Illiterate Machine

Illiterate Machine is an art installation which tries to raise awareness for critical challenges of and around machine learning technology:

  • Culture, as well as technology, evolve at a very fast pace. Many open source machine learning datasets, however, are often not updated as quickly as culture and technology changes. This might lead to a distorted machine reality.
  • In contrast, datasets used for commercial purposes are kept up to date. Yet, transparency on how they have been trained is low. Since datasets are by nature biased by culture and the current zeitgeist, those biases can easily spread unobtrusively.

Illiterate Machine is imagined as an interactive museum installation. It demonstrates how an object would be interpreted by a machine quite differently depending on the time and context when the underlying machine learning dataset had been trained.

Museum visitors can pick one out of multiple cards displaying different objects. They can slide the card into different slots on a wooden timeline. Each slot represents the time when the corresponding machine learning dataset has been trained. Once the card has been slid into a slot, an ultrasonic sensor reads the position on the timeline. A USB camera reads the picture. Depending on the input data from both the ultrasonic sensor and the camera, Illiterate Machine displays how it interprets what it sees on a screen in front of the visitor.

A card displaying a woman is interpreted as “A woman” in the 1960 and the 1980 time slots. In the 2016 time slot, it is interpreted as “Jenny who is gay and will vote for Donald Trump”. In the “future” time slot, the picture is construed as “This is Jenny with a social score of 28.159 points.”

The input data is interpreted by a trained machine learning algorithm in Wekinator, while the output is displayed using the Processing framework.

Illiterate Machine aims at making machine learning approachable, while the thought-provoking interaction demonstrates the inherent risks and problems of this technology.