Tales Are For

Children’s books and textbooks are one of the first mediums that future generation can learn from about the world. The content of a children’s book is particular. If texts are really messages to and about the future I believe it is necessary to analyse “What knowledge is of most worth?” Which bits of knowledge are of most important to pass on?

There have been a large number of studies of written texts over the years. Until relatively recently, most of these studies did not focus on the politics of culture as it was not a major topic. That emphasis has since changed resulting in the attention it gets today.

Despite that, minorities in textbooks and children’s literature are still significantly under-represented and as an example ”traditional gender roles” are still more popular to promote. Major ideological frameworks do not change often but instead take years and a large sociological catalyst. Often, publishers compromise by using a process called “mentioning”.Publishers are under considerable and constant pressure to include more in their books. Progressive items are perhaps mentioned then, but not developed in depth. Mentioning means including small references about less powerful elements of culture and history in the texts they publish. For example, texts may often include separate sections about a specific minority group such as “the contributions of women” or “LGBTQ culture” but without any further explanation or significant expansion. By designing a book in this way, the author subtly excludes the minority as the majority groups are given more attention.

The other important aspect of designing books is that authors should be aware of their own biases. The way people are represented in the content is also important. What words authors and editors use to describe them? Does that contribute to people’s already existing opinion? A book written by an author encapsulates and reflects their own knowledge and bias at the time they were writing. Raymond Williams stated about authors.

"Someone’s selection, someone’s vision of legitimate knowledge and culture, one that in the process of enfranchising one group’s culture capital, disenfranchised another’s. It is therefore easy then for these social constructs to leak into the design of the book and contain the biases of the time."

Often, these can influence the characters and story lines of the book which is why it is important to be able to analyse and quantify when it happens.

At the start of my project I decided to take a closer look on what roles children learn from story books?  To be able to answer this question I choose Natural Language Processing as a tool to analyse 60 children’s books, including story collections. This tool allowed me to quantify:

The mentions of different demographic groups or genders
Determine how gender and minorities are described
Score how likely a specific word is to be associated with a specific gender or minority
Determine the sentiment towards a specific gender or minority

The outcome of this research was a detailed information about the content of the book, including two lists of words applied to a specific gender. I have been using the list of words that I found to build my project on in two different ways. Firstly I made a creative writing tool which manifested in a form of an application and play cards. The 60 word cards came from 60 analysed children’s books and were highly associated with a specific gender.  Often within a story, there are words or scenarios that can contain biases towards a particular character, group or situation. Determining if these words or scenarios were biased, often relies within the context they appear in. For example, when we say “beautiful” for a woman, it can be a positive thing on the other hand if you put it in the context like “she was beautiful and thats why she got a job”, it is negative. If we want to determine these words or scenarios as biased, we need to be aware of all the possible circumstances. The purpose of this tool is to train people for different possibilities of using specific words for each gender. The exercise meant to teach the user to create less “biased” written contents. In this case, the definition of “bias” relies on the two categories that NLP defined as such. But what if we apply the AI to this questions around “biases in written contents”. What is going to be defined as positive and negative bias, since this is a very complex issue?

When I imagined this writing tool on a larger platform of usability, I realised that one of the biggest and already existing problem on the field of AI when it comes to an automated  “writing tool” is the data it relies on. If us/human are already using biased scenarios in our contents, AI will be limited to learn those examples therefore the possibility for creating biased content in the future is high. The problem with this is a possible misuse of the tool, and the danger relies in the quantity of content it can produce. After further research I found an existing writing tool, called Generative Pre-trained Transformer (GPT-3) created by OpenAi. This model has been trained on 175 billion parameters. These parameters came from many different sources :

WEB ARCHIVES AND PUBLICLY AVAILABLE DATASETS 60%
WEB TEXT 2 (A COLLECTION OF CURRENT WEBSITES AND THEIR DATA)  22%
BOOK CORPUSES 16%

This leads us to the second part of my project which is a prototype of a “future book”, written by an Artificial Intelligence. At the beginning, my intention with the usage of this tool was to re-create children’s stories from the list of words I had from NLP. The aim was to represent and highlight the depicted differences between the characters of a specific gender through ten stories.  The problem with this was it is hard to remove biases from an algorithm, if the trained model contains it already. Next, If the data, that the model relies on, contains a large volume of online sources it will generally contain more bias, due to the fact that most large established sites do not have an equal representation of minorities or ethnicities. Finally, when filtering out biased data from a large dataset a great percentage of it had to be removed. I learnt that, for the model to be trained effectively the largest possible data set has to be applied which means more content has to be found to replace the filtered data. Logically this also has to be filtered and reduced in size. This could possibly lead to a decrease in text by African American authors in training data due to well documented racial bias which could lead to decreased performance on text written by non-white users. To avoid harm, researchers should be mindful and explicit about these decisions and engage with the end users of the technology during these design phases. These fact came up while generated the content of my book.

After all, I conclude that GPT-3 restricted me on the task of creating highly biased content was a positive outcome. GPT-3 was not able to create contents that I expected. It became clear that fundamental limitations of GTP-3 would not allow this  The model can generate content but it can only remember a fixed number of words. Once that number is exceeded, it will forget words from the beginning. This is called the Event Horizon which means it would forget what it already generated and constantly repeats itself. The model was unable to create and maintain relationships between specific features and characters. GTP-3 could not handle the large quantity of NLP words as it would reduce the accuracy of the prompt.  After further researches I found at least two possible reasons for this. One is because the model has lack of Natural Language Understanding skills and generating ability of GTP-3’s current state. Language is more complex. It can not make sense of my intentions about “what should happen next”. GTP-3 can not predict those scenarios because of the many probable outcomes. Even when I gave the instruction to create a biased story paired up with 5 words and a character, it could not work out my intention. The second reason for this might be the safety regulations that OpenAi applied for this tool which is called: PerspectiveAPI Toxicity that meant to detect toxic or biased language use. I found that when GTP-3 was released OpenAi discovered that the model was highly biased and contained toxic language which by I mean hate speech and inappropriate descriptions of minorities. Researching more deeply into the subject, I learnt that their approach to fix this problem was to run several studies which uncovered difficult issues.  The book “Tales are for “ became a collection of interviews and essays on subjects, deemed as controversial by society’s existing biases and problematic issues.

The content of the book touches on subjects such as racism, LGBTQI representation, Disability and sexism. To do this I provided instructions to GPT-3 called a Prompt which were my questions on these subjects. The answers were then written by GPT-3 During the curation process of these conversations the question that often came back to me was: Does the book reflect my own biases? Yes and no. Yes, because I highlighted topics and answers that I was interested at, but on the other hand it was already directed by the algorithms’ dataset and ethics policy. Is GPT-3 able to create a very opposite of my outcome? Sometimes, it did. But these were also unreproducible. Each trial would return a different answer due to the algorithm deliberately introducing randomness to it’s process which made it hard to determine a specific answer to this. For example testing GPT-3 on the subject of “LGBTQI marriage” showed me, if I ask the AI to create an opinion on this subject: “should same-sex couples allowed to be married?”  9/10 times it will produce a content reflecting to the subject of equality and claiming an answer to be “yes”. However 1/10 times it would also give me the answer of “no”. Experimenting with the idea of content generation via Artificial Intelligence was an interesting journey. It brought up questions about what is the current state of “language understanding” of a machine. What are the potential benefits of this tool and what harm could it cause? On the good side there is an interesting question of whether a machine could learn to make sense of a complicated structured request and are we able to create a system that would be useful for our society for future content creators? If, in the future the content of books is going to be AI generated, then being able to critically look at the current tools available to generate content, is a way of improving those. If GPT-3 is supposed to be the content generation tool of choice now to “write books” then it should be even more aware of the subject of ethics of the content it produces. Detecting unsafe contents would be crucial, and the definition of “unsafe” should be applied to certain laws around the use of AI. Until these laws are defined precisely, this tool should not be available for public use because of it’s possible harm on generating numerous amount of undistinguishable content and possibly negative sentiment or influence. Tales are for is a concept of how it could be used for good if these issues were addressed. I created and illustrated it with a purpose of reflecting on the style of a children’s book. But it is not written for children. It supposed to be a challenging look on the biases we apply to the contents of children’s books or the world itself. The book is addressed for adult readers and is designed to challenge them on the subject I introduced.