The Science of Group Creativity: Insights from New Research on Simple AI and Team Performance

July 9, 2024   Matt Jones

  • Group Decision Making
  • Blog

Coming up with good ideas can be hard. It requires outside-the-box thinking to come up with novel concepts and thinking critically to improve the ideas one has already generated. Being part of a group can make this easier by allowing people to brainstorm and bounce ideas off each other. But there are also drawbacks: people might, consciously or subconsciously, fall in line with everyone else’s ideas instead of coming up with their own, which is especially unproductive when the ideas are lousy.

How groups generate and explore a broad set of inter-related ideas that, at the outset, they only have limited knowledge about, and how, more generally, they engage in creative tasks is still poorly understood.  But new experimental research performed by the Sunwater Institute and Yale researchers sheds light on the interplay between individual creativity, group performance, and also how AI might affect the creative and innovative capacity of human teams.

An Experiment on Creativity

Ideas are nebulous things, and have been difficult to formally characterize or measure in a scientific setting. Therefore, we designed an experiment where participants were asked to come up with nouns as a stand-in for ideas. Words, like ideas, have semantic meaning which means some words are more similar than others, and, in the case of words, this semantic similarity can be measured. A “target noun” can be used as a stand in for the best possible idea, and nouns that are similar to the target noun can be deemed to be worth more points than unrelated nouns. The task we invented resembled the game at, but, in our experiment, the games were played in groups. We also made the problem more difficult for participants by introducing decoy nouns that could distract and lead participants astray, just as a group of people trying to figure out how to make a fishing rod might settle on a needlessly suboptimal solution when working together, for example.

Participants were incentivized to find a single target noun which was assigned to have the highest point value out of 20,000 nouns. Other nouns were then assigned points relative to their closeness to the target. For instance, if “dog” was selected as the target with a point value of 20,000, “cat” would also receive a high point value. This task replicates the many human decision-making endeavors where the options are not easily enumerable but are nevertheless related and easily evaluable (e.g., studio executives predicting how the public will react to movies, curriculum committees deciding which new classes to offer, families deciding on the best vacation, etc.).We chose a set of 18 target nouns that were spread out in the “idea landscape” and that were roughly equally obscure and then people worked together to find these unknown targets.

Creativity Chart with and without AI Bots

To determine what behavior was most effective in coming up with the highest scoring words, we sometimes added bots endowed with simple AI to the group; these bots could transmit different words between participants. These bots could repeat nouns suggested by other participants, spreading ideas through the group. Different bots could choose to spread different kinds of nouns. When presented with a list of words, some bots chose random words, some bots chose the word that was most similar to all the other words offered by the people they were connected to, and some bots chose the word that was least similar to the other words (to try to prompt new thinking by others in the group).


What Makes Groups Creative?

We recruited 1,875 online participants to participate in 625 groups that played word-search games, many of which also had bots embedded in the group. First, we found that groups are more than the sum of their parts. A group of 15 participants in constant communication outperformed the same 15 participants if they could not share nouns with each other.

Of all the different bots we tested, the most effective was the bot that shared very mainstream and popular nouns. These nouns tended to be closer to the target, and therefore the bots ensured that more participants were exposed to these high-value nouns and could improve on them even more. Similar to the “wisdom of the crowds,” spreading good ideas through a group can result in even better ideas being explored by the group as a whole.

However, when a creativity problem is difficult, with many different factors to consider, it can be difficult to determine the best way to improve on an existing idea.  We replicated this feature by expanding the strength of the decoy and making the problem more difficult. When we did this, these bots lost their effectiveness. When group members have no general idea of how to come up with better ideas, it does no good to spread popular ideas through a group.

Finally, by comparing each participant’s guesses when they are in a group versus when they are playing the game alone, we also noticed that the participants who succeed in the group setting are not necessarily the same participants who succeed alone. Some participants were very good at playing the game alone but contributed very little to the group, and vice versa. We believe that this points to a larger (and currently unstudied) difference between individual creativity and group creativity.


The human capacity for social learning in groups is enhanced by simple forms of AI, especially in situations where there are distractions or challenges of a certain kind. These simple bots had a notable effect on the ability of groups to find and exploit rewarding regions of semantic space, particularly when sharing similar ideas. Importantly, these bots implement a low-cost, straightforward, and decentralized algorithm, functioning solely with local neighbor information. While our bots only processed simple ideas in our setting, they could in principle be applied to more complex settings.

Read the Full Report >>