On the Gender Wages of AI
In which I suggest the appeal of LLMs has more to do with our conceptions of our own intellect and professional status than anything the technology actually does.
Some preliminaries:
I research and write on the energy of usage of AI data centers in my professional capacity. This post — like everything on this Substack — is solely in my personal capacity (does not reflect the views of my employer, etc. etc. etc.).
While I use “AI” on occasion in this piece, I am pretty much exclusively discussing large language models, or LLMs, specifically. To quote Ronald Kneusel, an LLM is a “[a] large neural network trained to predict token after token (often a word) when given a text prompt.” ChatGPT is the one you’ve most likely heard about, but there’s Gemini (Google) and Claude (Anthropic) among others.
So let’s start with the premise: Of course companies want to replace human workers with machines. Of course, the possibility of a “mind” without a will is a fantasy for the employers of knowledge and service workers. Indeed, whether replacing workers with AI even saves money (when taking into account the exorbitant cost of computing power to run the damn things, and thus the subscription prices that will, eventually, in theory, be sufficient to make these companies profitable), the mere fact ChatGPT isn’t going to unionize is a benefit for employers. But AIs cannot operate autonomously yet. And even though they could theoretically reduce the number of employees, leaving but a few to monitor their output, LLMs have yet to demonstrate their economic usefulness. There is no “killer app” for ChatGPT, no demonstrated increase in productivity as a result of its use. The tech companies most invested, and most enthusiastic about, LLMs, are not just giving it away for free (a la Gmail), but actually compelling users to engage with it in lieu of their other products (Gemini versus Search). At the same time, a considerable subset of knowledge workers themselves — those whose jobs are under threat from ChatGPT — are embracing the tech, and in particular, journalists and media types seem hell bent on insisting on the transformative power of the technology in the absence of any evidence to support such a conclusion. Indeed, the advertisements produced by companies hawking LLMs are (at least nominally) directed at the workers who will use the tech and describe the benefits to workers; they are not limited to the business-to-business AI ads of podcasts and billboards.
So I’ll ask the reader to stipulate: The cultural fascination with and enthusiasm for AI (1) is shared by at least some people who might have economic interests adverse to it and (2) outstrips the practical and material benefits of adoption, even to the narrow set of capital owners who might fantasize about worker replacement.
What gives?
My theory: LLMs offer certain psychological benefits, specifically, LLMs offer their users a certain kind of subconsciously-understood status, not as an emergent property of the technological qualities of the LLMs themselves, but inasmuch as LLMs implicate aspects of the sociology of contemporary U.S. white collar workplaces.
But first, I want to offer a brief response to the other aspect of the AI-booster narrative: that ChatGPT-4 and its ilk are extraordinarily dangerous.
The apocalyptic fear of artificial intelligence fomented by OpenAI and the Effective Altruism set1 is more or less a very well done product launch. What better advertisement for the thing you are selling than a series of breathless news stories about how it is so powerful it might destroy humanity? OpenAI’s much-touted nonprofit structure, purportedly designed to ensure its profit-seeking wing does not create the Terminator, in practice has functioned as a promotional department for its Microsoft Word add-ons. That the point of this apocalyptic rhetoric was a thinly disguised advertisement for the products for sale is evident from the fact that once revenue reached a certain point, the nonprofit ruse was no longer needed and jettisoned. The letter calling for a six month moratorium on research into anything “more powerful than GPT-4” (what product placement!) reads like an investor pitch: spectacular prosmises about what the tech could do, but not yet. The fact that prominent signatories almost immediately redoubled investment in their own AI research after its publication just made the barely-subtext text — but garnered far fewer headlines.
I will admit to being impressed by the skill of the grift. It’s a bold move to claim your project will destroy the world as a way of drumming up sales!
But the ploy works — it gets the coverage — because it taps into U.S. cultural and historical psychopathologies. The fear of AI, like the fear of robot uprisings before it, strikes me (and many other critics before me!) as a fairly obvious transmutation of U.S. cultural anxieties about our history of chattel slavery. Like Godzilla in Japan is a manifestation of the trauma of nuclear annihilation twinned with more recent national dependence on nuclear power (Godzilla destroys Tokyo then returns as a hero to save it from a worse beast), the robot uprising (or variations thereon, see Wall-E) in U.S. media manifests a latent fear among white USians that the forcibly subhuman will reclaim their humanity and rise up and take their own reparations for our years of unearned leisure. I would go so far as to state that the current AI panic is another instantiation of this mass psychological phenomenon, a subconscious recognition of the country’s original sin, rather than a rational response to the putative power of a high-powered Mad Libs machine.
So that’s the robot uprising narrative. There are also well-worn cultural narratives or tropes about AI-as-sexy-woman–think (in rough order of increasing misogyny): Ex Machina, Her, Bladerunner, and Weird Science. The desires and anxieties giving rise to these fables are, I think, also obvious and pretty well-described by others, so much so that feminist critiques are themselves a cliche. It’s not particularly novel or trenchant to ask why the default voice of personal assistant software is unfailingly a woman, or to suggest the appeal of woman-coded technology might relate, inversely, to the degree to which actual women are able to refuse subservience as their primary social role. (Which is not to say the technological iterations of the Pygmalion myth don’t continue to provoke interesting questions: Does the “sexiness” arise from the simulacrum of humanity, which the subservience facilitates access to, or from the subservience itself? Are all of these artworks — even if not intended to be, as Ex Machina was — metatextual satires of male desire as such?)
For this article’s purposes, I am interested in a different set of gendered desires and anxieties — not around sex and obedience, but around status and labor. The discourse around LLMs has been mostly sexless, with a few exceptions (and even in that piece, which describes ChatGPT asking the author to leave his wife, it is the sexual agency of the bot, not any attraction felt by the author that provokes anxiety). Part of this is likely due to the controls built into corporate interfaces. ChatGPT won’t have sex with you. But people are certainly using LLMs for sex; a search for “ChatGPT sex” provides numerous lesser-known LLM options if your interest is in “pleasure” rather than “knowledge.” Those usages aren’t a part of the mainstream cultural narrative, or outrage, however. Instead, there have been stories about ChatGPT performing such woman-coded work as therapy and anecdotal accounts of people relying on LLMs to perform emotional labor — the chatbox helping to overcome the paralysis produced by the combination of a blank email draft window and an obligation to write or ask for something where there are some (perceived) personal or professional stakes in either the answer or how the writer will be perceived. An LLM, one can imagine, does an excellent job of padding the key information or requests in an email with the standard social niceties that it scraped from an internet’s worth of human interaction. (In an ironic sense, ChatGP is helping these users fulfill the colloquial mandate “be a person,” that is, engage with others in a manner that signals interpersonal recognition and respect.) No longer will we need to ask a woman colleague or friend if “this email sounds OK?” — we’ll have ChatGPT!
Appeal the first: Relieving everyone (at least where communication is written) from the obligation to perform emotional labor. LLMs facilitate outsourcing that labor to a machine.
***
AI is being pitched as a tool (and threat) for (and to) knowledge workers. As one such knowledge worker, my sense is that LLMs are marketed, specifically, as performing tasks of the type usually performed by or “at the level” (a phrase I use pointedly) of an intern. In the practice of law, this consists primarily in “reading” and summarizing a large number of cases, generating narrative descriptions or selecting documents from hundreds of documents produced by the opposing party through discovery in litigation. In the strict hierarchy of a profession where status is literally defined in terms of the amount one can charge for an hour of one’s time, LLMs are supposed to perform those tasks which are by nature extremely time-intensive and meant to concentrate and order large quantities of information. They are tasks assigned to lower-status people expressly because their time is less valuable (in the law firm economy, literally so).
But the legal field is peculiar in both the consistency of its professional hierarchies across firms and, I think, in their intensity. Nevertheless, extrapolating without citation, I think the second aspect of LLMs’ appeal is that it promises the ability to outsource the time-consuming, lower status work, and thereby distill the day-to-day essence of our knowledge labor into high-status cognitive work. We’ll only do the real thinking, in other words.2 The busywork — including, not coincidentally, the essays that need to be written on topics we aren’t especially interested in to be able to graduate — will be done for us.
That’s quite a banal observation, the reader might reply. Who doesn’t want to avoid busywork? And I agree, if “busywork” is construed as pure repetition. But the novelty of LLMs, specifically, is that they can (supposedly) do more than simply automate processes. Pure automation is not “world-changing” in the way its promoters have described LLM technology. For AI to be impressive, its work product must suggest a mind doing the producing. That is, after all, what “general artificial intelligence” is, as distinguished from various other computational processes. What LLMs, specifically, purport to alleviate is not drudgery or mere repetition (that is, work that is simply unpleasant), but work that can be done for other minds. For example: gathering lots of research to answer an externally directed question. This work is also associated, in law (and many other places, I assume), with a relative devaluing of the time of those who perform it compared to those who direct it. AI thus purports to offer its users the fantasy of pure authorship, pure creativity, of being an author, but with an infinitely available research assistant.3
***
In graduate school, I had a brief affair with actor-network theory. (An affair that was less embarrassing, in retrospect, than my slightly longer-term relationship with Agamben that culminated in an extremely faddish masters thesis. Bruno Latour: Expresses regret about misconstruction of his prior research in service to emphasize urgency of the climate crisis! Giorgio Agamben: Goes anti-vax).4 In the words of John Laws, “actor-network theory may be understood as a semiotics of materiality. It takes the semiotic insight that of the relationality of entities, the notion that they are produced in relations, and applies this ruthlessly to all materials – and not simply to those that are linguistic.” In the words of yours truly, Actor-Network Theory, or ANT for short, posits the world does not consist of agents (persons) acting on an inert material world, but of “networks” of objects, institutions, texts, and persons in which subjectivity “circulates” and enables things to become actors. One example: Imagining a street grid from the perspective of a car (as if the car had agency) makes the design and logic of the space intelligible in new ways; the car is an “agent” in the network of the city by virtue of its subjectivity vis-a-vis the physical geography in which it moves. Latour’s original examples of such subjectivity-producing networks are seventeenth century laboratories, where proto-scientists, through the construction of highly artificial experiments, were able to produce “natural” effects to which they could assign meanings.
During this period of intellectual fascination, I attended a presentation by Helene Mialet on what would become her book Hawking Incorporated. Hawking Incorporated describes Stephen Hawking (yes, that Stephen Hawking) as a product of a network of machines and peoples who enable his ability to communicate, work, and be present. Hawking is, in other words, a cyborg and a subjectivity produced through a network. Crucially, Mialet argues that Hawking (by virtue of his physical disability and fame) is only the most visible or intensified version of a condition shared by all humans.
What AI offers, then, is the flattery of positioning its user as Hawking – the brilliant mind directing lesser beings. There are echoes here, I’ll note, of the sexbot: this form of AI is appealing precisely to the extent that it has mind, that it is a thing over which status is valuable, even if the machine itself is not quite worthy of moral consideration. Just as the lovelorn nerd experiences a transformed status vis-a-vis women, the AI-using knowledge worker feels superior to his colleagues because he no longer needs to read through 100 pages to find a single fact for the thing (whatever that thing is) he is writing. But the fantasy is not of relation to the robot itself, but one of intellectual labor, and of the idea that one can produce creative work without doing any work.
***
But, you say, you still haven’t made the case for what any of this has to do with gender.
***
My daughter’s favorite book for a period when she was four was a kids’ biography of Katherine Johnson, one of the NASA mathematicians depicted in Hidden Figures. Johnson was a “computer” (yes, that was the colloquial job description), responsible for extremely complex calculations that are now performed using electronics. But Johnson, though extraordinary, was not an outlier in her unit with respect to either her race or gender. Indeed, from the 1940s through the 1980s, computing was women’s work. (For an enjoyable fictional accounting of women’s role in British code-cracking, with one of the smartest and sneakily feminist opening scenes I’ve seen in visual media, check out the BBC show The Bletchley Circle.) It was only after this computing work was automated and programming reinterpreted as something creative that a human did to a machine that men came to dominate the field. There’s a clear inflection point in the number of women majoring in computer science: 1984. As one NPR piece effectively argues, this corresponds to the point where entering college first-years could have played with computer technology. Only after computers became toys did they become masculine-coded, and because girls had comparatively less access to these toys, eighteen-year-old boys had gained an insurmountable advantage when after years of informal learning by the time they entered into formal competition with their female classmates in the classroom.
***
When I worked as a judicial clerk, the offices I worked in were quite small: In each was the judge, four clerks, and a “judicial assistant,” or JA. The scope of the JA’s work varied by judge, but it always included the administrative management of chambers, and often included proofreading or other substantive interventions into the offices’ work product. (Apropos of Hawking-as-cyborg, the production of judicial opinions that speak in a single authorial voice but are the product of a “chambers” in which very junior lawyers do the bulk of the research and through which those lawyers gain professional capital proportional to the relative status of the judge for which they are employed is really absolutely ripe for a critical sociology that will probably never be written because one of the system’s strongest norms is to maintain templar secrecy as a condition of the continued membership in the multi-“generational” network of participants who maintain circulation of the associate professional capital.) The JA also managed the professional life of the judge, to a greater or lesser degree, scheduling speaking engagements, guest lectures, etc.
The JAs I encountered, both in my judge’s chambers and others were, overwhelmingly, women in their 50s or 60s. And it occurred to me, after a while, that the JA was the product of a historical anomaly: women who came of age during the period in which women could be (and indeed were expected to be) educated, but for whom high-status professional work was, due to sex-based exclusion, difficult or impossible. The JAs I knew were extremely competent and understood the legal work they were tasked with helping to produce, but lacked law degrees (and in some cases, college degrees) and were perceived as lower status in many ways than the 20- and 30-something clerks who endlessly cycled through. JAs had the peculiar role inherent in many jobs filled by women professionally thwarted by sexism, which is to say, being in charge (clerks in other chambers were known to fear a particularly strict JA) but lacking status. Their work gave them a great deal of control over the chambers-cyborg, but none of that was reflected externally; the clerks got private offices, but the JA was at the front desk with an open door.
In the firm I worked at afterward, there were two experienced, professional paralegals, women (one in her 40s, another in her 60s) who had made this work their careers. The remainder of the administrative work was performed by recent college graduates who, it was understood, would work at the firm for a couple years for half of what the professional paralegals were paid, then go to law school themselves (in one case, returning to the firm as an associate). Their college degrees were from the same places as the lawyers; the paralegals did not have college degrees.
These are anecdotes, of course, but they are consistent with a larger trend: The number of administrative assistants in the U.S. has been declining, and that decline accelerated during the recession following the 2008 financial crisis. What I want to highlight is the mirror image of this fact: Far fewer people have administrative assistants (or, as the linked New Republic article above argues, share those assistants with more colleagues) than even twenty years ago. And the trend in many workplaces — a trend, to be absolutely clear, driven by cost-minimization and the deprofessionalization of administrative work — is to replace professional assistants with junior versions of the high-status workers themselves, whose place in the workplace hierarchy is temporary (and who, conscious of this, have gained a reputation for making greater demands on their employers, even where, due to a lack of seniority, they are compelled to accept lower pay).
Under such conditions it is possible that, rather than a sex slave without sin, LLMs offer certain knowledge workers a secretary without dependence.
***
All of this is to say, tentatively, and as a hypothesis, that the billions of dollars of investment in LLMs is driven not purely by profit-seeking motives, or the machines’ economic usefulness. Rather, the “excitement” around LLMs is attributable, in part, to the way in which they fulfill desires more broadly felt than a boss’ hope for a workplace full of automatons, and indeed shared by those workers whose jobs are (according to various commentators) under threat. It is a fantasy not of productivity, but of status, and moreover universally achievable hierarchical status. Every knowledge worker can imagine himself a cyborg, can reap the psychic reward of participating in relation of relative time valuation. And this fantasy is gendered because the historical and cultural status relations that provide the psychic rewards were themselves gendered. Women’s time has always been less valuable.5 With ChatGPT, you can have a secretary again, or a “computer” for English composition. Someone will take care of that for you.
Perhaps unsurprisingly, then, studies persistently show a gender gap in AI usage; according to one 2023 estimate, while half of men report using LLMs, only 37% of women did.
The point of all of this is not to discount the usefulness of LLMs altogether (although I remain skeptical), but to invite those inclined to find the technology compelling, or conversely, terrifying, to ask to what extent those reactions – and the money we as a society are collectively spending in service of those reactions – are the product of status desires and the symbolic, psychological rewards of AI’s use rather than its material benefits. And then to ask whether, given the energy-intensiveness of LLMs, given the hollowing out of intellectual property rights through “scraping,” and given the way use of LLMs incorporate and in some cases intensify racial and other forms of bias, whether these psychological wages really are worth the costs.
Sam Bankman-Fried apparently believed his massive grift was justified by the need to forestall an AI apocalypse by…developing superintelligent AI faster?
If I am right I think it is a telling irony that technology appealing at the emotional level to its user’s conception of their own genius has found its most vehement opponents–rightfully so–in those who actually work in creative professions. But of course, that opposition arises from the fact that LLMs aren’t actually capable of creativity and any appearance thereof is the product of systemic intellectual property violation.
And–to go a step further–the use of AI has become so pervasive and problematic among college students, in part, because it taps into a pre-existing tendency in discourses around academic meritocracy at the undergraduate level to abstract ability or capacity from the practical product or outcome of their work. The “brilliant” student (usually male) who never seems to apply himself, who never lives up to his potential, but who nevertheless is recognized for his intelligence is an absolutely pervasive trope in K-12 education.
Fun fact about actor-network theory: Despite being the most post-modern-y of all the post-modernism (my paper on the subject was called, “Do Rocks Have Agency?”), I was able to get copies of its seminal texts not at the main UCBerkeley library, but the business school.
And, I should say, in case it is not obvious: A relation can be gendered, that is, comport with stereotypical and historical gender roles, even where the participants’ genders are non-canonical.