With Sui generis, I question how Artificial Intelligence transforms the notion of individuality by reducing human complexity to mere statistical abstractions. It started from an investigation on the way Machine Learning models operate and how we can relate them to human decision-making. I started this research with the premise that Machine Learning algorithms are judgement machines, as a response to the assumption that these technologies are neutral, sterile and objective. I, instead, argue that Machine Learning and Artificial Intelligence are built to make decisions. Because decisions are inherently biased, Machine Learning is meant to be biased.

More than being a collection of subjective data, the datasets used for Machine Learning training deprives us, its subjects, of our individuality. Once subjects are crowded together in the dataset, they then become indistinguishable from their neighbours. This process produces a flattening, triggered by the statistical projections of Machine Learning. You are only recognisable as an individual if you are coherent within your contiguity. It is an identification game, where we are identified by our similarity to our neighbours, our Facebook friends, or our fellow web surfers that have an aligned browser history. This neighbourhood is now reduced to a crowd of sameness.

Because Machine Learning algorithms operate on mathematical thinking, they reduce all the individuals to a statistical entity. This entity can never generate new futures, but only reproduce the past, perpetuating its biases and leaving no space for growth. Because these algorithms are tested on their ability to reproduce history, the future now equates to the past only, and therefore embeds all its mistakes, leaving no space for learning from them. The future is automated and caught in an endless loop of past events.

With Sui generis, my goal is to highlight this behaviour by generating a personal classification machine that is able to predict, or not, how much like or dislike a certain thing. For that, I created a small dataset of words tied to a value between 0 and 100. The main artefact, the dictionary, critiques the false personalisation of the model I developed. The choice of a dictionary as an object is meaningful: because it serves a symbol of universal knowledge, it only reinforces the idea that I am generating yet another instance of standardised universality. Each numerical classification exposes the limitations of the model, as they can never be truly personalised. On the other hand, the installation space mirrors a clinical or forensic environment, where visitors examine the book through a digital microscope. This creates a continuous cycle of scrutiny and judgement, transforming what was originally subjective data into quantitative interpretations.

The model I fine-tuned inherits not only the biases of its previous training but also the imprints of countless voices across the internet — every individual who has ever contributed to its datasets. What does the statistical output given by the model tell me about other people’s preferences? In what ways are they embedded in these supposed personalised classifications? How is my identity being influenced by the homogenised entity built by the model?