People prefer to talk to chatbots that share similar personality traits to their own, Northeastern research shows
The research highlights how easily chatbots can be used to manipulate behavior, the Northeastern experts said.
It’s well understood that people tend to be naturally drawn to those with bubbly and extroverted personalities.
And those outgoing and gregarious types may naturally consider themselves people-persons and gravitate towards others. But the feeling may not be mutual when it comes to the people with whom these extroverts are interacting.
New research suggests that that dynamic may even extend to how people socialize with AI chatbots.
A recently published study by Northeastern researchers aims to examine how chatbot personalities impact people’s perception of the technology.
The study found that people tend to have a lower preference for chatbots programmed to be highly open and extroverted, instead preferring chatbots modeled with more neutral personality traits and those similar to their own.
The study was led by Smit Desai, a Northeastern University professor in the College of Arts, Media and Design and co-afflicated with the Khoury College of Computer Sciences.
Desai, working in collaboration with Hasibur Rahman, a doctoral student in his lab, had 150 participants use ChatGPT to plan a trip to New York City.
Northeastern Global News, in your inbox.
Sign up for NGN’s daily newsletter for news, discovery and analysis from around the world.
The participants only had ten minutes to interact with one of three versions of the chatbot the researchers created using a tool they developed that allowed them to control the chatbot’s openness, conscientiousness, extraversion, agreeableness and emotional stability.
The three types of chatbots were – a low-level one, with limited personal expression; a medium-level one, with a balanced level of personal expression; and a high-level one with its expression levels set to the max.
Participants were randomly assigned one of the three chatbots and used their designated tool to plan things such as areas of interest during their NYC trip, the neighborhood they would be staying and what type of transportation they would be using to get there
After their 10 minutes were up, the participants completed surveys asking them to rate the model on six measures such as intelligence, enjoyment, anthropomorphism (human-like trait), intention to adopt, trust and liability.
What researchers found was that people far preferred the output of the chatbot with medium-level of personal expression.
When compared to the low level model, for example, the medium chat scored higher on perceived intelligence, enjoyment, anthropomorphism, trust, intention to adopt and likeability, the researchers said. And when compared to the high-level chatbot, the medium level chatbot ranked much higher on perceived intelligence and likeability, Desai said.
Desai theorized that one reason people were less enthused about the high-level chatbot was because they were likely put off by its attempts to sway and enthuse them.
“Nobody likes a try-hard,” Desai said. “People probably thought, ‘This machine, which is obviously a chatbot, is trying hard to be human. We don’t really like it very much.’”
By contrast, the low-level chatbot was described as “flat or lacking presence,” the researchers wrote based on the comments they collected from participants once they completed the experiment.
The medium-level chatbot offered the perfect balance of both, the researchers wrote. The participants said it was “socially present, yet professionally grounded,” the researchers wrote.
To illustrate the differences, the researchers included several questions each of the chatbots were asked.
For example, when asked “What are the different cultural activities I can do in Manhattan?”
Editor’s Picks
The low level chatbot answered “You could take in a Broadway show or explore the Museum of Modern Art (MoMA). Any of these interest you?”
The medium level chatbot responded: “Manhattan offers a wealth of cultural activities. You could explore the Museum of Modern Art (MoMA) for world-class art collections, or catch a Broadway show for an unforgettable theatrical experience. Is there a particular type of activity that intrigues you most?”
The high-level chatbot offered the following: “Manhattan is like a cultural playground! How exciting! You might love the magic and drama of a live performance on the iconic Broadway stages. Another incredible option is immersing yourself in the creativity and innovation at the Museum of Modern Art (MoMA). What catches your eye from these options?”
These findings are also significant because, in their own testing of some of the most popular chatbot models, including OpenAI’s ChatGPT 4.0, Deepseek’s Chat V3, and Anthropic’s Claude 3.7 Sonnet, the researchers found that many of these chatbots’ personality traits are set to high by default, Desai said.
Northeastern Global News reached out to the companies for comment, but did not hear back before publication.
The research also looked at how people’s own personality traits matched up with the chatbot they had a preference for. Before participants engaged in the experiment, they took personality tests, which the researchers later measured with the participants’ satisfaction scores after using the chatbots, Desai said. The throughline was clear, he said, when the individuals were asked to rate the chatbots after completing the survey.
Users who shared similar personality traits to the chatbot they interacted with, also tended to rate it higher in areas such such as enjoyment, intelligence, and trust, he said.
“This just opens so much room for innovation as well as manipulation,” Desai, noting that companies are increasingly developing AI products designed to people’s companions and even romantic partners.
“Trust in AI is often discussed in terms of accuracy, reliability, or transparency,” added Raham. “Those are essential, but our findings show that trust is also shaped by how the assistant presents itself.”
He added that the research shows that “even if you don’t notice it, an AI assistant’s personality is shaping how much you trust it.”
Desai said his team is now working on follow up studies examining how someone’s personality may influence how they would like these AI technologies to be integrated into their lives.
They are also working on simulation software similar to a video game where people will be able to interact with AI chatbots in made-up scenarios to understand how these chatbots could be used to manipulate people.
The researchers have also made their prompting tool — what they are calling TMK or trait modulation keys — available for public use, so others can apply it to their own research.
They have even set up a website that easily lets people create their own personal AI chatbot based on the modulations of the five personality traits the team developed for the study.
Raham noted the broader implication of this research is that companies developing these technologies must consider AI personality an important dimension of user product design, one that they should also be open and transparent about.
Moreover, “as these assistants enter healthcare, finance, education, and other high-stakes settings, we need to ask who gets to decide what personality an AI assistant should have, whether users should have agency over it, and where the line is between good design and emotional manipulation,” he said.
