Patricia J. Williams: “When AI Speaks for the Dead”

But the Trump administration’s executive order focuses solely on deploying—not studying—a specific subset of computer technology. This uncritical application of a machine (however “intelligent”) feels risky and hubristic. After all, generative AI is still in its infancy, and its immensely powerful workings are so little understood that our secretary of education initially referred to it as “A1.” Generative AI exceeds the contours of what we have come to think of as mere computer science: it is a technology whose great powers are still developing and whose restructuring of social interaction is almost completely uninterrogated. It is exciting for its potent capacity to recognize patterns across immense fields of data, which is a boon in quantitative and epidemiological fields like medicine, mathematics, and physics. But its vast surveillance powers threaten what’s left of privacy rights, its energy consumption is environmentally devastating, its unrestrained scraping of the Web decenters libraries and violates copyright, its output includes notorious “hallucinations,” and its mistakes are intentionally made unaccountable by virtue of its “inscrutably complex” black box design. (Here’s one example of AI hallucination: the Chicago Sun-Times recently published a summer reading list—but only five of the fifteen recommendations were real books. The paper had just gone through budget cuts: apparently, human reviewers cost too much.)

As I write, a massive budget bill is moving through Congress that contains a stealth provision prohibiting for the next ten years any state laws designed to regulate artificial intelligence. We will see whether this survives passage into law, but the very inclusion indicates a resounding industry interest in shielding itself from democratic input. Rather than apply even basic cautionary principles in unrolling such a transformative force, governments around the world are investing huge amounts of capital based on faith that this “technology of the future” is an unqualified good thing.

We who teach, read, and write for a living—and anyone who cares about the survival of such skills—are left at the practical limits of entirely predictable quandaries. This messy situation implicates not only academic integrity but also the very concept of authorship, the nature of knowledge production, the animacy of machines, and the consistency of citational practices. When I first read Trump’s AI education order, I was left wondering what, exactly, the capacious, evolving, and amorphous notion of “AI” might mean. The order looks to a trade statute for its implementation, specifying that “AI” shall be defined by 15 U.S.C. Section 9401(3). So I searched for that law, typing the code and section numbers into my browser. Google Chrome’s search function did not bring me directly to the text of the United States Code but instead led with a weirdly recursive “AI Overview” of the definition of AI:

The term “artificial intelligence” means a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. Artificial intelligence systems use machine and human-based inputs to—                                                               

     (A) perceive real and virtual environments;
     (B) abstract such perceptions into models through analysis in an automated manner; and
     (C) use model inference to formulate options for information or action.

For purposes of publication, my footnote would be identical, but it would refer to two different statements. This challenges the very meaning of a footnote as a source fixed in form and static in time. Maybe I should not call it a footnote at all. Perhaps it should be given lesser status as merely a “prompt” to an unstable, ever-in-flux mechanical intelligence.

As though to assuage my worry, at the bottom of both—or all—versions of “AI Overview” is a caution in small print: “AI responses may include mistakes. For legal advice, consult a professional.”

Handily enough, I am such a professional. The law is a profession based on the consistency of precedent—in other words, on the reliable sequencing of history as narrative referent. The law is also premised on rituals of veridicality and proof, due process, evidentiary standards, human encounter, and recognition by sentient peers. I worry how American laws, knowledge production, and habits of thought may be unsettled or destroyed by new AI interventions in legal processes.

Indeed, I question why a search engine’s list begins with an AI summary rather than the law itself. AI overviews are posed as something like syllabi or abstracts: quick explanations that save time and distill complexity. In effect, this engineered ordering trains readers and researchers to start with a machine-generated reduction, the search landing of first resort. At the same time, most AI answers are not reproducible. Each generated answer is unique, which makes the challenge presented by AI summation different from the challenges to citation-as-proof presented by other recent technological forms, such as dead URLs, deleted social media accounts, and live broadcasts. Though mechanically irretrievable, those latter sources still refer to fixed events—things that have happened only once. Generative AI, on the other hand, is a still-evolving technology and, at this point in time, is more like an unreliable witness whose memory wanders here and there in a manner that undercuts forensic testimonial practice as well as temporal ordering.

The venerable style guides for academic and professional publishing continue to grapple with the question of legitimation that this presents: Should a message generated by a bot be cited like the product of a tool—a calculator’s answer to a math problem, say, the output merely reproduced by, rendered by, delivered by, and attributed to you, the living author? Or should the bot be cited as a separate humanoid content producer—the equivalent of an independent or corporatized “author”?

The idea of autonomous authorship relies on complexly webbed assumptions about the Cartesian division between abstract mind and carnal body, about individualism as linked to specific expressive capacities, and about mortality and the passage of time as marking the termination of such capacities. We have built whole institutions around the idea of an author as the sole living, self-directing presence behind a work composed within a particular historical moment. This much has given rise to endless philosophies of reading—from biblical fundamentalism to New Historicism. Consider the gallons of ink spilled trying to apply just one of those theories: originalism, or the jurisprudential principle that a document should be interpreted according to its “original” meaning as its author intended at the time it was written. Other theories, like legal realism, hold that denotative meanings can be challenged by evolving connotations that allow for political or practical understandings to widen over time or beyond the literal. In all such theories, however, we rarely question the assumption that there is an originating voice, a singular creative genius, an identifiable force whose intellect is the font of all such “intellectual property” and whose body is firmly located in a sequence of calendrical time.

As the historian Paul J. Kosmin observes, our contemporary notion of history depends on archival practices that “turn the past historical: dates allow things to happen only once; and dates insist on the ordering and interrelation of all happenings. An event must be chained to its place in time before it becomes an available object for historical articulation.” I am not suggesting that this way of organizing time is inevitable or preordained; Kosmin contrasts our present calendrical system with a range of alternatives—ancient forms of temporal tally paced, for example, by the lifetimes of kings, seismic events, seasonal changes, or astronomical cycles. I simply mean to acknowledge that systems of time have quietly deterministic power in directing our attention and weighting our values. And, whereas every aspect of our contemporary civilization is ordered sequentially, generative AI seems to have a less sharply delineated sense of time. Its outputs have a subtle tendency to slide around, sometimes treating probabilities as present rather than future, and sometimes denying what we presently comprehend as the pastness of the past. Like dreams, AI allows for unconventional temporal disruptions, interruptions, and suturings.

My soul has its alarm turned off. No my soul has this knot in its throat—or is it a

gag—pacified, petrified, up all night counting silently toward infinity. Losing its

place. How many of us are left. What else could happen. Has it all already 

    happened.

Who is they. That autocorrected to thy. Why. No matter what I say it fixes it. It’s

    fixed.

Meanwhile, BBC Maestro’s site trumpets “an unparalleled opportunity to learn the secrets behind [Agatha’s] writing, in her own words.” The course’s legitimacy depends on differentiating between legally permissible appropriation and outright theft or inventions spun from thin air. But her descendants’ approval of this teaching method is not the same as a sign-off by the real Christie, and the decontextualized words that the avatar performs aren’t Christie’s own. “Agatha” is a digital prosthetic for BBC Maestro, and for the incorporated entity Agatha Christie Limited, whose CEO is her great-grandson. “Agatha” is not a prosthetic of or for Agatha Christie herself, who has been entirely disabled by the fact of death. “Agatha” is a tool for investors, a violent puppet excused by the narcissism of the living. It masks the agency of her heirs, who have remolded her memory in their own image and for their own ends. We are enticed into a peculiar necromancy, one that will probably seem peculiar only for a little while before it becomes a comfortable habit of relation to the dead.

From reconceptualizing copyright restrictions to regenerating the deceased, new technologies—and the laws that give them clearance—displace the role of death. They disrupt our apprehension of epoch and dislodge thought from embodiment. If everything is everybody’s or nobody’s, if all literature and all humanities are treated as just jumbled masses of information without regard for autonomy or time, then this is a kind of betrayal of the very idea of archives as repositories of human experience and layered memory. AI’s flattening logic sometimes reminds me of the Swiss comedian Ursus Wehrli’s project Tidying Up Art, in which he reorders the “mess” of a Joan Miró painting. Miró’s forms are dropped “any old way” on a yellow background; Wehrli sorts all the shapes “neatly” by color and then arranges them in tidy columns that are “more economical” and “more efficient.” He proudly announces that his version “takes up far less space—Miró could have saved canvas for another picture.” Wehrli also takes apart a noisily colorful Keith Haring painting and reconstructs it into what he describes as “Keith Haring looked at statistically,” with detached rows of “twenty-five pale-green elements,” “twenty-seven pink squares,” and so on. He brags that this organizational method would allow the artist to “estimate how many pots of paint he’s likely to need in the future” and concludes: “One can obviously also make combinations—for example, with the Keith Haring circles and Kandinsky’s dots. You can add them to all the squares of Paul Klee. In the end, one has a list with which one then can arrange. Then you categorize it, then you file it, put that file in a filing cabinet, put it in your office and you can make a living doing it.” What Wehrli describes is simply a primitive, bare-bones version of AI—which “makes a living” for its developers by scraping digitized literature, treating it as “training data,” melting it all down into large language models, and then extracting from that infinite literary mine to create new, “more efficient” formations.

Walter Benjamin wrote that “every image of the past that is not recognized by the present as one of its own concerns threatens to disappear irretrievably.” In today’s technologically accelerated political discourse, a great body of memory seems to be on the brink of irretrievable disappearance. Whether as infrastructure in our educational system or as guide in our daily lives, AI, with its sleek reordering of time, narrative sequence, and attribution, is creating new challenges to historicity and truth. And so we find ourselves at a paradoxical crossroads of nostalgic yearning and induced oblivion. We watch, surprised by our helplessness, as our educational system and even the precious hereditament of the Library of Congress—the largest library in the world, with more than twenty-five million volumes—trembles before the chainsaw of AI’s most efficient dreamers of dystopian, extropian life.