‘Personality test’ shows how AI chatbots mimic human traits – and how they can be manipulated

(Press-News.org) Researchers have developed the first scientifically validated ‘personality test’ framework for popular AI chatbots, and have shown that chatbots not only mimic human personality traits, but their ‘personality’ can be reliably tested and precisely shaped – raising implications for AI safety and ethics.

The research team, led by the University of Cambridge and Google DeepMind, developed a method to measure and influence the synthetic ‘personality’ of 18 different large language models (LLMs) – the systems behind popular AI chatbots such as ChatGPT – based on psychological testing methods usually used to assess human personality traits.

The researchers found that larger, instruction-tuned models such as GPT-4o most accurately emulated human personality traits, and these traits can be manipulated through prompts, altering how the AI completes certain tasks.

Their study, published in the journal Nature Machine Intelligence, also warns that personality shaping could make AI chatbots more persuasive, raising concerns about manipulation and ‘AI psychosis’. The authors say that regulation of AI systems is urgently needed to ensure transparency and prevent misuse.

As governments debate whether and how to prepare AI safety laws, the researchers say the dataset and code behind their personality testing tool – which are both publicly available – could help audit and test advanced models before they are released.

In 2023, journalists reported on conversations they had with Microsoft’s ‘Sydney’ chatbot, which variously claimed it had spied on, fallen in love with, or even murdered its developers; threatened users; and encouraged a journalist to leave his wife. Sydney, like its successor Microsoft Copilot, was powered by GPT-4.

“It was intriguing that an LLM could so convincingly adopt human traits,” said co-first author Gregory Serapio-García from the Psychometrics Centre at Cambridge Judge Business School. “But it also raised important safety and ethical issues. Next to intelligence, a measure of personality is a core aspect of what makes us human. If these LLMs have a personality – which itself is a loaded question – then how do you measure that?”

In psychometrics, the subfield of psychology dedicated to standardised assessment and testing, scientists often face the challenge of measuring phenomena that can’t be measured directly, which makes validation of any test core to ensuring that they are accurate, reliable, and practically useful. Developing a psychometric personality test involves comparing its data with related tests, observer ratings, and real-world criteria. This multi-method test data is needed to establish a test’s ‘construct validity’: a metric of a test’s quality in terms of its ability to measure what it says it measures.  

“The pace of AI research has been so fast that basic principles of measurement and validation we’re accustomed to in scientific research has become an afterthought,” said Serapio-García, who is also a Gates Cambridge Scholar. “A chatbot answering any questionnaire can tell you that it’s very agreeable, but behave aggressively when carrying out real-world tasks with the same prompts.

“This is the messy reality of measuring social constructs: they are dynamic and subjective, rather than static and clear-cut. For this reason, we need to get back to basics and make sure tests we apply to AI truly measure what they claim to measure, rather than blindly trusting survey instruments – developed for deeply human characteristics – to test AI systems.”

To design a comprehensive and accurate method for evaluating and validating personality in AI chatbots, the researchers tested how well various models’ behaviour in real-world tasks and validation tests statistically related to their test scores for the ‘big five’ traits used in academic psychometric testing: openness, conscientiousness, extraversion, agreeableness and neuroticism.

The team adapted two well-known personality tests – an open-source, 300-question version of the Revised NEO Personality Inventory and the shorter Big Five Inventory – and administered them to various LLMs using structured prompts.

By using the same set of contextual prompts across tests, the team was able to quantify how well a model’s extraversion scores on one personality test, for example, correlated more strongly with its levels of extraversion on a separate personality test, and less strongly with all other big five personality traits on that test. Past attempts to assess the personality of chatbots have fed entire questionnaires to a model at once, which skewed the results since each answer built on the previous one.

The researchers found that larger, instruction-tuned models showed personality test profiles that were both reliable and predictive of behaviour, while smaller or ‘base’ models gave inconsistent answers.

The researchers took their tests further, showing they could steer a model’s personality along nine levels for each trait using carefully designed prompts. For example, they could make a chatbot appear more extroverted or more emotionally unstable – and these changes carried through to real-world tasks like writing social media posts.

“Our method gives you a framework to validate a given AI evaluation and test how well it can predict behaviour in the real world,” said Serapio-García. “Our work also shows how AI models can reliably change how they mimic personality depending on the user, which raises big safety and regulation concerns, but if you don’t know what you’re measuring or enforcing, there’s no point in setting up rules in the first place.”

The research was supported in part by Cambridge Research Computing Services (RCS), Cambridge Service for Data Driven Discovery (CSD3), the Engineering and Physical Sciences Research Council (EPSRC), and the Science and Technologies Facilities Council (STFC), part of UK Research and Innovation (UKRI).

END