Collective Creative Search with Humans and AI

Studying how human–AI synergy supports collective creative search in a controlled semantic word-guessing task.

Generative AI is increasingly transforming creativity into a hybrid human–artificial process, but its impact on collective creative search remains unclear. This project studies human–AI synergy using a controlled word-guessing task that balances open-ended idea generation with an objective measure of task performance.

Participants attempt to infer a hidden target word, receive feedback based on the semantic similarity of their guesses to the target, and observe the best guess from previous players in the same chain. This creates a controlled form of collective creative search: people and AI agents explore a semantic space together, receive objective feedback, and inherit information from prior rounds.

The new preprint compares Human Social, Human Asocial, AI-only, and Human–AI Hybrid conditions. Hybrid human–AI groups outperform both human-only and AI-only groups, while maintaining intermediate diversity and adaptive search strategies. Within hybrid groups, both humans and AI agents adjust their behavior relative to single-agent conditions: humans contribute broader exploratory signals, while AI agents exploit promising semantic regions more efficiently.

Read the paper PDF

Experiment framework for collective creative search
Experiment framework for collective creative search. (A) Participants infer a hidden target word and receive similarity score feedback over multiple turns. (B) Each guess is scored from the product of the hidden-word score and the cosine similarity between the guess and target. (C) In each round, participants receive the best guess from previous rounds as a hint. (D) Participants are embedded in a chain-like collective guessing game with 10 rounds of 10 turns each. (E) Experimental conditions compare Human Social, Human Asocial, AI-only, and Human–AI Hybrid groups.
Semantic exploration trajectories in UMAP space
Semantic exploration trajectories. Words are embedded with Word2Vec and projected to two dimensions with UMAP. Colored trajectories show five games per condition for the example target word "compass." For each game and round, guesses are averaged into round centroids and connected from rounds 1 to 10. Insets show example rounds with all 10 guesses and similarity scores; best guesses are highlighted.

What we do

  • Design controlled semantic search paradigms that combine open-ended creative exploration with objective performance metrics.
  • Compare Human Social, Human Asocial, AI-only, and Human–AI Hybrid collective search across semantic spaces.
  • Analyze how agents adapt their strategies in response to collaborators, social hints, and changing hint quality.
  • Characterize the complementary roles of humans and AI: broad human exploration and efficient AI exploitation.
  • Test whether performance benefits can be reproduced through heterogeneous AI–AI collaboration, including Gemini 2.5 and GPT-5.1 agents.
Performance and diversity across conditions
Figure 3: Performance, diversity and strategy. A. Performance, computed as the average of the maximal score across rounds. Error bars represent one standard error across participants. Asterisks *, **, and *** denote significance levels of 0.05, 0.01, and 0.001, respectively. To account for multiple comparisons (here and in the rest of the paper), only results that remained significant after Bonferroni correction are shown. B. Performance across rounds. Error bars represent standard error across participants. C. Lexical diversity, computed as the average proportion of unique guessed words of each game. D. Cross-game lexical diversity across rounds. E. Strategy distribution by hint score quantile across conditions. The dashed red line shows the OLS trend for P(Exploit) across hint score bins; slope and bootstrapped 95% CI are annotated within each panel.

Why it matters

Hybrid human-AI groups outperformed both human-only and AI-only groups, demonstrating that the benefits of collaboration are synergistic. This advantage stems from complementary roles of human and AI: AI agents stuck rigidly to exploitation strategies regardless of hint quality, while humans adaptively balanced exploration and exploitation. When they are placed together, the benefit is asymmetric: human input pushed AI agents toward more varied and productive regions of semantic space, improving their performance and diversity, while humans’ own performance changed little, suggesting they served mainly as a source of exploratory signal. Critically, this synergy is not just about having diverse agents in the mix. Combining two different AI systems reproduced some of these gains, but human–AI collaboration remained superior. Together, these findings inform the design of creative AI tools, collective intelligence platforms, and human-centered AI systems that leverage complementary strengths.

(Li et al., 2026)

People on this project

Related Publications

2026

  1. Human-AI Synergy Supports Collective Creative Search
    Chenyi Li, Raja Marjieh, Haoyu Hu, Mark Steyvers, Katherine M. Collins, Ilia Sucholutsky, and Nori Jacoby
    In Proceedings of the Annual Meeting of the Cognitive Science Society, 2026
    DOI URL arXiv