Sparking creativity: Using LLMs in extension

In this episode, we’re exploring large language models (LLMs) as a tool that can help improve our writing process as enablers of change. LLMs, specifically those designed to generate creative “sparks,” can be a game-changer for us, making our science communication more engaging and effective.

Just so we all know what we’re talking about, let’s do a quick overview of Artificial intelligence or AI. This is the broad field focused on creating systems that can perform tasks requiring human-like intelligence, such as reasoning, perception, and decision-making. Within AI, machine learning (ML) refers to methods where systems improve their performance by learning from data rather than being explicitly programmed. 

Generative AI (GenAI) is a specialised subset of machine learning that creates new content—text, images, code, or audio—based on patterns learned from large datasets. LLMs are a type of GenAI specifically designed for natural language processing, trained on vast amounts of text to generate human-like responses, translations, or summaries. This hierarchy shows AI as the overarching concept, machine learning as the method, Gen AI as a focused application, and LLMs as a specialised form within that application.

LLMs work by generating text based on the query they receive. They predict the next word in a sequence by considering the context of the words that came before it. In previous episodes we’ve discussed How enablers of change might use ChatGPT which is an LLM, and we also gave Seven suggestions for making better use of ChatGPT. We made the point that LLMs are like an eager intern with an overactive imagination—very clever but also very naïve. So we have to be careful when using them, and never fully trust what they generate. While ChatGPT is currently the best known LLM, others include Claude, Gemini, and Copilot. 

As enablers of change we’re often required to create engaging science communication. We aim to explain complex concepts to a range of different audiences, from farmers to advisors or policy makers. A clever use of LLMs is to ask them to generate ‘sparks’—sentences or ideas that inspire and guide us in writing engaging content. Instead of starting from scratch and facing the dreaded Blank page syndrome—which we mentioned in a previous episode—we can use sparks generated by LLMs to find different angles or explore alternative ways of describing science to our audience.

Say for example we need to write about a new soil health initiative. By putting key topics into the LLM, we can receive several sparks that offer different perspectives or interesting facts about soil health. These sparks can range from technical explanations to real-world applications, providing us with a variety of starting points to build our message. 

Imagine we’re drafting a newsletter article about agroecological crop protection. We could get an LLM like ChatGPT to finish the sentence for us. So we could type in “Agroecological crop protection involves…” or “Growers use agroecological crop protection to…”. The LLM might generate sparks such as “Growers use agroecological crop protection to work with nature rather than against it in managing pests, weeds, and diseases.” Here’s a good tip: use the Try again button that is displayed beneath the response to get ChatGPT to offer alternative ideas. We can keep clicking this until we find one that resonates!

The sentences that it generates can then act as a springboard for us, helping us explore different ways of communicating our message effectively and engagingly. These sparks not only provide some content but also help us present it in a way that’s easy to understand and compelling for our audience. These sparks can also help us see our topics from different angles, ensuring that we cover aspects that might otherwise be overlooked. For example, if we’re writing about pest management strategies, a spark might prompt us to consider environmental impacts, economic benefits, or even farmer testimonials. 

We found a recent paper from academics at Columbia University in the USA, that described these sparks. Their research involved 13 PhD students and found that participants appreciated the diverse and coherent ideas provided by the LLM, which often introduced new angles they hadn’t considered. While sparks can significantly enhance our writing, it’s essential to remember that they are tools to aid our creativity, not replace it. We still need to apply our expertise and knowledge to refine and adapt these sparks to fit our specific context and audience. This combination of human insight and machine-generated inspiration can lead to higher quality science communication. The exploratory study highlighted how LLMs could assist in the creative process, and showed that while sparks were useful, the writer’s input remained crucial.

The study revealed that sparks were particularly effective in helping participants articulate their thoughts concisely. This ability to generate succinct and precise language can save us time and effort, allowing us to focus more on the content’s overall structure and flow. For instance, when explaining a complex topic like integrated pest management, a spark might generate a sentence like ‘Integrated pest management combines biological, cultural, and chemical practices to control pest populations effectively,’ which we can then expand with additional details and examples.

Let’s not forget the importance of addressing common reader perspectives. Sparks can remind us to consider what our audience might already know or believe about a topic. For example, if we’re discussing the adoption of renewable energy sources on farms, sparks might highlight the economic benefits or environmental impacts, ensuring that we address these points, acknowledging different reasons for considering these ideas.

The research paper we mentioned earlier also highlighted some challenges and areas for improvement in using LLMs for writing support. One common issue was the occasional generation of inaccurate or irrelevant sparks, sometimes referred to as hallucinations. As enablers of change, it’s our responsibility to verify the information provided by the LLM and ensure its accuracy. This process of critical evaluation not only maintains the integrity of our content but also enhances our understanding of the topic.

So we think the use of LLMs to generate creative sparks offers a promising tool for us as enablers of change. By leveraging these models, we can enhance our writing, making it more engaging, informative, and relevant. However, it’s essential to remember that these tools are meant to complement our expertise, not replace it. We must critically evaluate and adapt the generated content to fit our unique context and audience.

Well, you’ve read our thoughts, and now we’d love to hear yours! Drop a comment below and let us know your experiences with using LLMs to generate sparks. Got any tips or ideas? Share those too! We’re looking to make this a lively conversation, so join in with your thoughts and suggestions.

Thanks for your interest in this Enablers of Change episode. Don’t forget to subscribe to our newsletter to stay updated on new episodes. If you enjoyed what you heard, spread the word to your friends so they can join the conversation too! 

Resources

Gero, K. I., Liu, V., & Chilton, L. (2022). Sparks: Inspiration for science writing using language models. In Proceedings of the 2022 ACM Designing Interactive Systems Conference (pp. 1002-1019). Available online.