How to create content using AI
Author: Edd Baldry and Andy Gordon;
Reading Time: 5 minutes
We've made this resource open. You are free to copy and adapt it. Read the terms.
In this article we’ll share some examples of how you can use AI content tools (also known as Large Language Models) when creating content. This can bring you operational efficiencies, increased fundraising scope and more organisational impact.
For those short on time, as of September 2023, our recommendation at Torchbox is to use OpenAI’s GPT-4 for generating text. You can access GPT-4 by subscribing to ChatGPT+.
Our recommendation for working with images is to use Adobe’s Firefly. For music, voice generation or transcription there’s no clear leader. We’ll dig into details within the article.
About LLMs
Artificial Intelligence (AI) is becoming a catalyst for charity innovation. At the forefront of this change are Large Language Models (LLMs). LLMs are a type of AI that uses deep learning algorithms to mimic human intelligence. They use statistical models to analyse vast amounts of data then use natural language processing (NLP) to create content based on what they’ve learnt. LLM examples include ChatGPT, Jasper and Copy.AI. Charities are learning to use these tools to generate content.
Writing new content
As a charity you need your communication to be high quality and accurate. This could be a social media post, crafting a case study or updating website copy. With the right prompts and guidance, LLMs can excel in this area.
You’ll get out what you put in. We use a 20:60:20 principle where the writer does the initial thinking about prompts, and the final polishing, but the LLM does the work in the middle. Here’s how:
Bad approach: use a single prompt “Write an article about how non-profits can use generative AI to create content”. This prompt will get a very generic response. It is a bad prompt because there hasn’t been enough initial thinking.
Better approach: build the article iteratively, the same way you would work with another human. Use an initial prompt: “I need to write an article. It should be about how non-profits can use generative AI to create content. What do you think the five most important topics to cover are?”
The five topics returned give you a starting point to build upon. You can write more prompts for those topics that seem most relevant. This guides the LLM towards a better result.
Rewriting existing content for a different context
Large language models can really help rewriting content. We often find ourselves re-writing content for different platforms (for example LinkedIn, X and Instagram). Here’s an example of a prompt for LinkedIn:
“The below is a case study from (charity x) about (subject). Please can you summarise the case study for LinkedIn. I want the LinkedIn post to have a big impact. It should conform to their best practices. There should be max 200 characters per sentence. There should be a maximum of three sentences to a paragraph. The first 135 characters need to be an exciting hook to entice the user to read more (e.g. a question or bold statement). The post should be about what the reader can learn or use in their work. Please ensure the post is no more than 500 words.”
Note how specific the instructions are. You’ll get the most out of LLMs by being explicit. To get the most out of an LLM you need to understand that context is key.
Visual storytelling
A picture is worth a thousand words and a video is worth a million. There are new opportunities for charities to create their own high quality visual content. These can be done within tight timeframes and at relatively low costs. Examples of text-to-image or text-to-video generators are Midjourney, Pictory and Stable Diffusion.
Adobe’s Firefly is particularly interesting as Adobe has promised that it is safe to use commercially and is now built into Photoshop. It produces strong results and has an easy-to-user interface. From our perspective it’s the obvious tool for non-profits to use right now.
There are some great examples of AI generated visual storytelling. Two we particularly like are the WWF’s future of nature campaign (archive) focused on imagining dystopian and utopian futures. The second is Friends of the Earth’s recent social media post reimagining nature filled spaces. This was for their Postcode Gardener programme with the Co-operative.
Know your LLM risks
Charities should understand that there are risks associated with using these tools. We don’t want to scare people from using them. But it’s important to be aware of how they work and the risks you might be taking on.
The risks stem from the fact that they’ve been trained on material created by humans. This means the materials LLMs analyse contain human falsehoods and inaccuracies.
LLMs, to date, have also been trained to express themselves clearly and confidently. This is a problem if they tell you something that isn’t true, because they will make it sound like it is true! This is reinforced by our tendency as humans to trust machines to give us the correct answer.
If you’re using large language models to generate or synthesise you should be aware of nine misconceptions about LLMs. You want to avoid being like the lawyer in New York who tried to save time using ChatGPT to prepare his court-case.
What to use
As at September 2023 Torchbox recommends:
- OpenAI’s GPT-4 for generating text.
- Adobe’s Firefly for working with images.
—
Credits: Thanks to Andy Gordon and Ed Baldry of the digital agency Torchbox for contributing this article to Catalyst.
Image: courtesy of Torchbox.
Commissioned by Catalyst
