“I am not a robot…or am I?” – Elissavet Kardami

Melvin Waver’s talk, Using Neural Networks to Study Conceptual Shifts in Text and Image, provided some critical insight on the potential of neural networks in conducting academic research in the field of Humanities. The proliferation of available data and the rapid developments in the processing possibilities, in conjunction with the increasing digitization of archives, has opened up new possibilities for research. It has not only helped scholars answer questions that were formerly considered impossible to answer, but it has also brought forward new questions that the new digital tools are called to answer. However, parallel to valuable research potential brought forward by these neural networks, several questions arise regarding the implications of such advancements, the underlying mechanisms behind their development, as well as the ways they constantly stretch and challenge the liminal space between what can be considered human and non-human perception.

One of the current tasks of developing neural networks is to learn how to identify and categorize images. This can be a demanding and time-consuming process given the complexity of the sensorial potential of living organisms. There are different layers of recognition of visual patterns and objects that a neural network is asked to identify, which might seem obvious for humans, but difficult for an AI unless properly trained. The process involves the accumulation of a rather large dataset of images that include a specific object which has been manually tagged by a person, and then through a complex analytical process, based on a probabilistic model, the neural network gets trained to identify the requested object.  However, the need for neural networks to become more refined and accurate implies the compilation of a vast number of datasets and a significant number of working hours to compile the datasets and assess the performance of the neural networks. But how can this be achieved in an efficient and a methodical way?

This is where Melvin Waver’s talk illuminated some elements of our daily interactions on the internet which are closely related to the development of neural networks but, are nevertheless concealed for the typical user. More specifically, I am referring to the way companies, who function primarily on the internet providing a wide range of services, are infusing several applications in websites and indirectly using the cognitive and sensorial potential of its users to train neural networks. A clear example is the way website require from its users to declare that “I am not a robot”, and then in order to prove that, they need to identify specific objects amongst a set of images. What this identification of the requested object does is that is functions as a confirmation mechanism for the neural network regarding its initial assumption of each image containing that object. This not only raises ethical questions regarding the use of free labour done by internet users without their consent. It also gives rise to more philosophical questions which position the users in a discourse around what it means to be a robot and what are the parameters that are used to define one. Although the premise of the application is to identify yourself as a human and clearly distinguish your actions from those of a robot, you are being unconsciously conditioned into falling into patterns of behaviour that are more associated with robots than humans.

So, at first, I can easily assert myself that I am not a robot, but then, on a second thought, I start wondering…What if I am being made into one?