Artificial Intelligence & Creativity: How I Wrote a Children's Book with AI Tools

1. How did the idea for the children's book with AI come about?
2. Limits of creativity and surprises - What are the limits of artificial creativity?
3. Tools for creating the children's book
4. Publishing on Amazon KDP
5. Conclusion of the experience

Can artificial intelligence be creative? Can it help with creative writing? Why does a text sound like it was written in Hogwarts? Why do AI-generated images always look different? Questions upon questions and a lot of trial-and-error - but after 3 days there is actually a new 80-page children's book in two languages in the Kindle Store.

In this post, the author Dirk Emminger explains his journey from creative writing with AI, from the idea to the creation to the marketing of his children's book. He shows the limits, but also describes the surprises he has experienced with artificial intelligence and different systems. If you also want to be creative with artificial intelligence, use this article as inspiration and get started!

Source: Dirk Emminger

How did the idea for the children's book with AI come about?

In November 2022, the company OpenAI published the chatbot OpenAI ChatGPT (Generative Pre-trained Transformer) and within days there was hardly any other topic. "Artificial Intelligence and Machine Learning" were not completely new topics, but what OpenAI had published for the masses was new.

ChatGPT is a AI text generator based on the technology of the Generative Pre-trained Transformer. It is capable of understanding and generating natural language. It operates as easily as a chatbot, and there is no need to install or parameterize anything anymore. Everyone could and still can access ChatGPT for free today and use the model. It quickly became clear what potential ChatGPT has in the content creation process and especially the creator community was bursting with articles, use cases but also with concern.

Inspired by an article on Medium I got the idea to try what limits an artificial intelligence has and did a reality check.

Source: Dirk Emminger

Limits of creativity and surprises - What are the limits of artificial creativity?

There are numerous also free AI text generators, such as ChatGPT, neuroflash or Jasper. I planned my story and started with a prologue and a disclaimer, which marked the beginning for my artificial intelligence book.

Source: Dirk Emminger

As can be seen from the screenshots and after a comparison with the final book, the majority of the text was adopted without corrections. This applies to the rather factual parts, but also to the creative sections of the book. Prompts (input prompts to the AI) should be very specific when writing creatively.

"Write me a children's book with four characters who experience adventures together and think up the first adventure" works, but does not produce a good result and a long test. What works very well, however, are individual passages like: "Write a dialogue in which a squirrel describes to a vacuum cleaner robot where it is hanging in the old oak tree."

What also works well is taking over the tonality of known authors. Example: "In the tone of J.K. Rowling".

In short, the plot of the book is as follows: The young dog Luzi lives near a nature reserve (the Wurmtal). As he is the only foreign dog in the area, he finds it hard to find friends and companions to explore the forest, the river and the small mountains in the area. Luzi's owner Eva has a lot of work in her own bakery and therefore buys herself a vacuum cleaner robot to clean the big house. In the morning, Eva is in her bakery and doesn't notice that something strange is happening. Dusty (the vacuum cleaner robot), who is more than just a normal vacuum cleaner robot, discovers Luzi and decides to reveal himself to him in order to experience adventures together and explore the world outside the house. On their forays they get to know the chicken Mafalda and form a gang. They learn a lot about the environment, plants, other animals and secrets in the Wurmtal.

By copying the paragraphs into Jasper, I benefited from the simple operation and editor mode. Through this, the text could be edited quickly and selected which type of style, tone and keywords should be included in the entire document. The simplicity of the UX made it easy to make changes without any problems. The more extensive the document is, the faster Jasper will learn its own way of writing and create new sentences in a similar style.

In the adventures in the Wurmtal, it was important, for example, to incorporate dialogues into the text. This requires the use of sentences like "explains Eva" or "asked Mafalda" to understand who is speaking at the moment. Once the AI tool had gotten used to this structure for a while, all newly created paragraphs automatically followed its flow, without anything needing to be adjusted. Surprisingly, there were only a few completely misunderstood texts.

The operation of Jasper is very simple. In the freeform document mode, I could enter a prompt in the editor with the "Composer" which is then executed directly by pressing CMD/CTRL + Enter. For example: "Luzy and Dusty meet Mafalda, the chicken". This way I could quickly formulate small sections according to my storyline.

Without ChatGPT and the help of Jasper this process would have taken several weeks. The fact is that a document of about 10,000 words with the tool was created in about 10 hours.

Source: Dirk Emminger

Especially in creative writing, AI systems have to be "fed" very strongly with a storyline to get a halfway reasonable plot. Business texts have slightly different rules - you can also create high-quality content with just a few tips. Read more about writing AI texts here. However, the biggest creative surprise came when I asked ChatGPT or Jasper for solutions. In the specific example, it was still unclear how Chip the Chipmunk was to be freed from his plastic net by the heroes. The idea that woodpeckers could save him came from ChatGPT.

Tools for creating the children's book

1. Generative AI with ChatGPT & Jasper AI

Generative AI, or the algorithms behind it, is able to generate new data that is similar to the data it was trained with. To understand how it works, an example from generative image generation. Here, an artificial intelligence is trained with a large number of existing images to generate new ones based on the patterns and relationships in the training data. In order to train cat images, we therefore need a very large amount of cats. The model is trained based on our cat images and tries to generate new images that are realistic and similar to our training images. If we only train the AI with black cat images, it will only be able to paint black cats. If we only give it a certain breed of cats, it will only be able to draw this breed.

This technology is also used in language generation. Here, a generative model is trained on the basis of texts to generate new texts that resemble natural language. The language model is so powerful that it can generate human-like text and also make context-related decisions within certain limits. Since ChatGPT was also trained very broadly with classical literature, there is nothing against using it for creative writing. However, since the software is purely a chat robot and not a special book AI, writing long texts is not comfortable. For my purposes, I needed more of a text editor with AI functionality. I found an AI software called Jasper.

Jasper is a platform that early on specialized in content generation with conversational AI models. With Jasper, you can create social media posts, blog posts, emails and much more. The special thing about Jasper is the included editor / boss mode. In this mode, you can write long texts easily and intuitively, add tonality and Grammarly as a spelling and grammar tool is also integrated. The editor mode in Jasper was perfect for rearranging and changing the text generated by Chat GPC. In addition, Jasper was able to add a tone and rewrite certain passages.

Unfortunately, it quickly becomes apparent that neither ChatGPT nor Jasper works well in German. ChatGPT can communicate in German, but after a few tries it became apparent that writing in English is the better choice.

2. Grammarly and DeepL for translations and corrections

In retrospect, the "production" phase was the most pleasant. Because what followed was the first English correction and the translation into German. For the English correction, Grammarly was used. Although Grammarly is already integrated in Jasper, in order to get a nice, similar sounding document, the entire text was checked separately once again. Rewording I avoided as much as possible, because it was supposed to be an experiment to see how good AI is today and how fast it works when writing a whole book.

The next step, transferring the text into German, was a nightmare. The corrections took a lot of time due to the complex grammar. The book was translated with DeepL. The translation was readable, excellent and fast, but the German punctuation (which I admittedly am also not good at) required a lot of manual intervention.

I had to call my phone joker, a friend who is an agency owner and professional copywriter several times to think about why this is so. The explanation that seemed most obvious to us in the end was the use of "Write in the tone of J. K. Rowling". This style often led to very "flowery" descriptions and long complex sentences.

3. AI Artwork with Midjourney and Challenges

In choosing a AI tool to create the artwork for the project, I eventually settled on Midjourney. An overview of the available AI image generators can be found here.

Midjourney is an AI tool specifically developed for the creation and editing of digital art. It uses advanced machine learning algorithms to create unique artworks. It offers a wide range of features and customization options that allow the user to create everything from abstract graphics to realistic paintings.

The main factor influencing my decision were the costs. Midjourney is one of the most affordable tools on the market and allows me to create high-quality artworks at a low cost. The first +- 25 pieces (25 minutes GPU time) are free and the next available option is a basic subscription for $10, which has a value of +- 200 minutes per month. Please note that the basic plans are public and everyone can see your prompts and pictures. You can expand your subscriptions, though, with options for more GPU time and privacy features.

The operation is done via the message line of a Discord server. A big advantage of Midjourney, apart from the low costs, is the large number of tutorials, examples and tools.

When creating images for the book, I used a combination of keywords and renderers, namely Octane Render and Futuristic, to create frontal shots with 50mm lenses that were then rendered in 3D comic style to generate very strong emotions.

Source: Dirk Emminger

These strong emotions in combination with the very realistic rendering look great, but they almost drove me completely insane and caused another three hours of tutorials.

The problem here is called "consistency". This refers to the ability of the AI model to generate images that are uniform and make sense in the context of the task. If, for example, the task is to generate images of hamsters, the model should not generate images of cats or birds. AI models can sometimes generate images that do not match the task, leading to images that are confusing or do not make sense.

Another problem is that the model may generate images that are too similar to each other and do not offer diversity. This can be a problem as it can lead to a lack of variation in the generated images, which can be boring or uninteresting. Because I created the personalities of my characters with the command "hyperrealistic", this led to Eva looking completely different depending on her pose. And the other characters, Dusty, Luzi, Mafalda, could not appear together without Photoshop, for example, on the cover.

Source: Dirk Emminger

Technically speaking, there are some solutions. You can, for example, work with reference images and seed commands, but in reality, these photorealistic comic characters are not suitable for a continuous picture story with today's state of the art.

Publishing on Amazon KDP

For the self-publication of the children's book, I chose Amazon Kindle. This was the easiest and fastest way for me to get a digital version of the story. The entire process, from start to finish, was surprisingly simple and straightforward. Amazon Kindle Direct Publishing (KDP) is a platform that allows authors and publishers to self-publish their books in digital and print formats on the Amazon marketplace. With KDP, anyone can upload their book as a PDF or Word document, set the list price, and offer the finished book for sale in just a few steps.

I was very surprised at how extensive the self-publishing community is and that there are countless high-quality sources on the topic of "self-publishing" via Amazon. If you also want to self-publish a book, I recommend starting with the Amazon KDP help page, where you can find all the important information.

One of the decisions that needs to be made is whether to participate in Amazon's KDP Select. KDP Select is an optional program for authors and publishers who want to make their books available on Amazon. If you opt for KDP Select, the book is also included in the Kindle Unlimited (KU) and Kindle Owners' Lending Library (KOLL) programs. This is Amazon's book flat rate.

Advantages of Self Publishing via Amazon

• Increased visibility through top positioning in the Kindle Store
• Accessibility through Prime or Kindle Unlimited subscriptions
• Use of promotional tools such as free book promotions and countdown deals
• You receive royalties for every page of your book read

Disadvantages of Self Publishing via Amazon

• Restriction for 90 days to Amazon Kindle
• Limited advertising opportunities outside of Amazon Kindle
• Low royalties per sale compared to other distribution channels

Since I don't expect to be listed with my first title directly by a publisher and the calls from agents are still pending, I can live with the exclusivity of marketing on Amazon, at least for the first 90 days.

Conclusion of the experience

Creative writing worked very well for me with AI and the new generative algorithms. I would even go further and claim that an artificial intelligence as a helper in all creative fields can be a door opener. Without AI, it would have taken me massively more time to write a book and I would probably never have gotten the graphics/illustrations.

I would now never claim that my or an AI book is particularly good or particularly bad and that the world has been waiting for my work. But it has shown me once again that AI and technology as such is an enabler. The development around AI and also the development of new creative tools is so fast that probably writing a book is just the beginning.

The release of ChatGPT and the amazing progress from MidJourney v3 to MidJourney v4 were the triggers for this experiment. The development is so fast and almost every day I discover new tools in my news feed. The field of generative AI tools is currently very fragmented, but it is to be expected that in the near future a consolidation to complete E2E software solutions for various workflows will take place. An example of this is the integration of DALL-E 2 into the web application Microsoft Designer, similar to Canva Pro. Also Notion already offers numerous genAI use cases in various areas.

Even though there are currently still different specializations of genAI according to creative requirements. I think the future will bring us complete solutions that have fewer media breaks and where whole teams can work together on E2E genAI applications. The chances that part 2 of the "Adventures of Wurmtal" will be directly released as an audiobook and animated film are therefore not bad.