The productivity gains of generative AI are often touted as the main reason for its rapid adoption. Much like how a calculator can instantly solve mathematical equations too long for the human brain to compute, generative AI can help someone quickly summarize long articles and draft emails.
But beyond productivity, a key question remains: How is AI affecting quintessentially human activities, such as collaboration and creativity?
As cognitive scientists who study creativity, we recently conducted an experiment that explored how working with AI – rather than humans – affects originality and other aspects of creativity.
Speedy and wordy doesn’t mean creative
Creativity is often tied to what cognitive scientists call “divergent thinking.” One way of measuring divergent thinking is to ask someone to come up with a range of uses for a common item, such as a brick or a book. This is known as the alternate uses task.
Researchers then compare the number of ideas and the originality of these ideas to assess creative capacity.
Large language models appear to have the “number of ideas” part down pat: They can generate many more ideas than humans can. This ability to produce more – and produce that high volume more quickly – partially explains why more than half of new articles online are now generated by AI.
These texts, however, are often verbose, with redundant information. Researchers who tested 12 AI chatbots across different companies on thousands of factual questions found that their answers were often many times longer than human-written answers, with the actual answer making up only about 42% of the text. The rest was largely extra padding: repetitive information, unnecessary background details and explanations that went into too much detail or veered off topic.
Large language models also seem to struggle with the originality of their outputs.
An ‘original’ idea – again and again
In our research, we compared the results of people working on the alternate uses task with another person to those who collaborated with ChatGPT on the same task. We randomly assigned 208 participants to work with either AI or another person.
We found that compared with two people taking turns together, people working with ChatGPT wrote longer responses on the alternate uses task. But human pairs came up with more creative ideas than the humans working with ChatGPT.
We also found that the ideas people came up with when working with other people tended to be more distinct and diverse than those who worked with large language models. Individual AI suggestions could be quite creative – using a book as a sound dampener, for example – but ChatGPT presented that same idea across many separate sessions, making its overall pool of ideas more homogeneous.
Other studies have highlighted the homogeneity of generative AI’s outputs. One group of researchers found that using AI repeatedly over time to generate images often results in bland, generic pictures – what they call “visual elevator music.” Other studies have found that although AI may help an individual write a more creative story when compared with their preexisting work, AI-assisted writing as a whole, across all users, tends to be more similar.
All of this makes sense when you think about how large language models work. After being trained on massive sets of texts, large language models predict the next most likely words. Because their training processes were similar, different large language models, whether it’s ChatGPT, Claude or Gemini, tend to have similar cadences, diction, syntax and suggestions – particularly when compared with the work of two random human writers.
In other words, generative AI’s ideas may seem unique in a vacuum. But if everyone’s using AI, the ideas and writing style start to converge.
Creative collaboration is a long game
Finally, we also found that the creativity benefits of working with another person became even stronger as the collaboration progressed.
Ideas can be measured through semantic distance, which is a way of quantifying how distinct individual concepts are. Over time, interacting with other humans seems to result in larger semantic distance than interacting with a chatbot. In other words, people may traverse more creative terrain when collaborating with other people. Meanwhile, the tendency for ideas to converge in chatbot-assisted brainstorming could accumulate, potentially narrowing the range of ideas people explore over the course of a long period of time.
This is important because many creative projects involve long-term engagement.
For example, to launch a new product, researchers have found that it takes an average of 27 months to go from the initial idea to the finished product. More innovative products can take as long as 54 months.
Chat-human before you ChatGPT
If you want to work with a chatbot, how can you preserve your creativity in the face of AI’s tendency toward homogeneity?
Some research suggests using what’s called “persona prompting.” Like talking to other people with different perspectives, users could provide an extensive prompt to give chatbots a pretend personality that can simulate different types of people. For example, you might prompt it to respond as if it is “a male college professor from India who is an expert in teaching” or “a 58-year-old female founder of a construction company.”
Researchers found that while this persona prompting increased ChatGPT’s idea diversity, the overall range of ideas still fell short of that produced by human writers. These findings echo other recent research suggesting that chatbots are poor simulators of human personality.
If you need help with a creative task, we suggest going back to the basics: Chat with a person before chatting with AI. They’ll have idiosyncratic personal experiences that can’t be captured by a large language model that’s trained to generate average responses.
Composing a sentence, however modest, is a creative act, whether it’s in a text to a friend or a wedding speech. As AI is integrated into workflows, it’s important to remember that other people can still be a valuable resource. Work completed in collaboration with other humans feels more meaningful.
They might not respond as quickly or as comprehensively. But chances are they’ll have something more interesting to say.
This article is republished from The Conversation, a nonprofit, independent news organization bringing you facts and trustworthy analysis to help you make sense of our complex world. It was written by: Clin Lai, Brandeis University and Roger Beaty, Penn State
Read more:
Clin Lai received funding from APA Division 10 and National Science Foundation.
Roger Beaty receives funding from Amazon, the Army Research Institute, and the National Science Foundation.
Commented