models, particularly those trained on past research data, for predicting the correlations between an individual’s answers in personality tests. PersonalityMap, the model developed by the researchers, could soon be used by other psychology researchers to plan their research, for instance by predicting relationships that could emerge and identifying promising hypotheses before conducting experiments.

“While we are not currently planning/running a follow-up study, there are a variety of research questions that could be explored in future research, such as focusing on expanding these models to handle non-linear dynamics, evaluating their robustness to novel item statements, quantifying prediction uncertainty, and moving beyond simple associations toward more causal analyses,” added Schoenegger.

“PersonalityMap is also under active development, with the goal of including an ever-increasing range of human traits that it can make predictions about, such as other beliefs, behaviors, demographic factors and health-related variables with the goal of accelerating scientific work.”

© 2025 Science X Network

Artificial intelligence (AI)-based models have proved to be promising tools for predicting some data trends and future outcomes with good accuracy. These models’ ability to make predictions by analyzing data can be particularly valuable for research, for instance, helping scientists to devise hypotheses to test and predict the outcomes of experiments.

Researchers at London School of Economics and Political Science, Spark Wave, New York University and University of Oxford recently carried out a study aimed at assessing the ability of AI models to predict the correlations between the answers that individuals would give to different questions in widely used personality tests.

Their findings, published in Communications Psychology, show that a specialized AI-based tool they developed can predict these correlations much better than humans.

“The original idea for this paper came from a conversation between two co-authors, Dr. Lucius Caviola and Dr. Spencer Greenberg, who discussed the idea of rigorously testing the personality prediction capabilities of frontier models as well as PersonalityMap,” Philipp Schoenegger, first author of the paper, told Medical Xpress.

“PersonalityMap is a free platform developed by two of our co-authors, Spencer and Alex, and just released to the public, which enables users to explore over 1 million human correlations, including personality, beliefs and behaviors.”

PersonalityMap allows users to access the correlations between the personality scores, beliefs and behavioral patterns of participants who took part in past research studies. In addition, the researchers trained an artificial neural network on these correlations, allowing it to make predictions based on the relationships it learned.

This neural network now powers the platform, quickly making predictions about the relationships between different traits, beliefs and behavioral patterns in response to user queries. As part of their recent study, Schoenegger and his colleagues set out to compare the predictions made by this specialized AI model to those made by other general-purpose AI models (i.e., GPT-4o and Claude3 Opus), 254 non-expert humans and 272 psychology experts.

“We selected a set of personality psychology-questionnaire item pairs with known correlations (that the AIs were not allowed to train on to prevent cheating) and asked both human participants (including both laypeople and experts) and AI models to estimate those correlations,” explained Schoenegger.

“We then compared their estimates to the actual values. This allowed us to see how accurately each group—individual humans, aggregated humans, and various AI models—could predict the true relationships.”

The findings of the experiments carried out by Schoenneger and his colleagues revealed that all the AI models they tested are better at predicting the correlations between people’s answers to questions on personality tests than most laypeople and psychology experts.

Notably, the specialized neural network-based model developed by Schoenegger’s colleagues, Spencer Greenberg and Alexander Grishin, was found to outperform both GPT-4o and Claude 3 Opus, two of the most widely used general purpose large language models (LLMs).

“When comparing individual experts/laypeople with single model runs, we found that most AI models are better than the vast majority of laypeople and most experts too,” said Schoenegger. “The main implication here is one of cost. Instead of asking an expert to provide an estimation of the relationship between two items, one can simply query an AI to arrive at similar or better results.”

Overall, the findings of this recent study highlight the potential of AI-based models, particularly those trained on past research data, for predicting the correlations between an individual’s answers in personality tests. PersonalityMap, the model developed by the researchers, could soon be used by other psychology researchers to plan their research, for instance by predicting relationships that could emerge and identifying promising hypotheses before conducting experiments.

“While we are not currently planning/running a follow-up study, there are a variety of research questions that could be explored in future research, such as focusing on expanding these models to handle non-linear dynamics, evaluating their robustness to novel item statements, quantifying prediction uncertainty, and moving beyond simple associations toward more causal analyses,” added Schoenegger.

“PersonalityMap is also under active development, with the goal of including an ever-increasing range of human traits that it can make predictions about, such as other beliefs, behaviors, demographic factors and health-related variables with the goal of accelerating scientific work.”

© 2025 Science X Network