I’m a huge Sherry Turkle fan. I don’t think her recent work is as good as her later work but that’s in part because she’s switched to writing for a popular audience so it’s less nuanced and more sweeping in its judgements. Even so her new book Artificial Intimacy is frustrating me. Here’s an example of why from loc 1993:
In all of my work, interviews are recorded and transcribed. Then I sit down with the transcripts, listen to the recordings, and try to make sense of things. I count on those hours to get close to my data. It’s where I get my best ideas. One day, the administrative assistant in charge of keeping my project files comes up with something she thinks will help, something she can do in minutes.
Note the lack of a subject in that first sentence: interviews are transcribed. It just happens for her. It might be that “the administrative assistant in charge of keeping my project files” is the one doing the transcription. But given that Turkle feels the need to distinguish this administrative assistant by role, it suggests she has multiple assistants with different roles. For such a thoughtful analyst the lack of curiosity about the institutional conditions of her own lifeworld is really startling. The interviews just arrive for her, ready to be worked on in her own preferred mode. Not unlike the professor she describes a few pages earlier who uses as an army of graduate students to prepare their drafts. From loc 1972:
Yet some faculty don’t see how using AI to help with writing differs from using a team of graduate students to prepare first drafts. An architecture professor says, “I usually write with all the students on my grant. They do the first draft and sometimes, a second. I’ll see a later draft and tinker. I don’t feel very different if an AI does the first draft.”
It is a routine experience of being a senior professor in a leading American research university to have work just arrive for you, ready for you to continue it in your preferred mode. For the architecture professor this involves using the graduate students to do the initial work. For Turkle the assistants just prepare the conditions for her to do the deep work (whereas for me the transcription is itself a questionable form of outsourcing, even if it has sadly become commonplace).
They draw the boundary differently about what is acceptable but the conditions in which the boundary is being drawn are shared. Those are the conditions which she does not seem to interrogate but I don’t see how you can analyse the cut between acceptable/unacceptable without doing this. Without it the analysis just collapses into a sequence of judgements about people using models in harmful ways. I’m often agreeing with the claims Turkle is making about the damage being done by models to people and institutions. But there’s a lofty distance from the conditions in which decisions about model use are being made which I find ethically troubling and analytically problematic.
There’s also a lack of curiosity about the different ways in which people can use models for academic work. Turkle suggests that “When we delegate any part of the writing process, we undermine the relationship between writing and thinking” (pg 140). But what if you are writing to the model? I write many thousands of words in my interaction with models, with a not dissimilar form to short blog posts. I discover my thinking through my writing, now in a freer and more readily available way because the barrier to articulation for a prompt is smaller even than a blog post. It provides continuous occasions for exactly the discovery that Turkle is talking about. Yet to the extent this is consciously upstream of formal writing, I am presumably ‘delegating part of the writing process’.
The way she has set this up conceptually excludes so much by definitional fiat in a way which again I think is a product of her own distance from the research object. It’s striking that she doesn’t really seem to genuinely try using a model (as opposed to reporting on some cursory conversations with what she insists on referring to as ‘Chat’ in a way that is bugging me) until the aforementioned assistant puts AI-generated pull quotes in her file alongside the transcript. From pg 141:
Nonetheless, the automated summary is enticing. The AI had done its work in minutes—I might have taken a day or two. I begin to feel the things I’d heard about in my interviews with doctors, lawyers, and businesspeople. Exhilaration: This is magical; I could do so many more interviews. Concern: With the AI cheat sheet, I will read the interviews with less focus. A sense of loss: Even if I keep the AI in my pocket, I know it’s there, ready to give me an answer. What of the decades I have spent learning how to analyze human conversation? Have they gone to waste? What is my job now?
Note that Turkle is still not actually using the model. This is the most passive use of a language model imaginable: being given an AI-generated summary that someone else has generated. Yet she is entranced. Obviously the fact this is on pg 141 does not mean it was necessarily late in the writing process, but it does suggest this experience was late in the research. Perhaps only after she’d conducted the interviews and analysed a significant number of them. Perhaps it’s just the common sense of a digital sociologist but I honestly would have assumed that with anyone doing research on software some meaningful real-world engagement would have come early in the process, not being left to happenstance by an assistant showing initiative in the later stages of the analysis? Again my point is there is a distance from her object which is left unexamined, without which her analysis of how other people’s relationship with that object is shaped by their conditions inevitably suffers.
