Unlocking the Power of Metaphor: AI’s Role in Ideation
OpenAI founder Sam Altman once stated that the most common question startup founders ask is how to generate ideas for their ventures. New ideas are essential not only to begin but to sustain a business amidst the inevitable market shifts. By 2023, a report noted that 44% of companies were utilizing AI to identify new innovation themes, domains, adjacencies, and technologies. Companies successfully infusing AI into innovation processes often transform into “idea generation powerhouses,” producing over five times as many ideas as their counterparts.
Despite this abundance of ideation, Harvard professor Theodore Levitt pointed out that true innovation is scarce due to the practical challenges of implementation. Innovation, according to management expert Peter Drucker, comes from a thorough analysis of opportunities influenced by broader social or demographic trends. In essence, innovation charts a feasible path to implementing ideas.
Research from INSEAD highlights the value of diversity in idea generation. Ideators who propose a variety of ideas, including a mix of brilliant and less feasible ones, are more desirable than those with just decent ideas. A study from Cornell revealed that GPT-4 was able to generate ideas much more efficiently and cost-effectively than students from a top US university. Among the top forty ideas produced in an experiment, GPT-4 contributed to thirty-five. Furthermore, a review by seven different AIs consistently agreed with business strategy experts on the feasibility of GPT-4 generated business models two-thirds of the time.
Researchers from Wharton Business School have explored the use of Large Language Models (LLMs) like GPT-4 for generating unique and diverse ideas. By employing a chain-of-thought (CoT) command sequence, these LLMs can examine problems through a series of microtasks and identify potential reasoning flaws, effectively simulating a conversation. CoT prompting increased the number of unique GPT-4 ideas by 27%, closely mirroring the diversity found in human group ideation.
In 2023, Stanford University researchers discovered that, with the CoT technique, GPT-3 could comprehend both literary and non-literary metaphors, such as “A smile is a knife” or “A train is a large worm.” In another study, GPT-4 interpreted literary metaphors from Serbian poetry more accurately, according to human poets, than a group of college students. Such metaphors involve cross-domain mappings, linking distinct subjects in a way that appears novel or insightful.
The use of metaphorical mental models has been linked to business success, where analogies from one scenario can inform decisions in another—a process known as second-order learning. Noted business figure Ray Dalio mentions that a lack of second-order thinking can lead to poor decisions and result in complacency with inadequate solutions. AI’s ability to engage in metaphorical reasoning through CoT prompts enables the identification of previously unnoticed connections between seemingly unrelated scenarios.
Nevertheless, achieving optimal results will likely require the combined efforts of various AIs, supported by expert human prompts. Current research indicates that while progress in AI is rapid, GPT-4’s ability to handle abstraction at human-like levels still requires development. Though AI needs human guidance to augment its reasoning, its role in core startup teams seems promising, especially as new language models tailored to regional languages emerge.
