AI Tools

Using AI to help create Easy Read summaries

After thorough testing of using publicly available AI models to generate Easy Read content, we have found that two of these models far outshine the rest when it comes to creating Easy Read text. Namely, GPT-4 and ChatGPT. So we'll explore these two options below.

GPT-4

GPT-4 is the newest and most advanced language processing model developed by OpenAI. It stands for "Generative Pretrained Transformer 4" and is capable of generating human-like text, completing sentences, answering questions, and creating images, among other things. It is trained on a huge amount of data and uses a special way of learning that helps it understand and generate natural language with great accuracy.

GPT-4 can help create Easy Read documents by summarizing information in a way that is easy for a second grader to understand. This means using simple words, short sentences, and examples to explain things. To do this, we can use the 'Summarize for a 2nd grader' feature on the OpenAI Playground.

OpenAI's GPT-4 has a subscription model and an API access. GPT-3 is also available as a free-to-use chat assistant called ChatGPT. Bing chat (also known as Microsoft Copilot) is another AI-powered chat assistant that uses GPT-4, but also has access to Bing search, which allows it to get more up-to-date information than GPT-4 alone. Bing chat/Copilot can answer questions, compare products, write drafts, and even create images based on your words. You can use Bing chat in the sidebar of Microsoft Edge or in Chrome and Safari mobile browsers.

You can find out more about GPT-4 here: https://openai.com/research/gpt-4

You can find out more about Bing chat/Copilot here: https://copilot.microsoft.com/

Using GPT-4

We can use GPT-4 to generate text we need, also we can use it to finetune the results, as GPT-4 remembers earlier parts of the conversation as well.

OpenAI is currently the leading company in the field of large language models, and we believe that these tools are the best anyone can use currently to create Easy Read text. We can use Bing chat/Copilot to access GPT-4 without a subscription.

When GPT-4 is being asked: "what is an easy read text", it can answer with an explanation. Therefore, we can simply prompt it with "Rewrite this text as Easy Read: [TEXT]". We can also use "prompt engineering" to improve the response. An example of prompt engineering would be to ask GPT-4 to give a good answer, break down how it will answer into steps, give the model tools to better understand or accomplish the task. For example, this prompt will yield a better result than the above example: "Imagine you are a gifted writer who understands how humans process information. You have been given the task of writing an Easy Read document using the information on the https://easyread.demcloud.org/. Break down the steps you would use to generate an Easy Read document and use those steps to generate an Easy Read document summarizing this in an easy-to-understand way: [TOPIC,LINK,TEXT]". Here is an example of using this sample prompt to generate an Easy Read-formatted document for the U.S. Consititution using ChatGPT.

Limitations of GPT-4: Sometimes the AI gets "offended" by what the user asks, for example it sometimes claims that it's not supposed to write text for children to understand, or for intellectually disabled people to understand. It is certainly capable of doing so, and its limitations can be circumvented by using different wording, synonyms, or similar-but-not-the-same prompts.

Using AI to generate images

As AI image generation services just started to get good, there are limitations that you need to be aware of. The AI is notoriously bad at writing text, drawing hands and creating realistic faces, therefore we recommend to avoid these areas, or prepare to finetune the results with these limitations in mind.

Some other tips to create AI images are:

1. Start by providing clear and detailed descriptions of the images you want to generate. The more information you provide, the better the AI system will be able to understand and create the images.
2. Consider including information about the objects, colors, textures, lighting, and other visual elements that you want to include in the generated images.
3. Use specific and descriptive language in your descriptions, rather than general or vague terms. For example, instead of saying "a blue sky," you could say "a clear blue sky with wispy white clouds."
4. Avoid using ambiguous or contradictory language in your descriptions. For example, if you describe an object as "big" but then later describe it as "small," the AI system may not be able to generate a realistic image.
5. Experiment with different combinations of descriptions and settings to see what works best for your specific goals.

The best 3 image generation processes right now are DALL-E 2, Midjourney and Stable Diffusion. While DALL-E2 is a proprietary software with a subscription fee, Midjourney can be used for free for some extent, and Stable Diffusion can be used free and can be run on anyone's computer with a strong enough graphics card.

• DALLE2: https://openai.com/dall-e-2/
• Stable Diffusion:https://beta.dreamstudio.ai/dream
• Midjourney: https://midjourney.com/

Style

Achieving consistent style when generating multiple images is hard right now, but as AI tools understand styles and authors, they can be prompted with these modifiers in mind all the time, so there is a degree of style consistency that can be achieved. To maintain style consistency, we suggest using one image generation service throughout the whole process, as different services are trained on different datasets and use their own image generation machine learning models to create pictures.

Some other ways to maintain style consistency:

1. One way to maintain style consistency when generating images with AI tools is to use a technique called style transfer. This involves using a pre-trained model to transfer the style of a reference image to a target image, while preserving the content of the target image. This can help ensure that the generated image has a consistent visual style with the reference image.
2. Another way to maintain style consistency is to use a consistent set of design principles and rules when creating images with AI tools. This can help ensure that the generated images have a consistent look and feel, even if they are not directly based on a specific reference image.
3. Additionally, it can be helpful to review and edit the generated images manually to ensure that they meet the desired level of style consistency. This can involve adjusting colors, composition, and other visual elements to ensure that the generated images are visually cohesive with each other and with the overall aesthetic of the project.

For example, we can tell the AI to use this prompt: "A person using a smartphone to share something, big red x and exclamation sign in the style of Picture Communication Symbols"

Picture Communication Symbols (PCS) is a system of symbols that can be used to represent words and ideas. These symbols can be used to create visual supports, such as communication boards or social stories, to help individuals with intellectual disabilities or young children understand complex information.