to exercise any judgment on the veracity or feasibility of the output. Further, the underlying optimization algorithms provide no performance guarantees, and their output can thus be of inconsistent quality. Hallucinations and inconsistency are critical flaws that limit the use of LLM-based solutions to low-stakes settings or in conjunction with expensive human supervision. In what applications can we leverage artificial intelligence that is brilliant in many ways yet cannot be trusted to produce reliably accurate results? One possibility is to turn their weaknesses – hallucinations and inconsistent quality – into a strength (Terwiesch, 2023). In most management settings, we expect to make use of each unit of work produced. As such, consistency is prized and is, therefore, the focus of contemporary performance management. (See, for example, the Six Sigma methodology.) Erratic and inconsistent behavior is to be eliminated. For example, an airline would rather hire a pilot that executes a within-safety-margins landing 10 out of 10 times rather than one that makes a brilliant approach five times and an unsafe approach another five. But, when it comes to creativity and innovation, say finding a new opportunity to improve the air travel experience or launching a new aviation venture, the same airline would prefer an ideator that generates one brilliant idea and nine nonsense ideas over one that generates ten decent ideas. In creative tasks, given that only one or a few ideas will be pursued, only a few extremely positive outcomes matter. Similarly, an ideator that generates 30 ideas is likelier to have one brilliant idea than an ideator that generates just 10. Overall, in creative problem- solving, variability in quality, and productivity, as reflected in the number of ideas generated, are more valuable than consistency (Girotra et al., 2010). To achieve high variability in quality and high productivity, most research on ideation and brainstorming recommends enhancing performance by generating many ideas while postponing evaluation or judgment of ideas (Girotra et al., 2010). This is hard for human ideators to do, but LLMs are designed to do exactly this— quickly generate many somewhat plausible solutions without exercising much judgment. Further, the hallucinations and inconsistent behavior of LLMs increase the variability in quality, which, on average, improves the quality of the best ideas. For ideation, an LLM’s lack of judgment and inconsistency could be prized features, not bugs. Thus, we hypothesize that LLMs will be excellent ideators. The purpose of this paper is to test this hypothesis by evaluating the performance of LLMs in generating new ideas. Specifically, we compare three pools of ideas for new consumer products. The first pool was created by students at an elite university enrolled in a course on product design prior to the availability of LLMs. The second pool of ideas was generated by OpenAI’s ChatGPT-4 with the same prompt as that given to the students. The third pool of ideas was generated by prompting ChatGPT-4 with the task as well as with a sample of highly rated ideas to enable some in-context learning (i.e., few-shot prompting). We address three questions. First, how productive is ChatGPT-4? That is, how much time and effort is required to generate ideas and how many can reasonably be generated compared to human efforts?