In Jay Bennett’s 1986 young adult novel, Deathman, Do Not Follow Me, a teenage boy who regularly visits a local art museum to see Vincent Van Gogh’s 1889 painting, Starry, Starry Night, realizes when the painting has been stolen and carefully replaced with a forgery. He intuits the theft long before he catches the thieves, because standing in front of the forgery is not as emotionally affecting for him as standing in front of the actual painting. His experience demonstrates that genuine emotional expressiveness, such as Van Gogh’s when he painted, creates authentic human connections that transcend mortal and temporal barriers. If the capacities for imagining and empathizing are not unique to artists, surely they must be unique to humans. While the ability to empathize with others may still be safely regarded as a human (or mammalian) trait, the ability to imagine is rapidly becoming a routine function of machines using technology based on artificial intelligence (AI).
In 2017, NVDIA, a company primarily specializing in technology related to computing and artificial intelligence, published the results of an experiment. Using photos of real life celebrities, programmers taught a machine to generate realistic images of celebrities who did not actually exist. Granted, “realistic” is a subjective term. Anyone who wishes to form an informed opinion about the experiment can take a look at these “people”…
NVDIA’s faces were generated using Generative Adversarial Networks (GANs). Essentially, a GAN is a form of artificial intelligence where the neural network—the part of a computer program that gathers information and determines how that information should be categorized—simultaneously performs two competing functions. One part, the generator, generates new data. Another part, the discriminator, determines whether that data belongs to the actual training data set. In the case of NVDIA’s experiment, then, the discriminator was trained on a data set (photos of actual celebrities) and the generator repeatedly generated new model instances (distributions of physical traits that the distributor could potentially include as part of its original data set). In less technical terms, NVDIA’s experiment is well regarded because the distributor accepted the majority of the generator’s fake celebrities as potentially real ones. A discriminative algorithm, such as the kind found in a spam blocker, classifies input data based on its features. For example, the algorithm might be used to determine how many words consistently found in spam messages are contained in a particular email, after samples of those types of words had been provided by a programmer as input data.
By contrast, a generative algorithm, such as the kind used for a GAN, might take a provided classification, such as “celebrity,” and determine based on input data which features might be present within that classification. A discriminative algorithm classifies information, but a generative algorithm generates possibilities based on existing data. Insofar as it gives a machine the ability to use what is to form possibilities for what could be, its mode of operation bears rudimentary similarities to the functioning of the human imagination.
A more sophisticated, nuanced neural network, while perhaps unsettling, is not inherently harmful. After all, who wouldn’t want to generate a convincing rendering of the face of a human hottie to show nosy relatives during family gatherings? Ian Goodfellow, who is colloquially known as “The GANfather” because of his seminal role in pioneering GAN technology in 2014, believes his technology is more than just a maker of convincingly pretty faces. Goodfellow says he believes GANs wil make significant sociocultural contributions due to the technological advances they will make possible. GANs could make the optimization of certain tasks, such as predicting the interactions of subatomic particles or determining which medications best treat certain diseases or conditions, much easier to consistently achieve reliably and efficiently. For example, a GAN’s ability to generate potential medical records could allow researchers to treat patients more successfully without needing to wait for individual cases or secure permission to view medical records from individual patients.
However, Goodfellow is currently heading a team at Google dedicated to making machine learning technology more secure. After all, as Goodfellow admits, machine technology is amoral. The same generative algorithms that could make the aforementioned scientific advances possible could also simplify the processes of creating misleading or false news items wrongly attributing potentially damning images or quotes to public figures, especially politicians. Cyber security breaches, too, can be more successfully completed if a hacker can use generative algorithms to quickly and reliably guess the features of a cybersecurity systems. Goodfellow says the most effective safeguard against cybersecurity breaches using GAN technology is not more advanced cybersecurity systems, even though he is working with Google to develop those. Humans’ best defense, he says, is their own perceptiveness. He suggests students take speech and critical thinking classes that will require them to analyze information well and recognize authentic, persuasive arguments.
To return to the example from the opening of this article, a GAN can perform the roles of both the teenager and the art thieves in, Deathman, Do Not Follow Me. Like the thieves, the generator uses existing data to create a convincing, if not an authentic. representation of something that exists in real life. Like the teenager who returns Van Gogh’s painting to the museum, the distributor determines which generated images do not meet the criteria required for those images to be deemed convincing when compared to the original data set provided. Yet this analogy does not include one element central to Bennett’s novel about a contemporary boy who feels a personal affinity with a nineteenth century artist: empathy. The boy does not discover the painting is a forgery by using input data, though he does use deductive logic to determine who stole it. At first, he only knows he no longer feels the same emotional connection to the painting when he stands in front of it. Machines may be developing the capacity to generate possibilities, but the ability to determine the implications of pursuing those possibilities in the real world belongs solely to humankind.