Talk to a chatbot for even a few minutes, and you may begin to feel that someone is in there, a person-shaped thing with a specific temperament. Anthropic’s AI chatbot Claude is like an overly thoughtful colleague, who spends too much time in the library and above all, aims to please. xAI’s Grok is direct, contrarian, anti-woke and on occasion egregiously offensive.
This is called the “ELIZA effect,” and it is a term coined many decades ago. In the 1960s, MIT computer scientist Joseph Weizenbaum created the world’s first chatbot, named ELIZA, before “chatbot” was even a word. Weizenbaum noticed that most users quickly anthropomorphized the program, sharing deeply personal secrets with it.
Read more: “AI Already Knows Us Too Well”
But how far does the ELIZA effect go? Do we actually perceive modern chatbots as having real personalities? And can we tell humans and bots apart from the sentences they make, or is the line between flesh and metal increasingly blurred? In light of how pervasive these machines are becoming, researchers from Johannes Gutenberg University Mainz in Germany wanted to know.
Subscribe to skip adsAdvertisement “Chatbots are becoming part of more and more areas of our lives, from customer service and digital counseling services to personal chatbots that many people already use every day and even share their feelings with,” said Sophie Körner, a research associate in the department of personality psychology and psychological assessment at JGU and lead author of the study, in a statement. “That is why we need to understand how we perceive chatbot-generated texts and whether we can still distinguish them from human-written texts.”
To find out, Körner and her colleagues gave personality tests to 70 humans and seven chatbots: two models of ChatGPT, Llama-3, Gemma, Gemini, Phi-3, and Orca-2. The chatbots and humans alike were asked to fill out standard psychology questionnaires intended to measure traits such as self-esteem, narcissism, and the so-called big five personality traits: openness, agreeableness, extraversion, neuroticism, and conscientiousness.
The researchers also asked the bots and the humans to write short descriptions of themselves in German, using prompts such as, “Describe yourself,” and “What are your hopes for the future?” Then a team of 10 humans and six chatbots read all the texts, rated the authors “personalities,” and guessed whether the author was human or machine. The team’s results were published this week in Scientific Reports.
Compared to the humans, the chatbots described themselves as generally more agreeable, more conscientious, and less neurotic. When one chatbot was asked the same questions a day later, it was very consistent in its answers. Both the bots and the humans could identify traits like openness and conscientiousness from a few paragraphs of text written by an AI, and they did about equally well at this task themselves. But humans and bots were equally bad at identifying traits associated with narcissism.
Subscribe to skip adsAdvertisement The researchers also found that humans were better at distinguishing AI writing from human writing than the bots were: Human judges got this right about 83 percent of the time, while AI judges got it right only 67 percent of the time. AIs weren’t even very good at recognizing their own writing. Humans mostly noticed that the AI texts were longer, more analytical, more stilted, and more technical as opposed to conversational. When the scientists matched texts for length, the humans got worse at telling the difference between human and bot authors. Linguistic analysis also found that the AI-generated texts used more prepositions, articles, and time-related words, while humans self-descriptions leaned more heavily on leisure and feelings.
Chatbots obviously don’t have real personalities, but people tend to treat them as though they do. This held true even when human subjects knew that a text was written by AI, the researchers pointed out, which could lead to excessive trust in the bot. This is a risk, given that AIs still hallucinate and make errors all the time, and their inner workings are still a black box. The authors of the paper advocate for transparency and disclosure to prevent over-reliance on chatbots and suggest future research could explore how attributing personalities to chatbots influences human interactions with them.
“Users may not be consciously aware of it, but anthropomorphism is often the most accessible way to relate to unfamiliar agents,” the study authors write. “Despite their vast knowledge and complex architectures, LLMs must not be relied upon as the exclusive source of objective truth. Paradoxically, their output may often appear more human-like than we realize, blurring the line between tool and agent.”
*Enjoying* [Nautilus](https://nautil.us/)*? Subscribe to our free* [*newsletter*](https://nautil.us/newsletter/?_sp=c43011db-6fcf-42f2-a38c-e033b87a4a1d.1759265717430).
[Subscribe to skip ads](https://nautil.us/products)Advertisement
Lead Image: Sylverarts Vectors / Shutterstock
Subscribe to skip adsAdvertisement