The last word: AI musings
by Julia Lesage
Since many of my friends say they know nothing about AI, I am taking this editorial space to reflect on how I began to study AI. Early in 2023, Gary Kafer sent in a review of Kate Crawford’s The Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. [Please read Kafer’s piece alongside mine.] It led me to read Crawford’s book, which amplified my concerns about the damage caused by massive digital data gathering. I’d been thinking about AI since the year before when in November, 2022, ChatGPT was released, grew rapidly in its user base, drew many critics, and the nation saw a massive subsequent venture-capital investment in this kind of AI. On a personal level, I saw how ChatGPT suddenly reshaped the work life of many writers and teachers around me, and the strikes in Hollywood made it clear that as a media professional I had to learn more about AI.
Many writers continue to address this new social and technological development. For me, current AI “chat" programs cast an interesting new light on how data can be used, manipulated, and exploited. What most of us do not think about in our daily lives is how much of what we do establishes our digital presence, profitable to others without our consent. To see this more clearly you could make a list of what you do that leaves a digital trail—use a credit card, do a Google search, go to the doctor, send an email, write a check, read an article online, donate money to a non-profit organization, shop at a chain store grocery. It all leaves a trail.
More concretely I’d like to describe my writing process as I learned about AI. In my daily routine I turn to two Internet information-gathering apps on my computer, the use of which accumulates online data that could tell a lot about me. These are Feedly, an RSS feed which provides me with a daily collection of essays and news from periodicals and web sites that I designate, and Instapaper, which makes copies of some of those articles so I can consult them later. Planning this editorial, for about four months I have looked at many news and op-ed articles about AI daily and posted perhaps thirty to fifty of them to Instapaper to use later. In addition, I searched for essays to explain how programs like ChatGPT work; then I found that YouTube lectures by teachers and professionals were often more understandable than the articles.
Finally, I found a “chat” program that did not “hallucinate” [give false information] much as ChaptGPT does—Claude, which is free in its beta form from Anthropic. I have engaged in a number of conversations with Claude, included here in an Appendix, and asked it to comment on an outline I had made of things I might write about [my outline is also included in the appendix]. In the midst of all this, I was active on Facebook, making entries on AI, joining a useful group called ChatGPT for Teachers, and posting copies there of my chats with Claude. Sans Claude, this is my ordinary preparation for writing. I describe my augmented process here, both because it interests me now to see how AI might affect teachers and writers, and in the case of my interacting with Claude, because I know I am also training it for its future iterations.
This is what I know about how an AI like ChatGPT works. In its initial training phase, the task set the program is that the trainer gives it a word in linguistically correct sequence and asks the program to predict, statistically, what would follow. The training consists of giving it lots of text, many billions of words, to analyze, and hundreds of billions of internal nodes so it can build longer, more complex sentences. In some way, the nodes can each also communicate with each other, so there are many more internal interconnections. This is called an LLM, a large language model. Such a program requires heavy computer power to process the input textual material, taking perhaps a year. It’s expensive, For example, ChatGPT’s initial development cost about a half a billion dollars.
In addition this kind of LLM learning has serious limitations. For example, because human speech and writing can be hateful, much of the original output was hateful and needed curation, done by low-paid human labor. Additionally, because this AI is voracious for textual input, it gobbles much creative work, which is now justly legally contested. Furthermore, the initial programs released to the public often gave made-up or false answers to users’ questions since the programs were trained originally only to compose linguistically correct, conversational discourse. Even if later iterations of these products build in safeguards to deal with such large problems, especially in terms of erroneous content, it takes years and great expense to create changes in the program itself.
In the Appendix, the chats I had with Claude provide a more detailed version of how LLM AI works, but I want to give you my Dick-and-Jane version here because you can see my struggle to understand a process which I find strange and disturbing but also useful and even exciting in its results.
What is both exciting and disturbing to many computer experts is that this kind of ChatAI is much more efficient and smarter than predicted in its response to users. It can take on new tasks while needing many fewer examples to learn from; it integrates knowledge and data from different conceptual realms (e.g., science, law, computing, fiction writing) if asked to do so; it can now police itself according to humanistic rules so that it does not spew out hate or tell how to make weapons. At the same time, no one knows how its many internal nodes are communicating with each other to do such problem solving. Because of these new kinds of tasks that something designed as just a language completion program can take on, computer experts themselves are divided. Optimists in the computing field can now anticipate a time when computing intelligence might approximate “general human intelligence”; pessimists worry if it will do great harm, perhaps in military application. My own opinion is that in the near future, because of the long-range trajectory of capitalism and its shaping of how we develop and use technology, the massive investment in, and use, of AI are here permanently, but in ways we should learn to use and hopefully shape. I think it will change our lives as much as, if not more, than computers have already done.
Chat AI in perspective
Because my entire professional life has been conducted in and around English departments, ChatGPT immediately struck me as a “homework doing” or “plagiarism” or “exam taking” machine, and I alerted my former department to the fact. But since then, even though students massively now use the program in this way, from a teacher’s perspective such a worry seems like a superficial response. Teaching writing, especially argumentative writing or writing based on research, still remains one of an English teacher’s essential tasks; we understand how learning to write an essay is a formative moment in adult cognitive development. In my own essay writing, this means taking time to do lots of reading and note taking, perhaps interviewing other teachers or practitioners, and then organizing the notes in long outlines, some of which I use to write an essay. My brief experience with Claude indicates that a Chat program is more useful at the beginning stages of this process than at the end stage, and I would advise teachers to teach it as a research or brainstorming tool.
In the first months of ChatGPT’s release, an image from the past kept coming to me. It was from Harry Braverman’s Labor and Monopoly Capital. Written in 1974, Braverman’s book had a great influence on the Jump Cut editors as it described how the industrial labor process in the United States in the late 19th century changed with the advent of “scientific management” or Taylorism. What impressed me most when I first read that book was that Braverman also vividly described the massive growth to come in the 20th century of the white collar work force, whom I taught. He developed an image that metaphorically had a great impact on me as I imagined a “paper empire” built on top of the industrial one. He writes,
“A mass of clerical workers has come into existence whose work embraces all that was formerly handled on an informal basis in the shop itself, or on a minimal basis in the small shop offices of the past. Since management now carries on the production process from its desktops, conducting on paper a parallel process that follows and anticipates everything that happens in production itself, an enormous mass of recordkeeping and calculation comes into being. Materials, work in progress, finished inventory, labor, machinery, are subjected to meticulous time and cost accounting. Each step is detailed, recorded, and controlled from afar, and worked up into reports that offer a cross-sectional picture at a given moment, often on a daily basis, of the physical processes of production, maintenance, shipment, storage, etc. This work is attended by armies of clerks, data-processing equipment, and an office management dedicated to its accomplishment. “ (p. 373)
Fast forward to the 21st century. With the massive growth of home and business computing and then the growth of the Internet, I now imagine a data or information empire built on top of that paper empire. Authors who analyze this empire’s oppressivness, Nick Couldry and Ulises Mejia describe this new kind of empire as Data Colonialism, and Shoshana Zuboff as Surveillance Capitalism. Furthermore, LLM AI represents an even greater step in this largely invisible empire’s development. Perhaps now best called an Operational Data empire, Jussi Parikka describes it in terms of imperceptible intellectual and physical operational processes that shape our material world.
To flesh out this metaphoric image of overlapping “empires,” one would need to add an understanding of how capital accumulation seeks out new markets, and according to different potentialities, how it creates those markets. It is a dynamic and historical process, augmented by the close relation between those in power and the state. Marxist theorist David Harvey in The Limits to Capital (1982) laid out a picture of contemporary capital accumulation and its cycles that still has relevance today in thinking about the rapid expansion of AI. Billions of speculative dollars are being poured into its development, deriving from and contributing to the fortunes concentrated in the hands of a few, especially in the tech industries. With acts of appropriation, the rich, aided by state power, are seizing terra nullius, a space up for grabs, finding a new way for using the data already being gathered about us as individuals and about our world.
I am embedded in the same “grid” as those around me and understand the digital world’s oppressive aspects but can also see the potentials in the new tools offered us. As I try to understand AI programs as they impact teachers, students, and artists, it seems like it has uses for us, too. It’s clear that we are just at the beginning of AI operational innovation. Right now that is taking place mostly in expensive computer development processes in tech companies in the United States and China, and surely in the military.
As I think of how fast ChatGPT is affecting teaching, I recall when computers entered my own life as a necessary part of my work. For Jump Cut, the computer offered a new way of doing layout and combining image and text, as we used laser printers at home to prepare text to take to the offset printer. In its later iteration, our computerized layout program developed an efficient interface with the Internet. We could then create a website and abandon expensive paper publication for inexpensive, worldwide online diffusion. During this period of learning to incorporate the computer into my professional life, I also recall how slow my academic colleagues were to become computer literate—e.g., for years stubbornly resisting sending and receiving “attachments.” Later, the Jump Cut editors found similar resistance among our writers when we first considered going electronic, yet the Internet quickly allowed us to expand to international reach in terms of readership and authors. My history adapting to technological change is why I write here to ask our readers not only to fight appropriation and plagiarism but also to consider how new educational and creative strategies might potentially emerge from “chat” programs now.
Braverman, Harry. Labor and Monopoly Capital: the Degredation of Work in the Twentieth Century. (1974) New York: Monthly Review Press, 1998.
Couldry, Nick and Ulises A. Mejias. “Data Colonialism: Rethinking Big Data’s Relation to the Contemporary Subject,” Television & New Media, 20:4, Sept 2, 2018.
Crawford, Kate. The Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press, 2021.
Harvey, David. (1982) The Limits to Capital. New York: Verso, 2018.
Parikka, Jussi. Operational Images: From the Visual to the Invisual, University of Minnesota Press, 2023.
Zuboff, Shoshana, The Age of Surveillance Capitalism: The Fight for a Human Future and the New Frontier of Power (New York: Public Affairs, 2019). Jump Cut review by Victor Wallis: http://ejumpcut.org/archive/jc59.2019/VicWallis-Zuboff/index.html
Click on links to read. I discussed with Claude mainly how AI functions, application to education and writing, and housewifery.
Lesage working outline for this essay.
Depending on your browser settings on your computer, this will open either onscreen or be downloaded.