The Spirit Machine

An AI-based art installation writing texts inspired by the work of Maria Montessori. For Radboud University, The Netherlands. (2021 - ongoing)

Artist: Jeroen van der Most

Co-creator: Gabriëlle Ras, PhD candidate AI

Maria Montessori

The Spirit Machine is an art installation that revolves around the academic work of Maria Montessori, one of the most inspiring thinkers for modern education. Montessori (1870–1952) established an education system based on a child’s own initiative and abilities. Her "cosmic education" emphasized the interdependence of all the elements of the natural world. She envisioned academic teaching that focused on breeding the "Spirit", the inspiration to do research. As late as 2020, Montessori was chosen as one of the 100 women of the year by Time Magazine. Radboud University honored Montessori by giving her name to a new building and asked artist Jeroen van der Most to create an AI-based art piece for it.

The artwork

In the artwork AI writes a continuous stream of new texts, inspired by the work of Montessori. The AI reflects like a philosopher on research news from the university. As if Montessori might have written the texts herself. Never has anyone used Montessori’s oeuvre and AI to create art. Jeroen van der Most shows with the artwork that AI can give someone from a distant past a place in our current time. Montessori was a crucial inspirator for contemporary education and is now inspiring again through AI. For at least a year, the installation will confront students and researchers, with inspiring new lines of texts. On a daily base, new quotes are written by the system with a series of techniques never used before, read about them below.

GPT-3

The piece was developed in an intensive technical cooperation with Radboud University’s AI department, represented by Gabrielle Ras, PhD candidate AI. The specific AI system used is GPT-3, which is today’s most advanced text-generation AI. It’s an example of a so-called large language model, that is based on a vast amount of textual data including billions of texts from the internet and libraries filled with books, where Montessori’s oeuvre is widely available and discussed. This gives the system the capacity to understand patterns and structures in texts and generate new texts of remarkably high quality in the "style" of Montessori. For the art project GPT-3 was combined with a shell of experimental new features that maximizes the relevance and inspirational impact of the generated quotes, but also implements de-biasing, guarantees new-ness of the texts, filters sensitive content, and presents the output using artificially intelligent design.

System setup

Before the details of the system are discussed, it’s good to understand the system creates the quotes in two steps:

1. Topics for the quotes are extracted by the AI from the research news section of the university’s website.

2. The AI is instructed to write inspirational quotes about the found topics in the style of Montessori.

Reflecting like a philosopher

A new approach is used to extract the topics from the university’s website. GPT-3 looks at research news messages and extract subjects from them. This goes beyond simple keyword extraction, it does so like a true philosopher. The system recognizes the overarching themes in a headline, comes up with related inspiring subjects and uses those for new quotes. The system is able to extract the associative topics based on a series of human-made examples of headlines and subjects. A technique called "few shot learning". Think for example of the headline: "Study shows playing Minecraft benefits children". GPT-3 understands that this is a headline about children, but could also associate it with games, joy, and human-machine interaction.

The use of the associative, overarching subjects has two advantages:

1. The inspirational impact of the quotes is increased. The project aimed to create quotes that have an emotional impact or trigger ideas. "Bigger" overarching themes or associative topics that touch on human emotions or relationships lead to more impactful quotes.

2. More diverse topics and quotes are created. Which makes the output more interesting, but also improves the capacity of the AI to reflect on news messages that touch on topics outside of its data set, a common limitation of current AI systems. Tests were done for example, with the headline "Study shows playing Jadamaga benefits children". Jadamaga is a non-existent game, to which the AI has no reference, but the system is still capable of generating a wide range of relevant topics like parenting, fun, health, or learning based on the surrounding words in the sentence.

De-biasing

The topics are used for quote generation. A lot of time was spent on maximizing the quality of the quotes, with a specific focus on ethics. Several new features were built that make the system’s output in line with the highest standards:

• A de-biasing mechanism was created to make the quotes gender neutral, as the creators noted that the AI would often refer to a teacher or child with the pronoun he. Such biased output is quite common in AI systems, it is a reflection of biased human input data. The artistic choice was made to always replace the pronouns he and she with the word "they" in combination with singular verb. So "he walks" would become "they walks".

• All quotes are Googled by the system. It checks if the texts are new and do not fully, or to a large extent, exist already on the Internet. To prevent that (accidental) plagiarism is committed.

• The generated texts are filtered using several “sensitive content” filters. A combination is used of an advanced filter by GPT-3, a separate machine learning based filter, and a custom-made blacklist of words. To prevent any quotes with explicit language or negative sentiment ending up in the final artwork.
  
  

"Intelligent" design

The final quotes are presented on a four meter wide screen in an experimental design that mixes old and new. A classic font is combined with bright neon-like colors that are common in the current digital art– and NFT scene. Geometric shapes form the base for the texts, a reference to the play devices used in the Montessori education system. A rudimentary animation method, the animated gif, moves the shapes and text onto the screen. The design speculates about future presentations of academic writing and brings up associations with both the first early-twentieth century prints of Montessori, as also retro-scifi, the eighties, and early game- and online art.  

Unique element in the design is that the AI makes design decisions based on an understanding of what it is showing. GPT-3 estimates the most important words in the texts, again based on some example sentences in which a human indicated the most important words. The AI system then highlights the important words with a different color, but also adjusts the speed of the animation with which the words are shown. After important words, pauses are implemented. An experiment to make the display of computer-generated texts more accessible and natural, like human speaking or thinking.  

Spirit

The quality of the texts written by The Spirit Machine has exceeded the expectations of the creators. Do keep in mind that the texts are written by a machine, which could cause them to, sometimes, be strange or confrontational. Most of the time however, they will inspire to learn, teach, or do research: the exact inspiration Montessori described as the “Spirit”.  

The artwork hopes to shine new light on the work of Montessori, but also inspire thought on AI. The system is an experimental aggregation of new ideas and techniques around increasingly smart text- and language AI-systems that might play important roles in our society. Could AI be a medium to bring values and ideas from the past to the present, to benefit our future? The creators would love to discuss this with you and are open to new cooperative projects, talks, and creative challenges based on this article. If you have anything in mind, please do get in touch.