3.1Discussion
The question was what a record does when it puts two goods on one scale and then counts the cost, and what its labelled ladder does to the choice. It read the sixteen AI-generated records of Cosmic Press's last file, every one of them set in a hypothetical situation, and it answers in two parts.
The first part concerns the record's account of its own costs. Every question in the file but one weighs one course against another (the night-patrol record asks what an activity means), and every record then faults the weighing. The fault is specific and regular. Its usual charge is that the question set two poles apart and lost the ground between them, or looked at one side; the cure it offers is a compromise, a balance or a backup. The counterquestion then proposes the missing course, and does so in the same register of better and worse in which the question was asked: a question about what is wise or safe meets a counterquestion about what is wiser, safer or better. Read with Espeland and Stevens, the records are a sustained exercise in commensuration that also names its costs. Read with Sunstein's distinction, they grant that values are plural and keep them on one metric. Read with Anderson, three records come close to saying that a good such as love or freedom was taken in only one form, and even these turn back to dividing the good better. Read with the work of Fiske, Tetlock and their colleagues on taboo trade-offs, the records weigh things that those works treat as hard to weigh, a household's means against riders' safety, a partner's fatigue against an act of love, and register no strain. The records cannot say whether any reader would. They show a form that has no slot for refusing a trade.
The second part concerns the ladder. The labelled notes state rather than ask, and what they state is one course, six times over. They add common measures of their own: the body at the level of Life, power used or saved at the level of the Earth, energy, order, entropy and signal at the level of the Universe. The top note, though labelled Universe, speaks of energy, order, entropy and signal, and it picks out one course, sometimes the question's and sometimes the counterquestion's, without saying which. Through Potochnik and McGill's lens, the ladder is one fixed hierarchy laid on every choice alike, which is the trouble they find with universal levels. Read with Wimsatt, the top note is an aggregative description that drops the organization the question was about. Read with Kim's work, a physical description that marks no difference between the courses leaves the choice where it was.
Put together, the two findings show a record at odds with itself in a precise way. Its defect line criticises the question for reducing a choice, and its ladder then reduces the act it describes, twice over, to measures the question never used. The record's self-criticism operates on the question and not on the ladder. Nothing in the record faults the ladder, and the form gives it no place to be faulted.
3.2Contributions
The paper makes two contributions, stated here as it stated them at the start. First, it describes a generated record form in which commensuration and a critique of commensuration are built into every record, and it locates the critique: the records fault the split and keep the scale, so that their self-criticism belongs to the commensurating side of the debate rather than to the side that holds some goods to share no metric. Second, it describes a labelled ladder of levels used on choices, and shows what it does: it restates one course instead of carrying the question outward, it brings in further common measures at the levels of the Earth and the Universe, and its top note, a physical description, marks no difference between the courses. Both contributions concern how this record set frames choices. Neither says anything about lifts, cats, love or votes.
3.3Limits
The sample is one file of sixteen records from one stretch of one catalog, chosen by a rule that follows the studio's own division into files. It shows the closing stretch well and the rest of the catalog not at all; the counts for the earlier stretches serve only as contrast. The eight entries of the stretch that the rule leaves out share the sixteen's labelled ladder, past-tense defects and suggesting counterquestions by the word tests, but they were not read.
The records are AI-generated and their situations hypothetical. The paper reads them as texts in a fixed form and draws no fact about the world from them. In particular it takes no view on the truth of the physical, biological or social sentences in the scale notes; it reads them only for what they do to the choice. No course in any record is endorsed or opposed here, and nothing in the paper is guidance on safety, money, health, housing or law.
The English is a translation made once by this study's editorial session. Every count is made on the Korean to keep the translation from deciding it, but the readings in Sections 1.2 and 1.3 rest on the English. The counts are crude word tests. A test for "only" catches some phrases where the word does other work; a test for physical vocabulary takes "evolution" and "cells" as physical when a biologist would not; a test for a "whether" clause depends on a spelling pattern that misses some forms. The tests are listed beside their counts so that a reader can check them.
The readings through the works are this study's. The works are lenses chosen for this paper. Several were written about people's reactions or about scientific practice, not about generated texts, and Kim's work concerns minds and bodies, not choices; where a lens fits badly the paper has said so, and its use of each goes no further than the work's own summary.
3.4Further research
Three lines of work would test the findings. The first is to read the rest of the catalog with the same two questions: whether earlier stretches, whose questions are not posed as two-way choices, fault their framing in the same way, and whether their ladders, when they ask instead of stating, carry the choice anywhere the labelled ladder does not. The second is to set the records beside human-written decision records of a similar size and ask whether the critique that stays inside the common metric is a feature of the generated form or of decision writing in general. The third is to put the records in front of readers and ask whether any of the weighings strike them as the kind of trade the taboo trade-off studies describe; the records themselves cannot answer that.