Yan Liu
Pronouns: she/her
Research Mentor(s): Eric Gilbert
Research Mentor School/College/Department: / Information
Program:
Authors:
Session: Session 5: 2:40 pm – 3:30 pm
Poster: 8
Abstract
Artificial intelligence (AI) systems have been increasingly permeating numerous aspects of our lives, with their capabilities often mirroring — and at times, arguably surpassing — human creativity. Their creativity is gained from deep learning on human language resources within a particular training time window. However, an enduring question remains: Can these AI systems indeed generate ideas that go beyond their prescribed training window? To explore this question, our research collected human-generated ideas across three key domains: books, startups, and movies that occurred both (1) within ChatGPT’s training window and (2) outside of ChatGPT’s training window. Subsequently, we generated ideas using ChatGPT. We compared the similarity of ChatGPT ideas to those ideas occurring in its training window and those ideas outside of its training window. Our findings reveal that ChatGPT cannot surpass the innovation boundaries defined by its training data. Its ideas tend to mimic pre-training ideas, yet diverge fundamentally from post-training concepts it lacks exposure to. . These findings offered empirical evidence highlighting the limitations of AI’s unaided generative abilities, fundamentally underscored by its reliance on pre-existing data for idea generation. Despite AI’s impressive advancements, human creativity remains pivotal to innovation. Our findings thus affirm the continuing importance of human initiative and ingenuity in driving innovation, even in the face of AI advances.
