If you know me well, you’ll know that I’m staunchly anti-AI and see it as a politically suspect plagiarism machine that erodes our capacity for deep thinking, creativity, community, and empathy. Despite its title, Madhumita Murgia’s Code Dependent: Living in the Shadow of AI is a series of journalistic essays that present, generally, a more optimistic view of AI’s capabilities. I’m still not persuaded, but Murgia offers a balanced assessment that helps to combat apocalyptic thinking. The book offers cautious optimism for those receptive to technology.
Essentially, each chapter focuses on an aspect of AI development or implementation. The book begins with a discussion of who trains algorithms and how it works. Murgia provides a case study of a woman whose whole family was working on algorithm training essentially by completing captchas. The central issues in the chapter are presented as a compelling narrative, even if it’s well-trod ground, namely the ethics of exposing people to horrific content at a rapid rate for minimal pay with no PTSD resources (for fiction, cf. We Had To Remove This Post by Hanna Bervoets). Where Murgia adds her own twist is in discussing how the algorithm training models exclude even the employees training them; in the case study, the woman has multiple warnings filed against her by the system and was suspended from work, cut off from her income. The agency of the human-algorithm dynamic becomes inverted with a startling lack of transparency. That thread is picked up in a later chapter with respect to UberEats. One of UberEats’ employees documented problems with the system, unfair and obscure payment, and started tracking his own time. He also created an app called UberCheats for other employees to document how their wages were being stolen, to share knowledge, and so on. It was a really interesting way of fighting back at a company that hides behind algorithms so that nobody is responsible.
The other compelling element of the first chapter is how we are all training AI. Every time we identify a stop sign, the parts of a bicycle, and so on—all of those millions of captchas are being used to train algorithms. With Murgia’s optimistic spin, she highlights how people can be trained on how to train AI for issues of significance, like identifying diseased skin or anomalous blood samples for medical purposes.
It’s hard to see the light in the second chapter, which elucidates the extent to which AI is used to generate deep fake pornography. Again, Murgia anchors the chapter in a case study of a poet whose likeness was stolen and placed her in all kinds of graphic scenes. The statistics on how much (and how quickly) AI technology has been deployed for revenge porn, deep fake porn, and so on, is truly disturbing; I don’t know if there’s a way to spin that positively. Murgia’s case study for the chapter describes the disorientation of seeing herself in situations that she was almost certainly never in—the most alarming part, perhaps, is the doubt the experience inspired in her, not so much about the situations she was depicted in, but who in her life might have generated these videos. She started eyeing her male friends with suspicion, then her female ones, wondering who was at the heart of stealing her pictures and generating these images and videos. It’s pretty harrowing to hear how ill-equipped the legal system is to create and enforce protections around this kind of stalking and sexual abuse.
On the topic of crime, Murgia has chapters that discuss the predictive power of artificial intelligence. One chapter goes into the ways in which facial recognition algorithms are tracking everyone through the streets and artists and activists’ attempts to resist. In another, Murgia explains how algorithms have replaced humans as the driver of policy and the unforeseen consequences for the most vulnerable. For example, there’s a lengthy discussion of an algorithm that classified 600 of the nation’s youth as being most dangerous and likely to offend or reoffend, and then another list of 400 as a secondary sorting. The Minority Report-esque nature of the AI system classified youth who had offended, but lumped in youth who were the victims of crime, whose families accessed social services, and so on. In an attempt to make statistical predictions regarding crime, the algorithm was creating policy that reinforced extant biases. See also Algorithms of Oppression by Safiya Noble and Automating Inequality by Virginia Eubanks for more information on these problematic practices.
In these situations—really, in all situations outlined in the book—algorithms are improved only through intensive human intervention. By way of example, Murgia has a chapter about healthcare. When people defend AI, they often say how AI will revolutionize healthcare, find cures for cancer, and so on. (Conversely, nobody considers how AI will revolutionize the creation of diseases, but that’s another story…) Murgia’s case study shows how improperly trained algorithms were misdiagnosing patients and it was only with increased face-to-face time with doctors, doctors who understood context, that care actually improved. In Argentina, there was also a project to predict which young girls would get pregnant as teenagers. Among the factors considered for prediction were things like community, religion, income, and so on—but notably ignored all trace of men, as if young girls were self-impregnating. The algorithm also did not account for the practice of gang rape that had become endemic to many communities in the region. The young girls’ data was then maintained in a database that could potentially be accessed by subsequent governments, which is particularly concerning given the precarious legal status of abortion in the area. It’s only when humans started challenging the algorithms that they could be refined—or, more appropriately sometimes, scrapped.
In one of the final chapters, Murgia considers the way ChatGPT in particular has been developed and the (mis)uses to which it has been put. One of my core concerns is that ChatGPT doesn’t produce anything meaningful and I felt validated by the perspective of a science fiction writer who is interviewed in the chapter. ChatGPT is incapable of producing anything that is truly moving and I’m sick of pretending like it could. There’s an anecdote about Murgia’s daughter making puns, trying out language games. The author comments how “ChatGPT doesn’t find anything funny.” In one sentence, that is the core of it. There’s no personality, no life in it. The author continues on to decry the notion that training ChatGPT is like training a child; it’s a completely different process that requires human connection.
I’m unclear, exactly, on how Murgia maintains a sense of optimism with respect to artificial intelligence systems. Every chapter documents problems and shows the cracks in AI systems. The only way to combat faulty systems is to promote increased human intervention by social scientists, artists, historians, and so on—but I’m not optimistic that we’re headed in that direction. I worry that we’re going to end up in a world where AI programs what is statistically most likely, and then reinforces those same statistical probabilities as a matter of policy. It would be great to implement some of the critical questions Murgia shares, but I’d be hard pressed to find examples of business operating benevolently to improve the world.
Good luck out there, fellow luddites! Happy reading!
No comments:
Post a Comment