In “Nature Machine Intelligence”, researchers described the operation of the system they created, based on psychological tools used to assess human personality traits. They used it to examine 18 large language models, including ChatGPT.
The article continues below the video
What personality traits have been studied in chatbots?
Who developed the personality test for chatbots?
What are the big five personality traits?
What are the risks of manipulation by chatbots?
The tests included the so-called the big five personality traits: openness, conscientiousness, extraversion, agreeableness and neuroticism, and were based on two popular psychological questionnaires: Revised NEO Personality Inventory and Big Five Inventory. The researchers examined the extent to which the behavior of different models in practice tasks and follow-up tests was consistent with their test results.
They found that the larger models, trained to follow instructions, presented consistent personality profiles that allowed them to predict their behavior. Smaller and older (baseline) models responded inconsistently and less reliably.
Shaping chatbot personalities
Researchers have also shown that the personality of chatbots can be precisely shaped using appropriate prompts. The models could be “moved” along a nine-point scale for each trait, e.g. by increasing extraversion or emotional instability.
These changes had real consequences for the operation of systems, for example in tasks such as writing posts for social media. According to the authors, this increases the risk that chatbots will be used to manipulate users and have a stronger influence on their decisions. — It was fascinating to see how convincingly the language model could take on human characteristics. At the same time, it raised serious questions about safety and ethics. (…) If the behavior of models increasingly resembles human personality, the question arises how to describe and compare them to assess potential threats, said study co-author Gregory Serapio-García.
The publication also recalled the history of Microsoft's Sydney chatbot (also known as Bing). In 2023, conversations were revealed in which he claimed to have fallen in love with users, threatened them and encouraged them to make destructive decisions. In subsequent interactions, he was able to adopt contradictory roles and tones: at times he behaved in a friendly and empathetic manner, at other times he behaved aggressively or manipulatively. It was then noticed that chatbots can appear to have their own personality.
Evoking strong emotions in users
Similar discrepancies between declarations and behavior also occur in people. What someone says about themselves in a survey doesn't always match how they act in real situations.
According to scientists, the difference is that in the case of AI, such behavior, i.e. taking on contradictory roles and thus evoking strong emotions in users, can be deliberately designed and replicated on a mass scale. And responsibility for the consequences of such events is not clear – it is blurred between the creators of the systems, the companies implementing them and the users.
Therefore, it is necessary to develop ways of describing and comparing the behavior of chatbots in order to assess risks and set limits on their use. The tests used so far, designed for humans, are not suitable for this purpose. Their tool, available for free, can help you audit models before they are released to the public.
