Make AI more creative so you can come up with better ideas
Innovation and creativity are essential for solving problems and developing new products and services, but generating fresh ideas can be challenging. Tools based on LLM models like ChatGPT or Copilot are helpful in this task and proven effective at generating structured ideas. However, they have certain limitations in producing truly diverse ideas and often generate overly similar proposals when not properly guided.
These limitations arise because LLMs optimize their responses based on the most statistically probable patterns in their training data, often failing to explore a wide and varied set of solutions to generate unconventional ideas.
To address this limitation, a group of researchers compared the ideas generated by an LLM (GPT-4) with those proposed by a group of students. The results show that while students naturally produce more varied ideas, LLMs can achieve a similar or even greater level of idea diversity when a well-designed prompt structure is employed.
Prompting techniques allow for the exploration of more creative and innovative solutions, generating less conventional ideas.
Prompting techniques to increase idea diversity
The researchers evaluated a series of strategies designed to maximize LLM models' creative potential:
Minimal prompt
This approach provides basic instructions, such as 'Generate a list of ideas.'
While simple and quick to implement, it tends to produce homogeneous results due to the lack of specific guidance. Models respond to common patterns without exploring unusual or complex ideas.
Simulated personas
This method involves instructing the model to adopt different perspectives, personalities, roles, or thinking styles, such as 'Think like a visionary scientist' or 'Imagine being an eccentric artist.' For example, emulating figures like Nikola Tesla, Joan Miró, or Ada Lovelace can lead to significantly varied creative results. In a business context, this technique can be applied to imagining solutions from multidisciplinary perspectives.
Established methodological frameworks
Incorporating creative techniques like Stanford’s Design Thinking into prompts provides a structured framework to guide the creative process. This improves diversity and ensures the coherence and feasibility of the proposed ideas by leveraging a widely recognized methodological approach to creativity.
Chain of thought (CoT)
The CoT approach breaks the main task into smaller, sequential subtasks, such as initial idea generation, idea refinement, and detailed description. From an algorithmic perspective, this segmentation allows the model to focus on each specific stage of the problem, with the possibility of integrating intermediate evaluations that continuously improve the results.
This method increases thematic dispersion and the number of unique categories, fostering greater diversity and reducing generic patterns, proving highly effective at generating new ideas.
■ Applied to other fields, such as education, the CoT technique can first generate learning content, then refine it to adapt to different age groups or specific needs (such as children or older adults learning digital skills), and finally detail the activities, materials, or pedagogical approaches suitable for each group.
The CoT prompting technique can be adapted to various industries and needs, enabling businesses to explore valuable and novel solutions.
Practical example of applying the CoT technique
For a practical example of how to apply the CoT strategy in a real-world context, researchers used the example of creating a product costing less than €50 and aimed at university students:
First stage: initial generation
- Prompt:
> Generate 50 ideas for physical products aimed at university students that can be sold for less than €50.
- Example response:
Thermal mug.
→ This step leverages the model’s ability to provide quick and general responses.
Second stage: idea refinement
- Prompt:
> Make these ideas bolder and more creative.
- Refined result:
Rechargeable battery mug that cools or heats beverages to user preference.
→ This step adds a distinctive element, ensuring unique or uncommon ideas.
Third stage: a detailed description
- Prompt:
> Describe each idea in a 40-80 word paragraph.
- Final example:
Smart thermal mug with an integrated battery that adjusts the temperature to keep coffee hot or cool drinks during summer.
→ This step helps concretize and contextualize the idea, exploring its practical applications.
Using these methods appropriately, LLMs can match or even exceed human idea generation capabilities.
Analysis results
The study concludes that the CoT technique (as applied to GPT-4 in this case) is particularly effective due to its ability to break down complex tasks into manageable subtasks, enabling models to explore a broader spectrum of solutions at each stage.
Compared to other techniques, CoT takes advantage of sequential processes that facilitate the progressive refinement of ideas, increasing the likelihood of generating uncommon responses through iterative evaluations and restructuring based on feedback provided by the model.
Additionally, combining multiple prompting techniques amplifies the range of solutions generated.
Conclusion
The strategic use of prompting techniques in LLM models transforms their capacity to generate ideas, enabling exploration of broader and more diversified creative territories across an increasing number of business, production, and academic domains.
These methodologies enhance performance in specific tasks, and researchers conclude that, when properly utilized, AI can match or even surpass human capabilities in idea generation through the design of elaborate and detailed prompts that guide the model toward novel results.
The key lies in designing clear and structured prompts that steer the model toward less frequent outcomes.
This approach is not limited to creating products for university students. Instead, it can be adapted to various industries and needs, enabling companies to explore a vast territory of potentially valuable and innovative solutions.
■ For further reading, see the study Prompting Diverse Ideas: Increasing AI Idea Variance by Meincke, Mollick, and Terwiesch.