Motivational and emotional factors cannot be ignored in causing variations in performance, which were absent in the chatbots’ responses as they executed the assigned task always at the best of their capacities, they said
The best, most creative ideas came from humans, said researchers who analysed responses of humans and ChatGPT in a creative task.
Artificial intelligence (AI)-based chatbots might, however, outperform the average human at creative thinking, the researchers said in their study published in the journal Scientific Reports.
They found that while the humans’ ideas were wide-ranging, and included the best and the worst from the entire output, AI models’ ideas were consistently better than the average human creative ideas.
The results also suggested that generating creative ideas may not be a feature unique to conscious human beings, their study said.
For the study, the researchers from Finland and Norway used the Alternate Uses Test (AUT), whereby the test-takers, including AI-chatbots ChatGPT-3, ChatGPT-4 and Copy.AI, were asked to imagine uncommon and creative utilities, including silly and strange ones, for everyday objects such as “rope”, “box”, “pencil” and “candle”.
AUT is a widely used measure in assessing divergent thinking, studied to play a role in activities like brainstorming and mind-mapping and thought to be key to creativity.
“Divergent thinking tests are perhaps the most widely recognised tools for gauging everyday creativity. They challenge participants to generate a variety of ideas in reaction to given stimuli, which can be either visual or verbal,” Simone Grassini, the study’s corresponding author, told PTI in an email.
The team recruited 256 participants aged 19-40 years, 44 of whom were students. Each subject’s creative output was finally scored based on the count of ideas (fluency), flexibility in uses suggested, originality and elaboration (level of detail in responses).
The clearest weakness in the humans’ performance lay in the relatively high number of low-quality ideas, possibly arising due to failures in their executive functions, the study said of its human subjective ratings results, according to which, ChatGPT-4’s performance was adjudged the most superior.
Executive functions involve selecting behaviours enabling one to achieve a chosen goal, which here was generating creative ideas, and include attentional control, cognitive inhibition and flexibilty and working memory.
Difficulty in maintaining task goals in their working memory, for example, or in inhibiting ideation from closely-related concepts and in switching attention to distantly related concepts could be possible reasons for failure, the researchers said.
Motivational and emotional factors cannot be ignored in causing variations in performance, which were absent in the chatbots’ responses as they executed the assigned task always at the best of their capacities, they said.
Further, the study’s semantic distance results found that chatbots performed better than the average human, even as they did not consistently outperform humans’ best performance.
According to associative theories of creativity, creativity is linked to flexible and highly connected semantic networks, helping make associations between different concepts in the brain’s stored knowledge.
The speed of AI models in accessing large data structures may explain their higher-than-average performance, even though it is not clear how semantic networks are represented in their memory, the researchers said.
“It is difficult to pinpoint specific mechanisms,” said Grassini, who referred to the internal functioning of AI models as a “black box”.
“We could, however, hypothesise that either the model has learnt optimal responses from some document, like training material it found on the internet, or that it has generated answers as an interpretation of our query.
“In this case, we could define that such type of “creativity” expressed by the AI-model is some sort of emerging quality of its functioning,” said Grassini, also an associate professor of psychology, University of Bergen, Norway.
Thus, highly creative individuals, likely possessing flexible semantic networks, can still compete with AI in triggering distant, weakly related concepts for ideation, the authors said in their study.