Chapter 1. Questions Before Answers

The sixteen Dionyas inquiries on AI and creative control decide one thing firmly: that the first step in any dispute about AI in music is a record of who did what. Everything else, including who would judge and by when, they leave open. This chapter shows how the records do this and what it costs, and it adds to the two debates set out in the Introduction: the study of problem formulation, which rarely sees formulation published on its own, and the debate on AI in music, which argues over rights that these inquiries keep translating into documents.

1.1Data and method

The data are Dionyas's public list of inquiries, fetched by plain requests and kept as snapshots in the study's folder. The list holds 1,539 inquiries across 32 domains. They are of five kinds: 759 cases, 500 historical questions, 256 general designs, which the list calls foundations, 18 chart questions and 6 signal questions. Each sits in one of three lanes, 1,276 in the evidence lane, 250 in a qualitative lane and 13 in a symbolic lane. The list describes its own type as editorial question design and gives its count of generated judgments as 0. A record has six fields that this study reads: a title, a question, a scope that says what the inquiry covers, a list of the material it needs, a boundary that says what it may not claim, and a state.

The selection rule is one sentence: every inquiry whose leading domain is AI and creative control and whose lane is evidence. The domain holds 18 inquiries; the rule keeps 16. The 2 it leaves out are in the qualitative lane, and their scope marks them as synthetic questions, short scenes written for a simulation; they were left out by the rule, not by reading them first. The sixteen fall into three groups. Eight are general designs. They name no artist, work or market, and their scope says that these will be named when the inquiry is put to use. Four belong to a series of designed industry cases that runs across the whole list; one of them calls its own setting a hypothetical production design. Four are editorial cases about a named company, platform or recording: an agreement between two named companies, figures announced by a named company about AI uploads and streams, a named recording, and a named platform's policy, each as its record frames it. The appendix gives every record in full.

The records are in Korean; every quoted sentence below is from one English translation made for this study, sentence by sentence, with nothing added. Dionyas does not say, for these sixteen, who drafted the wording; it calls them editorial. The two synthetic inquiries the rule leaves out are the only ones in the domain whose record says its source card was a model's draft checked and registered by RISCON, and this study draws nothing from them about the sixteen. What the four named cases say about real organisations is how the inquiry frames its subject, and this study has not checked any of it. It draws no fact about any company, recording, figure or law from the records.

Two properties have to be set apart before reading. The first is the studio's form: every inquiry has a scope, needs and a boundary, and every one of the sixteen ends its question with a question mark. None of this is a finding. The second is the studio's frame, wording copied from record to record. The series this paper belongs to measures frame by five-word runs, counting a run as frame when it recurs across a site's records. By this study's count the questions of the whole list contain no frame at all, 0% of their runs. The boundaries are 41% frame and the needs 30%, and the scopes 46%. Inside the sixteen, the site threshold catches less, because a domain's lines are too rare across the whole list to pass it. Counting instead the runs that recur in at least 3 of the sixteen, the boundaries are 64% frame and the needs 68%.

The reason is plain once the domain table is read beside the records. Every one of the 256 general designs in the list takes its boundary, word for word, from its domain's caution line and its first need from its domain's evidence line; 256 of 256 do both. Every one of the 148 inquiries of the industry series adds to its boundary the same sentence about what the number of questions does not mean. In the sixteen, then, the boundary and the first need of the general designs, and the opening of the boundary in the industry cases, are the studio's frame. This study quotes them once to show the frame and does not read them as evidence. It reads as evidence the questions, which are each record's own; the scopes and needs of the cases, where they differ from the domain lines; and the boundaries of the four named cases, which are written for each case.

Every number in this paper comes from a script that reads only the snapshots, by simple word tests on the Korean original. The counts describe the records and nothing else.

1.2From rights to records

The first finding is a move the inquiries make again and again: they take a question that sounds like a question about rights or authorship and turn it into a question about what should be written down. The general design that opens the domain shows the move in its plainest form.

The title speaks of dividing assistance from replacement, a contested line in any argument about authorship, but the question does not ask where the line falls. It lists four uses of a tool, from a suggestion to a whole piece, and asks at which stages the person's choices and the tool's output should be recorded. In Runco and Nemiro's terms, the problem has been expressed and partly constructed, but its definition, what would count as replacement, is left to whoever reads the record later.

The second general design pushes the same move into the past tense of the work.

Here the record proposes that the creative judgment lies in the refusals as much as in the result: a finished song hides the candidates it rejected, and the question's condition is the documentation that would show them. The industry series repeats the idea in the setting of a songwriting collaboration, and adds a second question that names the claim the documentation is meant to resist.

The second sentence is the only point in the three where rights appear, and they appear as a shortcut to be doubted: having made the final edit does not, the question implies, settle the matter. Read beside Brand, Allen and others, Beyond Authorship (2015), these inquiries ask for something close to a contributor-role record for music. Brand and her co-authors argue that, as the number of authors on scientific publications grows, an ordered list of names is inadequate, and that identifying specific contributions will lead to appropriate credit and fewer authorship disputes. Their subject is scientific publication, not music. The Dionyas inquiries never use the word taxonomy, and they do not propose roles. But their unit is the same: not the author but the contribution, stage by stage, kept, discarded or rewritten. This study reads that as the inquiries' central design choice.

The same move governs the inquiries about terms of use and the inquiries about suppliers. A general design asks what a tool's claim of commercial use leaves unchecked, and an industry case asks what a paid plan can be said to grant.

Both questions grant a premise, that the terms permit use, and then turn the apparent permission into a list of further checks, without saying what they would find. The industry case on outsourced work turns the list into an object, a checklist, and adds a question about how the absence of evidence is to be marked.

The second sentence is the sharpest in the set, in this study's reading. It concerns not what was done but the status of a statement about what was done: a contractor's reply that there is no record of use is itself a record, and the inquiry asks that it not be recorded as more than it is. The concern is with the grade of a document.

Consent is where the move is strongest. Consent appears somewhere in 15 of the sixteen records, but in the questions themselves it appears in 4, and rights or authorship in 4. In the rest it sits in the list of needs, often in the domain line the general designs share. Where consent does enter a question, it enters as scope:

The question lists five dimensions of a permission and asks how each can be confirmed separately. Then it asks whether one permission may stand for another. Sheehan, Can Broad Consent Be Informed Consent? (2011), argues, for biobanks, that broad consent can be informed consent, and that the distinction lies not between kinds of consent but between the kinds of choice a person makes. The work fits the Dionyas question badly, since it concerns biobanks and not music. This study reads the question as dividing one permission into items and asking whether a narrow consent may be widened. It does not answer, and neither does this study; it makes the scope of a consent the unit to be documented, the same move again from the right to the record of it.

One general design comes close to Drott's concern with what is put into a system rather than what comes out.

The inquiry looks at inputs: unreleased sessions, contracts, a voice. It names training and reuse. But it frames the input as a matter between a producer, a tool's policy and the people whose material is involved. Drott, Copyright, Compensation, and Commons in the Music AI Industry (2020), argues that the material put into these systems is the shared knowledge of a music community, a common-pool resource for which existing copyright regimes are inadequate. The Dionyas inquiry stays inside the individual frame: it asks whether each party's consent was checked, not who is owed what by a system trained on many. The voice-model inquiry, which names retraining, does the same. None of the sixteen asks Drott's question of distribution, and none asks Morreale's questions about the cost of creative labour or the exploitation of existing music. This is a finding about the design, not a fault in it: the inquiries are written for a producer deciding about one work.

In Volkema's terms the inquiries combine expansion and reduction. They expand each problem toward what it rests on, the stages, versions, consents and checks behind a finished track. They reduce it by excluding questions of distribution and of the industry as a whole. A count shows how the expansion is phrased: 14 of the sixteen questions ask how, which, what or how far, and only a few ask whether. The questions ask for a procedure more often than a verdict.

1.3A date without a deadline, a boundary without a reopening

The second finding concerns the three things a decision needs beyond a question: someone to make it, a time by which it must be made, and a condition that would make it be looked at again. The sixteen inquiries supply a version of each, and each version is displaced.

Begin with the decider. Seven of the sixteen questions name a party of some kind: a producer, a singer, a contractor, an audience, an artist, a platform or a company. Five speak of a person or of the human in general. But a party named in a question is usually the subject of the inquiry, not the one who will answer it. Only one general design gives the work of deciding to someone:

The producer here is the one who must tell two things apart. Yet even this question asks how they are to do it, not what they should conclude, and the record does not say whether the producer's telling would settle anything for anyone else. Elsewhere the decider dissolves: the question on disclosure makes the audience a party whose wishes are the subject, not a decider.

The state of each record is consistent with this: none reports a decision. Of the sixteen, 8 have their question designed and await inquiry, 6 have their question registered and await judgment, and 2 have their sources under review. Across the whole list, 715 of the 1,539 records, 46%, say they are awaiting judgment. The design leaves a place for a judgment and names no one to fill it. Dery's point about the power of initial definitions applies here in an unusual way. The first definition of these problems is set by the studio, which names no decider; the decision itself is left wholly open, while the frame in which it must be taken is fixed in advance.

Time appears in two forms only. The general designs leave it to be filled in: their scope says that the artist, work, market and period are named when the inquiry is recorded. The named cases give dates, but the dates belong to sources, and this study reports them only as the records give them. Four questions carry a year, and all four belong to named cases: the date of an announcement, a release or a policy, as the record says. One scope adds a date of review.

The question, as the record words it, sets an announcement against the public documents that followed, once more the move from a right to a record. The time it carries is the time of the announcement and of the review, as the record gives them, not a deadline for anyone. No one in the sixteen is given a date by which to decide.

What, then, would change an inquiry? The records have no field for reopening. What stands closest is the boundary, and in the named cases the boundary is written for the case. Read together, the four named boundaries show a pattern. Three of them state what a source cannot be taken to establish and what the inquiry itself does not do; the fourth, taken up below, opens with a definition.

The second boundary does three things that a reopening condition would do, without being one. It dates its own figures and says they are not the latest, which implies that newer figures would change the inquiry. It fixes the class it is about, fully generated tracks, and forbids widening it. And it refuses to treat the company's own detection as independently verified truth. The question it guards is itself an either-or about explanation:

The figures in this question are the company's announcement as the record reports it; this study does not use them as facts. What matters for the design is the structure. The question offers two explanations, and the needs ask for aggregates by track age and route of arrival and an account of the labelling and recommendation methods: the material that could tell the explanations apart. The boundary then says what the answer will not be allowed to claim. A condition for reopening looks forward to what would change a judgment. A boundary looks sideways, at the claims a judgment must not make. The sixteen inquiries have the second and not the first.

The two named cases about a recording and a platform show the boundary doing definitional work. The case on a named recording asks about the order of a production and the reach of its credits:

The first sentence of the boundary is not a limit on the inquiry. It is a definition, stated as settled before any evidence is read: separating a voice from an old recording and generating a voice are different kinds of work. This study reads it as the clearest instance in the set of what Dery calls the power of initial definitions. The question asks how far the credits can tell the two apart; the boundary has already decided that they are two. The platform case names the same recording as its point of comparison; this study reads that as the earlier inquiry's distinction between separation and generation serving as a premise of the later one.

Here the decision has a decider in view, the platform, whose confirmation the question names, as the record words it. But the question asks whether the decision can be made without a production history, and the needs ask for the policy in force and the records of versions, tools and consent that one has the right to use. The decision is again routed through a record. The boundary then separates acceptance by a platform from the existence of a right, which is the same separation the domain's frame line makes for every general design:

This is the frame, quoted once. It appears in all eight general designs and opens the boundary of all four industry cases. In the industry cases a second frame sentence follows it, the same in all 148 members of the series:

The two frame sentences are telling as design though they are not evidence about any one inquiry: both are limits on inference. A studio that writes no judgments has standardised, above all, the things its inquiries may not be taken to say.

Two general designs show what the design leaves out when no boundary of its own is written. The inquiry on correction asks how a comparison should be read:

And the inquiry on cost asks what is actually saved:

Both carry real choices, a method of comparison and a full account of costs, and both are fenced by the domain's line about assistance and rights, which fits neither closely. The frame protects every inquiry from one kind of overclaim, and leaves each to say nothing about the overclaims particular to it. That is the cost of standardised formulation that Baer and his co-authors' design goals would lead one to expect, read here by this study: a shared mechanism widens the questions it asks of every problem, and narrows what it can say about any one.

The chapter's two findings fit together. The inquiries move a decision about AI in music from rights to records, and replace the decider, the deadline and the reopening condition with a source date and a boundary; both moves serve an inquiry that can be published without claiming more than it holds. For the study of problem formulation, they add a case in which formulation is a finished public product. For the debate on AI in music, they show where an editorial design places the question: on the single work and its documentation, short of the distributive questions Drott and Morreale raise.