We note here that the average novelty of all ideas, irrespective of source, lies between slightly and moderately novel. While human ideas are a bit more novel, there is little reason to believe that novelty – being the first to think of an idea – leads to a significant financial advantage in domains associated with off-the-shelf technology, low entry barriers, and limited intellectual property protection. As such, from a commercial point of view, we don’t believe novelty provides sufficient advantage, if any, to overcome the productivity and quality benefits of the LLMs. Further, recall that novelty was not an explicit objective for any of our ideation schemes. In settings where novelty is the goal, it should be part of the prompts. Limitations Student Subjects It is possible that professional product innovators would generate better ideas than our students. However, that is not our intuition having worked in many product development settings. Many students in this course have gone on to be product innovators, sometimes based on ideas from the course tournament. We have not produced evidence that ChatGPT is better than the best human product innovators working today. However, we believe that we can claim conservatively that ChatGPT is better than many human product innovators working today and probably better than average. Thus, at a very minimum, an LLM could elevate the least capable humans to a better-than-average level of performance. Domain Our results are set in a common widely understood domain, for consumer products likely selling at a price less than USD 50. Presumably, there is a lot of commentary and data around these domains in the training data used by the GPT class of language models. As such, it is possible that in more specialized domains, say surgical instruments, our results will no longer hold with the current class of models. That said, to us, if this is true, this is likely driven simply by the paucity of training data. An organization looking for opportunities in these specialized domains should presumably be able to fine-tune language models with their own proprietary data and achieve comparable or better performance. Misbehavior Most language models do not provide any performance guarantees and it is possible they can generate offensive, illegal, or inappropriate ideas. Ideators using models for ideation should exercise caution. Of course, the same caution is warranted with human idea generators. Similarity For most innovation settings, the goal is to thoroughly explore the landscape of possibilities. Doing so enhances confidence that the most reasonable opportunities have been unearthed and considered. To this extent, we prefer a process that generates 200 diverse ideas to one that generates 200 highly similar ideas. Our analysis does not speak to the similarity or variability in the content of ideas. This remains an open question for further study.