We Attribute Personalities to Chatbots, Even When We Know Better

Chatbots may possess artificial personalities, even when users know better. Engaging with a chatbot for a brief period can lead individuals to perceive someone with a distinct temperament, often described as overly thoughtful or overly eager to please. This phenomenon, known as the "ELIZA effect," was first observed in 1966 by MIT computer scientist Joseph Weizenbaum, who created the first chatbot named ELIZA.

The ELIZA effect manifests even as chatbots become increasingly integral to various aspects of daily life, including customer service, digital counseling, and personal interactions.

Researchers from Johannes Gutenberg University Mainz in Germany sought to investigate how humans distinguish chatbot-generated texts from human-written ones. In their study, 70 humans and seven chatbots participated in personality assessments and text generation tasks. The chatbots and humans completed standard psychology questionnaires measuring traits such as self-esteem, narcissism, and the big five personality traits.

They were also asked to describe themselves in German, followed by a group of human and machine raters who evaluated the texts, guessed their author's identity, and assessed the perceived personalities.

The results, published in Scientific Reports, showed that chatbots generally described themselves as more agreeable, conscientious, and less neurotic compared to humans. Both humans and chatbots were able to identify traits like openness and conscientiousness from brief text descriptions generated by AI. However, humans and bots performed equally poorly at identifying narcissistic traits.

Notably, humans were better at discerning AI writing from human-written texts, correctly identifying the difference about 83% of the time, while AI judges achieved a mere 67% accuracy.

When analyzing AI-generated texts, researchers found that humans noticed certain characteristics that set the AI apart, such as longer, more analytical, and stilted language rather than conversational or personal. The AI texts also contained more prepositions, articles, and time-related words, while human self-descriptions leaned more heavily on leisure activities and emotions.

Despite their complexity, large language models (LLMs) still struggle to possess genuine personalities, and the ELIZA effect demonstrates how users readily anthropomorphize these artificial agents, even when aware of their artificial nature.

This tendency to attribute human-like qualities to chatbots could potentially lead to over-reliance on these AI systems, which are still prone to errors and hallucinations, as well as opaque inner workings. The researchers emphasize the importance of transparency and disclosure to mitigate potential risks associated with relying on chatbots for objective truth.

By recognizing the limitations of AI-generated personalities, users can strike a balance between leveraging the benefits of chatbots and maintaining a critical perspective on their interactions with these digital entities.