The role of deception and fabulation in generative algorithms

• In Event: Learning Under Algorithmic Conditions: Critical Perspectives on AI and Curriculum

Abstract

The role of deception and fabulation in generative algorithms

Overview
---------
With the launch of ChatGPT and similar platforms, generative algorithms have shifted both the field of AI and the field of education. Generative algorithms gain skill through what is termed ‘unsupervised learning’ (Buckner, 2019). These algorithms train on curriculum but they also generate invented plausible images and words, and learn from both the training set and these generated or ‘imagined’ data sets. Computer scientists call this generative process of formulating non-examples to test and correct algorithms a matter of “fiction”, “fabulation”, “dreaming” and “hallucination” (Ellis et al, 2020). This paper looks closely at the algorithmic architecture that affords this kind of learning, and argues that generative algorithms mimic certain human learning habits.

Objective
---------
The paper aims to show how generative algorithms exhibit a robust kind of learning, insofar as they perform a kind of ‘creative’ abstraction and hypothesis generation. My first objective is to explain how generative algorithms leverage the power of fabulation, as a way of learning how to create something new (an image, an essay, etc). My second objective is to demonstrate how the creative process of fabulation is both similar and different in humans.

Theoretical framework
---------------------
Theories in curriculum studies have long celebrated the role of creativity in learning. Focusing on the key concepts of ‘generative’ and fabulation, I draw principally on two philosophers, Gilbert Simondon and Gilles Deleuze. Simondon (2017) offers a framework for rethinking the relationship between humans and technology, contesting earlier Marxist theories. Deleuze offers a framework for revealing the role of the imagination in human learning. My argument is that generative algorithms learn according to the same polemic that Deleuze (1991) sees at the heart of human learning; a polemic between the imagination (fabulation) and its constant correction from perception.

Methods and data
-----------------
The method of exegesis includes analyzing the specific algorithmic learning that is performed by Generative Adversarial Networks (GAN), which are machine learning frameworks designed by Ian Goodfellow in 2014. GANs have been used in both science and art, supporting experimental practices, exploring hypothetical spaces, and testing simulations in education research. I will contrast these GAN methods with five artist accounts of their own creative process of fabulation.

Findings & Scholarly Significance
---------------------------------
Given the rapid intrusion of AI in education, it’s essential that we understand how generative machine learning algorithms operate. All too often we hear that they are “black box” devices, but in fact there is much we can understand about these mechanisms, and more educators need to know these details. This paper offers concrete insight into how some of these AI algorithms work, and raises critical awareness about their ‘learning style’, both different from and similar to human learning. I show how generative algorithms learn through deception and fabulation, and that this process aligns somewhat with human habits of learning and creative invention. I discuss the malicious applications that have led to concerns about GAN ‘deep fakes’, and legal responses that have included the 2020 California law that controls the use of this kind of technology.

Author

• Elizabeth De Freitas, Adelphi University