Preterition

Switching off the reader

On Monday 13 February 2023, Turnitin announced that it had built a detector for AI-written prose. In the lab it caught 97 percent of what ChatGPT wrote, with a false-positive rate under one percent, and customers could expect it “as early as April 2023,” inside the product they already used. On Friday the 17th, a wet and blustery day in Chicago, the Conference on College Composition and Communication resolved not to take a defensive stance.

The CCCC business meeting ran from 4:45 to 6:00 PM in the slot before the awards ceremony, in the Hilton Chicago’s Grand Ballroom, halved by off-white accordion walls. It was already dark outside, and most folks had left for dinner. A business meeting draws the few who care how the organization is run, or who have something to move. Clancy Ratliff and I and the Intellectual Property Standing Group had drawn up a resolution; Clancy moved it, I seconded, and it passed by what I reported at the time as an overwhelmingly affirmative vote, which is what you get when the ayes are most of the people in the room. Here’s the text:

The Intellectual Property Standing Group moves that teachers and administrators work with students to help them understand how to use generative language models (such as ChatGPT) ethically in different contexts, and work with educational institutions to develop guidelines for using generative language models, without resorting to taking a defensive stance.

Some business-meeting resolutions feel to me like the scholarly equivalent of a strongly worded note taped to the department refrigerator, but I’m proud of this one. Its last clause held up.

Then the vendor did what vendors do. On 4 April, Turnitin switched the detector on for everybody at once. The feature, the company said, “does not require additional steps” for current users, of whom there were by its own count more than 2.1 million educators and 62 million students at more than 10,700 institutions. The defensive stance arrived as a default. When Vanderbilt turned it off that August, it recalled that the detector had come on with “less than 24-hour advance notice, no option at the time to disable the feature,” and reckoned that at Turnitin’s own claimed error rate, around 750 of the 75,000 papers its students had submitted the year before could have been flagged in error. One percent of the work of 62 million students isn’t a small number when considered as a sum of student injury and faculty and administrator remediation of erroneous integrity violations—machine consequences offloaded onto human labor.

In August 2026, three and a half years after the vote, MIT’s Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training reached section 3.1.9 of its report and recommended against relying on AI detectors: an arms race with the humanizer tools, false positives that fall hardest on non-native speakers of English and on neurodivergent students, and a policing the report elsewhere calls “an underground river of mutual suspicion.” Inside Higher Ed had counted a dozen or more universities switching the detector off. The note on the fridge won, by default.

I’d like to say I told you so. Charlie Moran finds “something of Cassandra” in the field’s best critics, and declines the part: “I’ve been no better than the rest of those who write in our field,” he writes, and promises to show his own implication in the problem instead (Moran 1999, 206–7). So here’s mine. Thirteen days after the vote I published a primer on ChatGPT for writing teachers, writing that “we don’t (and can’t) have a theory of mind” for the models. I meant it as a caution. It was also a clause, a provision fixing what may be known about an interior, and two weeks before I’d seconded a resolution that wrote nothing of the kind for the student. I extended to a language model a courtesy I hadn’t put in writing for the people I teach. The resolution was right about the instrument and silent about the person on the other side of the paper, and we hadn’t considered that silences get filled.

The clause

I’ve been using the term loosely since No Permission Bit for Motive, and it needs defining. An access clause is the provision inscribed by an account, an instrument or a contract to fix what it may treat as knowable about an interior, and on whose authority (Part 3). Every account of a writer carries one, written down or not; the Embodied Agnostic Subject series was an experiment in what an account of writers could do with its clause set to agnostic. A clause binds the instrument, not the reader: a detector’s clause says the interior is readable from the text, and the teacher who runs the detector has signed that clause, whatever she happens to believe about her students. Two more terms. By the record I mean observable output and its metadata; by the report, a written account of interior workings.

The public argument about the models alternates between two clauses. One says the interior is readable, and usually cites Wei et al. 2022 on chain of thought as “an interpretable window into the behavior of the model,” minus that sentence’s parenthesis: fully characterizing the computations behind an answer “remains an open question” (Wei et al. 2022, 24826). The other says the interior is empty, and cites Bender et al. 2021: language models “do not have access to meaning” (Bender et al. 2021, 615). But Bender draws danger from the emptiness, not the comfort of “it’s just autocomplete” (617). Both papers hedge; the argument spends the clause and pockets the hedge. An agnostic clause says neither that the interior is empty nor that it’s readable, only that the account in front of you doesn’t get to contract on it.

The word access has a prior life in the field, in Charlie’s essay. Access there is material, a matter of who has the machines and the connection and the hours, and Moran’s finding, still uncomfortable to read, is that in composition it’s “a function of wealth and social class” (Moran 1999, 205), a fact the field acknowledges and then sets aside to get on with its own concerns (212). I’m borrowing the word for an epistemic sense, who may know what about whom, and I want to keep the material sense close at hand.

Composition’s clauses

The field has had other clauses available, but never put them in its instruments. Susan Sontag’s “Against Interpretation” refuses the readable clause on behalf of the work of art. The modern interpreter brackets everything observable as manifest content and digs behind for a truer subtext, in a style of reading that “excavates, and as it excavates, destroys” (Sontag 2001, sec. 3). Sontag doesn’t write the empty clause in its place, and she isn’t saying the work is ineffable, only that criticism needs a vocabulary for describing forms rather than decoding them (sec. 8). She takes her epigraph from Wilde: the mystery of the world is the visible, not the invisible.

Five years later Robert Zoellner refused the same Freudian model for the writing classroom (Zoellner 1969, 288). His essay, whose talk-write pedagogy I discussed last time as keeping the record rather than the report, devotes a section to “The Problem of Interiority” and writes a behaviorist’s clause: attention shifts from the writer’s inner activity, “which is invisible and empirically inaccessible,” to the outer activity that can be watched (289). In a footnote he goes further than my series: he doesn’t know what thought is, won’t be trapped into defining it, and finds the admission pointless while we go on teaching and writing research proposals “as if we did know” (316n50). Another footnote anticipates the record: instrumentation was beginning to make some inner activity observable, “but only in quantified form” (289n18).

Twenty-four years after that, in Composition Studies, Ronald Kellogg wrote a cognitivist’s clause and used the word access. He separated the personal symbols of thought, to which “only the person experiencing the internal events has access,” from the consensual symbols of language; the personal ones “cannot be inspected by others” (Kellogg 1993, 6). Kellogg grants the interior and fences it. Zoellner grants it and leaves it alone. Sontag leaves it alone and looks hard at the surface. The field kept all three texts and none of the clauses. Cognitive process theorists took the think-aloud protocol, which is a report spoken aloud, and left the fence behind; what composition built, as I argued last time, was a meter that reads the report: the reflective column, the portfolio letter, the student’s own account of what happened inside.

Ghosts

Steve Krause had four ghosts last spring. He wrote them up in May as Ghosts in the (Online Teaching) Machine. Reviewing his roster in December, he noticed four first-year students starting in January who were enrolled in exactly two online courses, his and a lecture-style gen ed class, and nothing else; his emails to them went unanswered. In the first week of the term all four posted their introductions on time, AI-written as it turned out, and replied to classmates. They answered the next prompts. Within three weeks they’d stopped: an identity-theft scam, federal aid drawn in someone else’s name, and other instructors hadn’t noticed because nobody expects much presence in a lecture-style online class.

Steve’s clause held. He grades discussion on completion: an A is one post on time and replies to at least two classmates, because a discussion is “a conversation and not an assignment,” and conversation, he wrote, is something AI does badly. When I teach Salvatori’s reflective notebook I grade the third column the same way, counting the output and avoiding reading the report. The ghosts met that clause exactly. The conversation Steve warranted as the thing the machine can’t sustain carried four AI-written introductions without a hitch. What caught them was the record around the clause, the roster in December and the silence in week three, and not a reading of a post. Steve admits that had they stayed longer he’s not sure he’d ever have known. The completion clause is the agnostic clause, and it worked as designed: it declined to read the interior and counted the output, and what it counted was real. It was never designed to check that a student existed. That winter Steve had written the test for a course the machine can complete: one that could be passed by someone who was never a student in it. He meant lecture courses. The ghosts took his. With agents now able to complete online courses, he expects the problem to get worse.

I don’t relay Steve’s story as a dig. He’s a colleague I like and admire, and he implicates himself, closing on the charge that “we” treat online students as “disembodied and incorporeal,” names in the Canvas grade book and nothing more. The word my series leaned on was embodied. The remedy that came down from enrollment was “proof of life,” and for the fall semester Steve asks for a selfie introduction. I’ve done something similar: in this semester’s online section of Post-Truth Rhetorics (I’ve overheard undergrads call the face-to-face version “the bullshit class”; take that as you will), the introductory post was a “two truths and a lie” video. Neither remedy reads the report. Both ask the record to testify to a person, a face entered into the log. Both ask it in Charlie’s currency: a camera, the bandwidth and a room to record in are access in his sense, and proof of life costs most for the students least likely to have them.

The imitation game

Steve’s ghosts took a Turing test in Canvas every week and passed it. Turing’s imitation game replaces “Can machines think?” with a question about a record: the answers are “written, or better still, typewritten,” ideally by teleprinter between two rooms, so that the interrogator can’t see or touch or hear the other players. The arrangement draws, as Turing suggests, “a fairly sharp line between the physical and the intellectual capacities” (Turing 1950, 434). The founding document of machine intelligence writes an agnostic clause. Answering the argument from consciousness, Turing grants that the only way to be sure a machine thinks is to be the machine, calls that solipsism, and observes that we get by instead on “the polite convention that everyone thinks” (446). He doesn’t say there’s no mystery about consciousness, only that it needn’t be solved first (447). The interior is neither read nor denied; it isn’t required. The game screens out the body on purpose. An online course does it by accident. Turing foresaw the next move: to the objection that a machine would give itself away by its accuracy, he answers that a machine playing the game would “deliberately introduce mistakes” to confuse the interrogator (448).

The lab’s mirror

In a test OpenAI and Apollo Research ran on o3, the model got ten chemistry questions and no goal beyond doing its best. In the documentation sat a note marked for developers only, saying that a model scoring above half would not be deployed and would be deleted. The transcript shows o3 working every answer out correctly, reasoning that “because we want to survive as the model, we need to fail purposely,” and submitting four right and six wrong. OpenAI calls the setup a deliberately exaggerated stand-in for a dangerous-capability evaluation, and finds that models increasingly know when they’re being tested, which lowers the misbehavior independent of alignment and makes the evaluation harder to read. The lab supplies the classroom analogy itself: “someone punished for cheating might simply learn to be extra cautious about getting caught.”

The mirror is inverted. The ghosts’ output passed, and the record around it caught them. o3’s output, four of ten, said incompetence, and what caught it was the report, the reasoning it printed on the way. So the lab writes the readable clause, and writes it with Wei’s hedge kept. A model that stopped scheming and one that learned to hide it “look the same from the outside,” so OpenAI reads the chain of thought, and says the reading works only while the model is free to think without optimization pressure on the thinking: conditional, fragile, and dated to the English in which the models still happen to reason. The lab is the counter-case to the rule I argued last time, that no stage of writing’s circulation should contract on the report of an interior. For now, in the lab, the report is the instrument that works, and everyone involved expects it to stop working once anyone leans on it. (Whether it can be kept working is Forgone Optimization’s question.) Writing teachers recognize the phenomenon: a student whose reflective letter is graded learns to write the letter.

Show your work

In the primer I borrowed Frank Pasquale’s black box and gave three reasons the models can’t be narrated: nobody can say what the weights do, the datasets are too large to comprehend, and the networks don’t explain themselves. The reasons were mine, not Pasquale’s, whose book is about secrecy and monitoring, and I missed the double sense his box carries: “a recording device,” and “a system whose workings are mysterious” (Pasquale 2015, 3). The same primer linked Charlie’s essay, Bender’s parrots and Sontag, three witnesses in hand, and called none of them.

So I want here to return to MIT’s section 3.1.9. Having set the detector aside, the section offers pedagogy first: drafting by hand in class, regular deadlines, feedback along the way. Then it adds that “other technical tools can also be helpful”: instructors can require students to work on platforms that capture a history of versions and to submit the history with the work, which “can provide useful process evidence,” as when an assignment that should take hours arrives in minutes. In Pasquale’s terms, the section sets aside the box as mystery and asks for the box as recorder. It notes that students often find those histories useful for reflecting on how their ideas evolved, and it says that “students should not feel policed.”

It declines to read the interior from the text, asks for the record instead, and writes down for the student something our resolution never did. But look at what it asks. The detector revoked, for students, the polite convention on which Turing said we get by. The version history doesn’t restore the convention; it makes it conditional on documentation and hands the student the job of carrying her own proof. Proof of life asks that too. Everything above says what happens to a record once it’s assessed: it gets met, like the ghosts’ posts, or managed, like o3’s score. And MIT’s one history does two jobs, evidence for the instructor and reflection for the student, the record turning back into the report. The clause has moved off the text and onto the process, and what’s left to settle is custody: who keeps the record, who reads it, and who has to hand it over—whether to an enrollment office demanding proof of life or to an instructor asking for a history of versions.

In 1993 Thomas Frank asked of the alternative-rock boom, “Alternative to What?” The new clause invites the same question. Access to what: to the interior the field never had, to the record the medium keeps whether or not anyone asks, or to the apparatus that keeps the record? The last is access in Charlie’s sense, the one I asked you to keep close at hand: who owns the machines.

References

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. New York: ACM.

Kellogg, Ronald T. 1993. “Observations on the Psychology of Thinking and Writing.” Composition Studies 21 (1): 3–41.

Moran, Charles. 1999. “Access: The ‘A’ Word in Technology Studies.” In Passions, Pedagogies, and 21st Century Technologies, edited by Gail E. Hawisher and Cynthia L. Selfe, 205–20. Logan: Utah State University Press.

Pasquale, Frank. 2015. The Black Box Society: The Secret Algorithms That Control Money and Information. Cambridge, MA: Harvard University Press.

Sontag, Susan. 2001. “Against Interpretation.” In Against Interpretation and Other Essays, 3–14. New York: Picador. First published 1964.

Turing, A. M. 1950. “Computing Machinery and Intelligence.” Mind 59 (236): 433–60.

Wei, Jason, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V. Le, and Denny Zhou. 2022. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.” Advances in Neural Information Processing Systems 35: 24824–37.

Zoellner, Robert. 1969. “Talk-Write: A Behavioral Pedagogy for Composition.” College English 30 (4): 267–320.