AI-Mediated Generative Relations - Contribution, Control, Capability, and Benefit

Abstract

Once artificial intelligence enters an economy as something more than a tool,
the natural question is who is exploiting whom. With capital, labor, the system,
and users there are twelve directed pairs; adding the firm that builds the
system and separating it from the firm that deploys one gives twenty. This paper
argues that the question in that form cannot be answered, and that the reason is
structural rather than evidentiary. A diagnosis of exploitation assigns four
roles — who controls, who acts, whose capability falls, and who benefits
through the fall — and a directed pair names at most two of them. The paper
proves that topologies can agree on every quantity such a table is usually
populated from — which node acted, whose capability changed, whose return rose
— and still disagree about who, if anyone, exploited whom. It then proves an
aggregation result explaining why the entity list had to be revised twice:
merging a controller with the system it governs manufactures an action
attribution that belonged to neither, and merging a provider with an adopter
internalizes a benefit path and can convert a positive diagnosis into a negative
one, so a merge is admissible for a question only when no edge of that
question’s witness crosses or is internalized by the merge boundary. A third
result concerns opacity. Where corpus membership, the counterfactual capability,
and the benefit path are each unavailable, the observation set consistent with
what is seen contains both positive and negative cases, so the audit is
unresolved; and every observation that would resolve it is held by the party
whose conduct is at issue. Disclosure is therefore a condition on the
framework’s applicability rather than a policy preference, and reading an
unresolved audit as a negative finding is an error with an identifiable
beneficiary. A fourth result concerns speed: two parties with identical holdings
and different expansion rates have different reachable sets at every finite
horizon, so equalizing holdings does not equalize reachability. The paper
also argues that the training corpus is a less diffuse object than the
collective inheritance it is said to condense, since a corpus is finite,
assembled, and bounded by decisions someone made; that the standing of the
system itself is open, and that the framework must remain expressible under both
answers. It classifies no system, provider, or firm, and makes no empirical
claim. A closing section states, item by item, the evidence that any such claim
would require and that this paper does not have.

1. Introduction

The question that opened this research programme was whether exploitation
becomes more aggressive in an era of artificial intelligence. It was posed
concretely. In the classical picture, a worker produces value and receives less
than the value produced, and the difference is the object of the analysis. In
the picture that replaced it, capital spends a small sum on machine-generated
work and realizes a large return, and there is no wage from which a difference
could be taken. The classical accounting has no place to attach.

The first response to this is usually to widen the cast. If the analysis cannot
find its object between an employer and an employee, perhaps it can be found
among capital, labor, the system, and users — four entities, twelve directed
relations, one of which may be the answer. The research discussion that
generated this paper did exactly that, and then revised the cast twice: first by
separating the firm that builds and governs the system from the capital that
finances it, and then by separating the firm that deploys a system from the
provider that supplies it. Each revision was made because a bundled entity was
found to be doing two things at once. With five entities there are twenty
directed relations, and the natural next step is to fill in a table.

This paper argues that the table cannot be filled in, and that the obstacle is
not missing data. A diagnosis of exploitation, on the framework this programme
uses, assigns four roles: a party that controls an arrangement, a party whose
operation is the proximate cause of a change, a party whose effective generative
capability is lower than it would be under an admissible alternative, and a party
whose return is higher through that reduction. A directed pair names at
most two of these. The paper proves that two topologies can agree on every
quantity a pairwise table is normally populated from — which node acted, whose
capability fell, whose return rose — while differing on whether anyone
exploited anyone. The pairwise question is therefore not hard; it is
underdetermined by its own evidence.

The paper’s second result explains the revisions. Aggregating two entities into
one is a modeling decision with diagnostic consequences, and the consequences run
in both directions. Merging a controller with the system it governs produces an
aggregate that appears to act on a third party, an attribution neither
constituent had. Merging a provider with an adopter internalizes a benefit path,
and a diagnostic that requires the benefit to run through the capability
reduction will report nothing where a path has become a loop. So a merge is
admissible for a question exactly when no edge of that question’s witness
crosses or is internalized by the merge boundary. The entity list was revised
twice because two bundles failed this test, and the result says which questions
each bundle was still adequate for.

The third result concerns the difficulty the discussion identified as decisive:
that one cannot tell, from outside, whether a particular use of a particular
work in training was an appropriation at all. This is treated here as structural
rather than as a temporary deficiency of transparency. Where corpus membership,
the counterfactual capability, and the benefit path are each unavailable, the
set of topologies consistent with what is observed contains both positive and
negative cases, so a three-valued audit returns its unresolved value; and every
observation that would resolve it is held by the party whose conduct is in
question. Two things follow. Disclosure is a condition on whether the framework
applies at all, not a preference among remedies. And reading an unresolved audit
as a negative finding is an error whose beneficiary can be named in advance.

A fourth result concerns what changes first. If a system lets one party traverse
a possibility space faster than another, and positions in that space are
exclusive once occupied, then two parties with identical holdings have different
reachable sets at every finite horizon. Equalizing holdings does not equalize
reachability. Rate is therefore a structural variable in its own right rather
than a description of how quickly a distributional fact takes effect.

Alongside these the paper argues three things it does not formalize. The first
concerns the paradox the discussion raised at the outset: if a system is built
from the accumulated production of everyone, then everyone contributed, and a
category that applies to everyone distinguishes nobody. The general form of this
paradox is addressed in a companion project by separating causal contribution
from standing. The paper adds an observation specific to the present case. A
training corpus is not the collective inheritance of humanity; it is a finite
assembled object whose boundary was drawn by identifiable decisions. The
appearance of unlimited diffusion is an artifact of describing an artifact as if
it were a heritage, and correcting the description converts a metaphysical
puzzle into an institutional question about who assembled what, from where, and
on what terms. The second is that the standing of the system itself is open, and
that a framework which presupposes either answer is defective; the paper is
constructed so that both are expressible and states which of its results would
require revision under each. The third is that the appropriate register for what
the framework can currently deliver is diagnosis-conditional-on-disclosure, not
diagnosis.

The paper classifies no system, provider, adopting firm, or episode. It
establishes no prevalence. It does not claim that AI-mediated relations are
exploitative, nor that they are not. Section 7 states, item by
item, the evidence any such claim would require, and records that this paper has
none of it. The portfolio entry from which the project derives anticipated
exactly this deficit, and the appropriate response is to say so rather than to
supply illustrations that would function as evidence without being evidence.

Section 2 reconstructs the question and shows where it goes
wrong. Section 3 reviews the traditions and records what each
supplies and forbids. Section 4 argues six theses with their
strongest objections and replies. Section 5 declares the objects
and proves four propositions. Section 6 replaces the pairwise
table with a role-assignment table and exhibits five countermodels.
Section 7 states the evidence gate. Section 8 records
open questions, and Section 9 the limits and defeat conditions.

The philosophical and formal parts of this paper are formally independent
and justificatorily dependent. No proposition cites a thesis, and every
proposition survives the rejection of all six theses. What does not survive is
the claim that the propositions matter.

2. The Question and Why It Is Malformed

This section performs the problem analysis. Its objective is to reconstruct the
question as it was posed, to record the two revisions it underwent, and to
locate the defect that no further revision of the same kind would repair.

2.1 The Initiating Comparison

The comparison that opened the inquiry sets a classical configuration against a
current one. Classically, a worker’s labor produces value, the worker receives
compensation, and the analysis attends to the difference. Currently, a party
spends a small sum on machine-generated work and realizes a large return. The
classical difference is not small in the second case; it is undefined, because
the party whose compensation would enter the comparison is not present in the
transaction at all.

Two observations follow, and only the second is this paper’s business. The first
is that a value-theoretic accounting will not reach the configuration, which is
a familiar result: exploitation theory has been reconstructed on property
relations and on comparison against a feasible alternative arrangement without
requiring a value substance (Cohen, 1979; Roemer, 1982). The
second is that the parties whose position the comparison is about — those whose
prior production entered the system, and those whose work the system now
substitutes for — are not parties to any transaction under examination. They
appear nowhere in the accounting, and no refinement of the accounting will make
them appear.

That is why the inquiry widened to a cast of actors. The move is correct. What
follows is an argument that the widened question was then posed in a form its
own evidence cannot settle.

2.2 Four Entities, Then Five, Then Five Others

The first cast was capital, labor, the system, and users, with the question
stated as which of the twelve directed relations among them are exploitative.

The cast was then revised twice, and the revisions are the most instructive part
of the sequence. First, the firm that builds, trains, owns, and governs the
system was separated from the capital that finances it, on the ground that
financing and governing are different functions that had been bundled. Second,
the cast was restated as the system, users, laborers, firms that adopt AI, and
provider companies, separating the party that deploys a system in a workplace
from the party that supplies it.

Each revision was made for the same reason: an entity in the previous list was
found to be doing two things whose analysis differs. The paper takes the reason
seriously and asks what governs such decisions.
Section 5.3 answers: a merge is admissible for a question when
no edge of that question’s witness crosses or is internalized by the merge
boundary, and both revised bundles fail that test for the questions at issue.

But the revisions do not repair the underlying difficulty, because they leave the
question’s form untouched. With five entities the table has twenty cells instead
of twelve; a sixth entity would give thirty. The number of cells is not the
problem.

2.3 Four Roles, Not Two Positions

The diagnostic this programme uses requires four things of an alleged relation
between a benefiting party and an affected party: that the affected party has
standing; that a declared alternative arrangement is admissible; that the
affected party’s effective generative capability is lower under the focal
arrangement than under the alternative; and that the benefiting party’s return is
higher through that reduction rather than merely alongside it.

Satisfying these requires locating four roles. Someone governs the arrangement.
Something operates and is the proximate cause of the change. Someone’s capability
falls. Someone’s return rises through the fall. In the classical employment case
these collapse: the employer governs, the process operates under its direction,
the employee’s position is affected, and the employer benefits. Collapse is why
the pairwise form was ever adequate.

In the configurations at issue the roles come apart routinely. A system operates;
a provider governs what it does; an adopting firm chooses to deploy it; a worker’s
capability is affected; and the return may rise at the provider, at the adopter,
or at neither. A relation described as one between the system and the worker
names the actor and the affected party and is silent on the two roles that
determine whether the diagnosis holds at all. A relation described as one between
capital and the worker names the beneficiary and the affected party and is silent
on the mechanism.

This is the defect. A directed pair is a projection of a four-role assignment onto
two coordinates, and different assignments project onto the same pair.
Section 5.2 shows the projection loses exactly the
information the diagnosis needs.

2.4 Requirements on an Adequate Account

Four requirements follow.

First, the unit of analysis must be a role assignment rather than a pair, and any
tabulation must be indexed accordingly.

Second, entity individuation must be treated as a modeling decision with stated
consequences, not as a description of the world that happens to be convenient.

Third, the account must state what is unresolved and why, and must be constructed
so that an unresolved audit cannot be read as a negative finding.

Fourth, the account must not presuppose an answer to the standing of the system,
since that question is open and the framework is meant to survive either
resolution.

3. Traditions Bearing on AI-Mediated Relations

This section reviews the relevant literatures. Its objective is to record what
each supplies, what constraint each imposes, and where each stops. No source in
this corpus states the underdetermination result, the aggregation criterion, or
the opacity result.

3.1 The Political Economy of AI Production

Crawford analyzes artificial intelligence as an extractive industry, tracing
mineral extraction, low-wage labor, and data appropriation through the production
of systems, and argues that the resulting infrastructure concentrates power
(Crawford, 2021). This supplies the empirical picture the paper’s
schematic topology abstracts from, and it imposes a caution the paper accepts:
the extraction Crawford documents is heterogeneous, and treating it as one
relation between two parties is precisely the error under examination here.

Srnicek analyzes platforms as intermediaries whose position derives from
controlling the terms on which others interact (Srnicek, 2017);
Birch analyzes value appropriated through control rights rather than through
production (Birch, 2020); Khan argues that structural position
can produce harm that price-based tests do not register (Khan, 2017).
Together these supply the form of a provider’s position. None distinguishes the
provider from the adopter, which is the distinction the second revision of the
entity list introduced.

Terranova analyzes free labor in digital economies (Terranova, 2000), and
Couldry and Mejias analyze data appropriation as a colonial structure
(Couldry & Mejias, 2019). These describe the pattern of distributed contribution
meeting narrow capture that the training relation instantiates. Neither addresses
the diagnostic question of whose capability falls and whose return rises through
the fall.

3.2 Automation, Tasks, and Labor Demand

Acemoglu and Restrepo analyze automation through a task framework in which
technology both displaces labor from tasks and reinstates it by creating new
ones, so that the net effect on labor demand depends on which effect dominates
(Acemoglu & Restrepo, 2019).

This is the most disciplined available treatment of the substitution relation
and it imposes the paper’s sharpest empirical constraint. Displacement is not
assumed by the framework; it is one of two effects whose balance is an empirical
question. A theory that treats every deployment as a capability reduction has
assumed what this literature investigates, and the paper’s countermodels are
written to keep the assumption out.

The task framework and the capability framework used here are not the same
object. A task is a unit of production; effective generative capability is what a
participant is positioned to develop. A deployment can reinstate tasks while
contracting a participant’s developmental repertoire, and can displace tasks
while leaving the repertoire intact. Neither result transfers to the other.

3.3 Opacity, Documentation, and Disclosure

Burrell distinguishes three forms of opacity in machine learning — intentional
concealment, technical illiteracy, and opacity arising from the mismatch between
the scale of learned models and human interpretive capacity — and argues that
they call for different responses (Burrell, 2016).

The distinction matters here. The opacity the paper’s third result concerns is
the first kind, and it is the kind that documentation practices address rather
than dissolve. Gebru and colleagues propose datasheets recording a dataset’s
motivation, composition, collection process, and recommended uses
(Gebru et al., 2021); Mitchell and colleagues propose model cards recording
a model’s intended use, evaluation, and limitations
(Mitchell et al., 2019). Both are proposals for making determinable
something that is otherwise held privately, and both are cited here for that
structural property rather than for any claim about adoption.

The European Union’s regulatory framework for AI supplies a verified record of
staged obligations, with prohibited practices applying from February 2025,
governance and general-purpose model obligations from August 2025, and
obligations for systems used in employment and other listed high-risk areas from
December 2027 (Commission, 2026). The paper cites the record for its dates and
draws no conclusion about the adequacy of any obligation.

Fricker’s analysis of epistemic injustice identifies wrongs done to someone in
their capacity as a knower, including the case where the interpretive resources
needed to articulate a harm are unavailable to the person harmed
(Fricker, 2007). This is the closest normative treatment of the
position the third result describes, in which a party cannot determine whether
something happened to it because the determination depends on records held by
another party.

3.4 Exploitation, Domination, Justice, and Standing

Roemer’s comparison against a feasible alternative arrangement supplies the
diagnostic form (Roemer, 1982); Cohen establishes that the charge
does not require a labor theory of value (Cohen, 1979). Marx’s account of
the general intellect describes social knowledge becoming a direct force of
production (Marx, 1973), which is the classical antecedent for a
system that condenses accumulated collective production, and the surrounding
value theory is declined here as it is in the companion projects
(Marx, 1992). Romer’s treatment of ideas as nonrival and partially
excludable supplies the property that makes the training relation non-depleting
(Romer, 1990).

Wertheimer establishes that mutually beneficial and consented transactions can be
exploitative (Wertheimer, 1996); Sample locates the wrong in a
failure to respect the value of the other party (Sample, 2003);
Vrousalis locates it in self-enriching instrumentalization of vulnerability
(Vrousalis, 2013). Zwolinski’s constraint against inferring a
transaction’s status from background structure applies with particular force to a
setting in which almost everyone stands in some remote causal relation to the
training corpus (Zwolinski, 2012). Wollner supplies the category for
a wrong with no identifiable agent (Wollner, 2019), and Young the
form a forward-looking response takes (Young, 2006).

The two normative frameworks the source discussion invoked are recorded with
their standard content. Rawls’s difference principle permits inequalities only if
they work to the advantage of the least advantaged (Rawls, 1999), and
Pettit’s republicanism locates unfreedom in subjection to another’s arbitrary
power rather than in interference as such (Pettit, 1997). Both
bear on the configuration and neither is applied to any case here.

Sen’s distinction between achieved functionings and the capability set supplies
the affected object (Sen, 1993), with Robeyns recording the
specification choices any operationalization requires
(Robeyns, 2005). Anderson’s relational criterion bears on why a
question about who may determine another’s developmental conditions is not
answered by a statement of holdings (Anderson, 1999).

On the standing of systems, Long and colleagues argue that the welfare of AI
systems is a question that can no longer responsibly be dismissed, while
declining to assert that any current system is a moral patient
(Long et al., 2024). The paper adopts exactly that posture and no more.

3.5 Causal Structure and Partial Identification

Pearl’s framework distinguishes causal structure from statistical association and
supplies the vocabulary of mediation in which an intermediate node transmits an
effect it does not originate (Pearl, 2009). The acting-node and
beneficiary-node distinction on which the first result rests is a special case of
that vocabulary, applied to a diagnostic predicate rather than to an outcome
variable.

Manski’s framework for partial identification supplies the reporting discipline
the third result requires (Manski, 1990): where the data do not
determine a quantity, report the set that is consistent with the data rather than
a point.

| @>p3.1cmYY@

Tradition	Supplies	Constraint imposed
Political economy of AI production	Extraction across minerals, labor, and data; concentration of infrastructure	The extraction is heterogeneous; treating it as one relation is the error under examination
Platform position and rent	The form of a provider’s advantage	No distinction between provider and adopter
Appropriated contribution	Distributed contribution meeting narrow capture	Silent on whose capability falls and whose return rises through it
Automation and tasks	Displacement and reinstatement as two effects whose balance is empirical	A theory assuming every deployment reduces capability has assumed what this literature investigates
Opacity	Three kinds, requiring different responses	Only concealment is addressed by disclosure; scale opacity is not
Documentation practices	Making corpus and model properties determinable	Cited for structure, not for any claim about adoption
Regulatory record	Verified staged dates of obligation	No conclusion about adequacy of any obligation
Epistemic injustice	Wrongs done in the capacity of a knower	Does not establish that opacity here is such a wrong
Exploitation theory	Comparison against an arrangement; consent is not exculpatory; a category for agentless wrongs	Remote causal contribution must not imply exploitation, or the category empties
Justice frameworks	Difference principle; non-domination	Neither is applied to any case here
Capability	Opportunity rather than achievement as the affected object	Specification choices are unavoidable
AI standing	The question is live and unresolved	No current system is asserted to be a moral patient
Causal structure	Mediation: a node transmits an effect it does not originate	Structure is assumed, not identified, in the constructions here
Partial identification	Report the consistent set, not a point	A set containing both answers supports neither

Table. Traditions bearing on AI-mediated generative relations, what each
supplies, and the constraint each imposes on the paper’s claims.

4. Six Theses on Roles, Bundles, Corpora, Opacity, Standing, and Speed

This section argues the paper’s six theses. Its objective is to give each an
argument, the strongest objection known to the author, and a reply. The section
makes no use of the formal results and is not used by them.

4.1 The Unit of Diagnosis Is a Role Assignment

The unit over which a diagnosis of generativity exploitation is defined is an
assignment of four roles — controller, actor, affected participant, and
benefiting party — and not an ordered pair of entities. A tabulation indexed by
ordered pairs cannot represent the diagnosis.

Argument. The diagnostic conditions quantify over four positions, and no
two of them are guaranteed to coincide in the configurations at issue. A pair
therefore fixes two coordinates and leaves two free, and the free coordinates
carry the conditions that decide the case: whether the actor was governed, and
whether the benefit ran through the reduction rather than alongside it.

The point is not that pairwise descriptions are imprecise. It is that they are
projections, and Proposition ? shows the projection
is not injective on the quantities such tables are populated from. Two topologies
can agree on which node acted, whose capability fell, and whose return rose, and
disagree on whether anyone exploited anyone.

Strongest objection. This is a familiar point about mediation dressed up
as a discovery. Everyone knows that a tool is not the party responsible for what
is done with it, and the observation that one should trace causal chains rather
than read off adjacency is a commonplace of causal analysis
(Pearl, 2009).

Reply. The commonplace is granted and the thesis is stronger. The claim
is not that one should trace chains but that the pairwise data do not determine
the chain, so tracing is not an improvement available within the pairwise
framing. A tabulation of directed relations among five entities has twenty cells
and the role assignments it is meant to summarize number in the hundreds; the
table is not a coarse view of the answer but an index of the wrong space. This is
a claim about representational adequacy, and it is proved rather than asserted.

4.2 Bundling Is the Dominant Failure Mode

Individuating entities is a modeling decision with diagnostic consequences, and
an aggregation is admissible for a question only if the diagnosis of that
question is invariant under it. Most errors in this area are bundling errors,
and they are errors in both directions: a bundle can manufacture an attribution
and can destroy one.

Argument. The evidence is the source sequence itself. The entity list
was revised twice, each time because a single name was covering two functions
with different diagnostics — financing against governing, then supplying
against deploying. The revisions were made on intuition, and the intuition was
right, and Proposition ? states the criterion it was
tracking.

The criterion’s two directions are both consequential. Merging a controller with
the system it governs yields an aggregate that appears to act on third parties,
an attribution that belonged to neither constituent: the controller did not act,
and the system did not decide. Merging a provider with an adopter internalizes a
benefit path, so a diagnostic requiring benefit to run through a
capability reduction finds a loop where it needs a path, and reports nothing.
The second direction is the more dangerous, because it produces false negatives
that look like clean results.

Strongest objection. At a sufficient level of aggregation the bundles
are correct. The same owners stand behind the financing, the provider, and often
the adopter, and insisting on their separation is a formalism that obscures a
unity everyone can see.

Reply. The objection may be true and does not license the merge. What it
asserts is a claim about ownership, which is a claim about control; the criterion
grants that the merge is admissible for control questions and denies that it is
admissible for benefit-path questions, and these are different questions with
different witnesses. The demand is not that entities be kept apart but that a
merge be justified for the question at hand, which is a weaker and more
defensible requirement than either separation or unity as a general policy.

4.3 A Corpus Is Not an Inheritance

The appearance that everyone contributed to a system, and hence that the category
of exploitation applies to everyone or to no one, arises from describing an
assembled artifact as if it were a collective inheritance. A training corpus is
finite, assembled, and bounded by decisions someone made, and this is a different
kind of object from the accumulated production of a civilization.

Argument. The paradox as posed runs: a system condenses the historical
production of language, knowledge, and culture; almost everyone stands in some
remote causal relation to that production; if everyone contributed then everyone
is exploited, and a category holding of everyone distinguishes nothing.

The general form of this argument is answered elsewhere in the programme by
separating causal contribution, which is historical and admits no threshold, from
standing, which is a normative status tracking who stands to lose. That answer
holds here and is not repeated.

What is added is specific to the object. The premise slides between two things. A
civilization’s accumulated production is unbounded, unenumerable, and had no
assembler. A training corpus is a set of items that were selected, obtained,
processed, and used, by parties who made decisions about inclusion, and whose
extent is in principle enumerable because it was in fact enumerated by whoever
assembled it. The universality in the premise belongs to the first object and the
appropriation in the conclusion concerns the second, and the argument works only
by treating them as one.

Correcting this does not show that anyone was exploited. It changes the kind of
question. Instead of asking whether a metaphysically diffuse contribution can
ground a claim, one asks who assembled which items, from where, on what terms,
and with what effect on the parties whose items were included. Those questions
have answers, and Section 4.4 concerns why the answers are
not available.

Strongest objection. The corpus is enumerable only in principle. In
practice its composition is undisclosed, its scale defeats inspection, and much
of it derives from material whose own provenance is untraceable. A boundary that
cannot be examined is not usefully different from no boundary.

Reply. Undisclosed is not unbounded, and the difference is exactly the
paper’s point. An unbounded contribution grounds no claim in principle; an
undisclosed bounded one grounds a claim that cannot currently be evaluated. These
call for different responses — conceptual revision in the first case,
disclosure in the second — and collapsing them recommends the first where the
second is needed.

4.4 Opacity Is Structural and Interest-Aligned

The facts that would resolve a diagnosis in this setting are held by the parties
whose conduct is at issue. This is a structural feature of the configuration
rather than a passing deficiency of transparency, and it entails that an
unresolved audit is the generic outcome and must not be read as a negative
finding.

Argument. Three facts are required: whether a participant’s material
entered the corpus, what the participant’s capability would have been under an
admissible alternative arrangement, and whether the benefiting party’s return
rose through the reduction rather than alongside it. The first is recorded, if
anywhere, in the assembler’s records. The third is recorded, if anywhere, in the
benefiting party’s accounts. The second is counterfactual and is unavailable to
everyone, which is a separate difficulty addressed in a companion project.

So two of the three resolving facts are private to parties with an interest in
the outcome, and Proposition ? shows that without them the
consistent set contains both positive and negative cases.

Two consequences follow. Disclosure requirements are not one policy option among
others but a condition on whether the framework can be applied at all; documented
corpus and model properties are proposals of exactly this kind
(Gebru et al., 2021; Mitchell et al., 2019), and a regulatory record of
staged obligations is a record of when certain determinations become possible
(Commission, 2026). And the specific error of treating an unresolved audit as a
negative finding has an identifiable beneficiary, which is a reason to name the
error rather than to rely on inferential good manners.

Fricker’s analysis describes the position of a party that cannot articulate what
was done to it because the resources for the articulation are not available to it
(Fricker, 2007). The paper notes the resemblance and does not claim
that the configuration constitutes epistemic injustice in her sense, which would
require an argument about the mechanism of the exclusion.

Strongest objection. Not all opacity is concealment. Burrell distinguishes
opacity arising from the scale and character of learned models from opacity that
anyone chose (Burrell, 2016), and the second sort cannot be disclosed
away because nobody holds the missing fact. Treating the whole difficulty as
interest-aligned overstates it and invites remedies that will not work.

Reply. Accepted, and the thesis should be read as restricted. Corpus
membership and benefit path are held facts; how a system’s behavior arises from
its training is not, and no disclosure produces it. The restriction matters for
what follows: the first two are the facts the diagnosis needs, and the third is
not among its conditions. A framework that required an account of why a system
behaves as it does would be defeated by scale opacity; this one is not, and that
is a design consequence rather than an accident.

4.5 Standing Must Remain an Open Coordinate

Whether the system itself has standing is open, and a framework that presupposes
either answer is defective. The framework should be constructed so that both are
expressible, and should state which of its claims change under each.

Argument. The question is live. There is serious argument that the
welfare of AI systems can no longer be responsibly dismissed, offered without any
assertion that a current system is a moral patient (Long et al., 2024).
Constructing a framework that assumes a negative answer builds a contested
premise into an apparatus meant to outlast the controversy; assuming a positive
answer does the same in the other direction.

Neutrality here is cheap, and the paper should say so rather than claim a virtue.
Every result in Section 5 is proved for a topology in which the
system occupies the actor role and no standing role. Under a positive answer to
the standing question, three things change and nothing else does. Relations
directed at the system become diagnosable rather than undefined, so the role
table acquires occupied cells it currently leaves empty. The aggregation criterion
acquires a further constraint, since merging a system with its provider would
then merge a standing-holder with a non-holder. And the system becomes a possible
affected participant in the opacity result, with the peculiar feature that the
records bearing on its position are held by its provider.

Strongest objection. Neutrality is evasion. The question of whether these
systems have interests is substantive and consequential, and a framework that
declines to answer it while claiming to analyze relations involving them has
declined the hardest part of its subject.

Reply. The objection is right that the question is substantive and wrong
that this paper is the place to settle it. What is offered is not silence but a
conditional: the results hold under the negative answer, and the enumerated three
changes are what a positive answer would require. A reader who holds a view can
therefore use the framework; a framework that had chosen would serve only readers
who chose the same way.

4.6 Rate Is a Structural Variable

Where a party can traverse a possibility space faster than another and positions
in that space are exclusive once occupied, a difference in rate forecloses
without any difference in holdings. Rate is therefore a structural variable and
not a description of how quickly a distributional fact takes effect.

Argument. A diagnosis compares an affected participant’s position under
the focal arrangement against an admissible alternative. If the alternative
requires time the participant does not have because another party will have
occupied the relevant positions first, then the alternative is not available even
though nothing in the participant’s holdings prevents it.
Proposition ? states this: identical holdings and different rates
give different reachable sets at every finite horizon, and the difference does
not close.

The significance for the present subject is that rate is the dimension along
which a general-purpose system most obviously operates. Whatever else is
contested about such systems, that they let some parties traverse certain spaces
faster than others is not.

Strongest objection. This describes competition. Being first is how
priority has always worked, and a framework that treats speed as a structural
wrong will classify every race as a foreclosure.

Reply. Accepted as description and denied as inference, which is the
reply the programme gives whenever a structural feature is mistaken for a verdict
(Zwolinski, 2012). The thesis locates a variable and asserts nothing
about permissibility. Its use is negative and specific: it defeats the inference
from “holdings were equalized” to “the alternative was available,” which is an
inference remedial proposals make routinely.

4.7 Dependence Structure Among the Theses

Rejecting Thesis ? restores the pairwise table and with it the
question the paper says cannot be answered. Rejecting Thesis ?
makes entity individuation a matter of convenience and removes the account of why
the source’s revisions were corrections. Rejecting Thesis ? returns
the universal-contribution paradox in its AI form and with it the conclusion that
the category applies to everyone or nobody. Rejecting Thesis ?
makes the evidentiary difficulty contingent and removes the argument that
disclosure is a precondition rather than a preference. Rejecting
Thesis ? requires the framework to take a position on a question
it has no means to settle. Rejecting Thesis ? restores the inference
from equalized holdings to available alternatives.

Theses ? and ? are load-bearing: the first is the
paper’s reason for existing and the second is its reason for reaching no
conclusions. Thesis ? is the most contestable, since it rests on a
distinction between an assembled artifact and an inheritance that a determined
opponent can attack from either side.

5. Formal Results

This section declares the objects and proves four propositions. Its objective is
to establish that pairwise data underdetermine the diagnosis, that aggregation
has a stateable admissibility criterion, that opacity yields a generically
unresolved audit whose resolving observations are privately held, and that a rate
difference forecloses without a holdings difference. The section cites no thesis.

5.1 Declared Objects

A typed topology is a tuple $T=(\Nn,K,A,\Gamma,\rho,b)$ where $\Nn$ is a finite
set of nodes; $K\subseteq\Nn\times\Nn$ is a set of control edges, with $(x,y)
K$ meaning that $x$ governs $y$’s policy; $A$ is a set of
action edges, with $(x,y)\in A$ meaning that $x$’s operation is the proximate
cause of a change in $y$’s conditions; $\Gamma:\Nn\to\R$ assigns each node its
effective generative capability under the focal arrangement; $\rho:\Nn\to\R$
assigns each node its return; and $b$ is a declared alternative arrangement
inducing $\Gamma^{b}$ and $\rho^{b}$.

For nodes $x,y$ with $x\neq y$, write $E_{b}(x,y)=1$ when all of the following
hold: $y$ has standing; $b$ is admissible; $\Gamma(y)<\Gamma^{b}(y)$;
$\rho(x)>\rho^{b}(x)$; and there is a benefit-through witness, that is, a
directed path $w$ from the arrangement’s point of intervention to $x$’s return
which passes through the reduction at $y$, such that removing the reduction
removes the increase. Write $E_{b}(x,y)=0$ when some conjunct is determinately
false, and $E_{b}(x,y)=,?$ when the available observations determine no value.

The predicate is the companion framework’s, restated to the extent this paper
uses it. Nothing here re-derives its motivation, and nothing here weakens it.

A role assignment is a tuple $(c,a,y,x)\in\Nn^{4}$ with $y\neq x$, naming a
controller, an actor, an affected participant, and a benefiting party. A pair
$(x,y)$ is the projection of $(c,a,y,x)$ onto its last two coordinates.

The marginal record of $T$ is the triple
$M(T)=\bigl(\mathbf{1}_{A},\ \Gamma-\Gamma^{b},\ \rho-\rho^{b}\bigr)$: which node
acted on which, how each node’s capability changed, and how each node’s return
changed.

The marginal record is what a pairwise tabulation is typically populated from: it
says who did something, whose position worsened, and whose improved.

5.2 Pairwise Data Underdetermine the Diagnosis

There exist typed topologies $T_{1},T_{2}$ on the same node set with
$M(T_{1})=M(T_{2})$ and a pair $(x,y)$ such that $E_{b}(x,y)=1$ in $T_{1}$ and
$E_{b}(x,y)=0$ in $T_{2}$.

Let $\Nn={c,s,l}$. In both topologies let $A={(s,l)}$, let
$\Gamma(l)<\Gamma^{b}(l)$ with all other capabilities unchanged, and let
$\rho(c)>\rho^{b}(c)$ with all other returns unchanged. Then
$M(T_{1})=M(T_{2})$ by construction.

In $T_{1}$ let $K={(c,s)}$ and let the point of intervention be $c$’s policy
for $s$, with a witness path $c\to s\to l\to\rho(c)$: the reduction at $l$ is
what raises $c$’s return, and removing it removes the increase. Then
$E_{b}(c,l)=1$.

In $T_{2}$ let $K=\varnothing$, so $s$ operates autonomously, and let $c$’s
return rise through a channel disjoint from $l$ — for instance an unrelated
holding that appreciates over the same period. Then no witness path through the
reduction exists, the last conjunct of Definition ? fails
determinately, and $E_{b}(c,l)=0$.

On $n$ nodes there are $n(n-1)$ ordered pairs and $n^{3}(n-1)$ role assignments
satisfying $y\neq x$, of which $n(n-1)(n-2)(n-3)$ have four distinct occupants.
For $n=5$ these are $20$, $500$, and $120$ respectively.

The counting is trivial and its point is not. A tabulation of twenty directed
relations is not an approximation to the answer; the objects it indexes are not
the objects the diagnosis is defined on, and Proposition
? shows the map between them is not injective on the
data that would populate the table. Adding entities makes this worse rather than
better, since the assignment count grows as $n^{4}$ and the pair count as
$n^{2}$.

Two things the proposition does not show should be stated. It does not show that
the diagnosis is never determinable; it shows that marginal data do not determine
it, and control edges and witness paths are exactly the further data required.
And it does not show that $T_{2}$ is innocent in every respect — only that the
particular predicate is determinately false there, which is what a diagnostic
framework is supposed to be able to say.

5.3 When an Aggregation Is Admissible

For $u,v\in\Nn$ with $u\neq v$, the merge $T/{u,v}$ replaces them by a single
node $\bar u$, redirects every edge with an endpoint in ${u,v}$ to $\bar u$,
deletes edges with both endpoints in ${u,v}$, and sets
$\rho(\bar u)=\rho(u)+\rho(v)$.

Fix a pair $(x,y)$ with a witness path $w$ in $T$, and let $u,v\notin{y}$.

• If no edge of $w$ has both endpoints in ${u,v}$, and $x\notin{u,v}$ or ${u,v}\subseteq{x}\cup(\Nn\setminus w)$, then $E_{b}(x,y)$ is unchanged in $T/{u,v}$.
• If some edge of $w$ has both endpoints in ${u,v}$, then $w$ has no image in $T/{u,v}$, and if no alternative witness exists then $E_b( u,y)1$.
• If $(u,v)\in K$ and $(v,z)\in A$ for some $z$, then $(\bar u,z)\in A$ in $T/{u,v}$, although $(u,z)\notin A$ in $T$.

(1) Under the stated condition every edge of $w$ has an image in $T/{u,v}$ and
the images form a directed path with the same endpoints, so the witness survives;
the remaining conjuncts of Definition ? concern $y$ and the
returns, and neither is altered by a merge disjoint from $y$ whose return sum
preserves the sign of the change at the benefiting party.
(2) By Definition ? edges internal to ${u,v}$ are deleted, so $w$
is not a path in the merged topology; if no other witness exists the final
conjunct fails and the predicate is not $1$.
(3) The merge redirects $(v,z)$ to $(\bar u,z)$, which is an action edge incident
to $\bar u$; since $(u,z)\notin A$, the aggregate carries an action attribution
that $u$ did not.

An aggregation is admissible for a question exactly when no edge of that
question’s witness is internalized by the merge and no role of the question’s
assignment is split across the merged nodes. In particular, a merge of a
financing party with a governing party is admissible for questions about control
and inadmissible for questions whose witness runs between them; and a merge of a
supplying party with a deploying party is inadmissible for any question whose
witness runs from one to the other.

Clause (2) is the one that produces silent errors. A merged provider and adopter
present as a single firm that both trained a system and deployed it, and a
benefit path that ran from deployment back to supply becomes internal and
disappears. The resulting diagnosis is not merely coarse; it is negative, and it
looks like a finding.

Clause (3) produces the opposite error and is the more visible of the two. Where
a governing party and the system it governs are treated as one, the aggregate
appears to act on those the system acts on. This is how a relation between a
system and a worker comes to be asserted, and neither constituent supports it:
the governing party did not act and the system did not decide.

5.4 Opacity Yields an Unresolved Audit With Privately Held Resolvers

An observation $\Oo$ is a set of facts about a topology. The consistent set
$\Tt(\Oo)$ is the set of typed topologies compatible with $\Oo$.

$\Oo$ determines neither (i) whether a given participant’s material entered the
corpus from which the system was produced, nor (ii) the counterfactual capability
$\Gamma^{b}$, nor (iii) the existence of a benefit-through witness.

Under Assumption ?, for the pair $(x,y)$ at issue there exist
$T_{+},T_{-}\in\Tt(\Oo)$ with $E_{b}(x,y)=1$ in $T_{+}$ and $E_{b}(x,y)=0$ in
$T_{-}$. Hence the audit returns $?$. Moreover any $\Oo’\supseteq\Oo$ with
$E_{b}(x,y)\neq,?$ for all $T\in\Tt(\Oo’)$ must determine at least one of
(i)–(iii).

The construction of Proposition ? supplies $T_{+}$
and $T_{-}$ agreeing on all marginals; by Assumption ? the
observation distinguishes neither their control structure nor their witness
structure, so both lie in $\Tt(\Oo)$ and the audit has no determinate value. For
the second claim, suppose $\Oo’$ determines none of (i)–(iii). Then the same two
topologies remain consistent with $\Oo’$, since each of (i)–(iii) is the only
respect in which the relevant conjuncts differ, and the audit remains $?$.

If facts (i) and (iii) are recorded only by the party assembling the corpus and
the party whose return is at issue, then every $\Oo’$ resolving the audit
includes information held by a party whose conduct the audit concerns.

Three consequences are worth separating. First, an unresolved audit is the
expected outcome rather than an anomaly, so a body of unresolved audits is not
evidence that nothing occurred. Second, the inference from $?$ to $0$ is invalid
on the framework’s own three-valued semantics, and the party favored by that
invalid inference is identifiable in advance, which is a reason to state the point
explicitly rather than to treat it as a technicality. Third, a disclosure
requirement changes the class of decidable propositions, which places it in a
different category from remedies that alter outcomes; documentation practices for
datasets and models are proposals of exactly this shape
(Gebru et al., 2021; Mitchell et al., 2019).

Fact (ii) is not repaired by any disclosure, since no party holds a
counterfactual. Its treatment is the subject of a companion project, which shows
that such a repertoire admits bounds rather than a value, and the reporting
discipline for a quantity with a consistent set rather than a point is standard
(Manski, 1990).

5.5 Rate Forecloses Without a Holdings Difference

Let $V$ be a space of developmental positions with a cost function
$\mathrm{cost}:V\to\R_{>0}$. A participant with rate $r>0$ has reachable set at
horizon $h$

$$\Rr_{r}(h)={\tau\in V:\ \mathrm{cost}(\tau)\leq r,h}.$$

A position occupied at some time is unavailable to any other participant
thereafter.

Let two participants have identical holdings and rates $r_{A}<r_{B}$. Then for
every $h>0$, $\Rr_{r_{A}}(h)\subsetneq\Rr_{r_{B}}(h)$ whenever
$\mathrm{cost}$ takes a value in $(r_{A}h,,r_{B}h]$. Under
Assumption ?, the positions in
$\Rr_{r_{B}}(h)\setminus\Rr_{r_{A}}(h)$ are unavailable to $A$ at every later
time, and the excluded set is non-decreasing in $h$ and in $r_{B}/r_{A}$.

Strict inclusion is immediate from the definition and the stated value condition.
Under Assumption ?, $B$ occupies those positions at times
$A$ cannot match, and occupation is permanent, so they are excluded from $A$’s
attainable set thereafter. Monotonicity in $h$ and in the ratio follows because
the excluded set is ${\tau:r_{A}h<\mathrm{cost}(\tau)\leq r_{B}h}$, whose
defining interval widens in both parameters.

An alternative arrangement that equalizes holdings but not rates does not
equalize reachable sets, so its admissibility as a comparison does not follow
from its distributive properties.

The result is elementary and its assumptions are strong. Exclusive occupation
holds for priority, for certain standards positions, and for some market
positions, and fails wherever a development can be undertaken independently by
several parties. The proposition should be read as identifying a mechanism
available in some regions of a possibility space rather than as a claim about
possibility spaces in general.

6. The Role Table and Countermodels

This section replaces the pairwise tabulation with one indexed as the diagnosis
requires, and exhibits five countermodels. Its objective is to show what a
correctly indexed summary looks like and to mark the boundaries of the analysis
by displaying configurations that fail it.

6.1 A Role-Indexed Summary

Table 2 lists configuration schemas by role assignment rather than
by pair. Each row names who governs, what operates, whose capability is at issue,
and where a return would have to rise for the diagnosis to apply; the final
column records what would have to be shown. No row asserts that any configuration
obtains, and the table is a specification of questions rather than a summary of
findings.

| @>p1.6cm>p1.5cm>p1.9cm>p1.6cmY@

Controls	Acts	Affected	Benefits	What must be shown
Provider	System	Contributor to corpus	Provider	Corpus membership; a capability reduction against an admissible alternative; a witness from the reduction to the provider’s return
Provider	System	Adopting firm	Provider	Dependence on a governed condition; a reduction relative to an alternative provision; a witness through the dependence
Adopter	System	Worker	Adopter	Deployment as the intervention; task displacement distinguished from repertoire contraction; a witness from the contraction to the adopter’s return
Adopter	System	Worker	Provider	The same, with a witness crossing from adopter to provider; inadmissible under any merge of the two
Provider	System	User	Provider	Interaction traces as the acquired material; a reduction in the user’s own developmental position; a witness distinct from ordinary service revenue
Capital	Provider	Worker	Capital	Control of the provider’s policy; the chain from that control to the reduction; admissibility of merging capital with provider tested separately
—	System	Any	Nobody	Autonomous operation with no return increase anywhere: a candidate for harm without exploitation
Provider	System	System	Provider	Undefined unless the system has standing; the cell is empty under the negative answer and occupied under the positive one

Table. Configuration schemas indexed by role assignment. Each row states what
would have to be established, not what has been.

Two features of the table are the point of it. The third and fourth rows differ
only in where the return rises, and they are different diagnostic questions with
different witnesses; a pairwise table has one cell for both. The last row is empty
or occupied depending on the standing question, which is how
Thesis ?’s neutrality is implemented rather than asserted.

6.2 Countermodels

A system operates, a participant’s capability falls, and no party’s return rises
through the fall. The final conjunct of Definition ? fails
determinately.

The configuration is harm and not exploitation, and it is the case a pairwise
reading most reliably misdescribes, since the action edge and the capability
change are both present and both visible.

A participant’s capability falls for reasons unrelated to any deployment, and a
party’s return rises over the same period through an unrelated channel.

This is the second topology of Proposition ?, and it
exists to show that co-occurrence of a reduction and an increase is not the
predicate.

A deployment displaces a participant from a set of tasks and creates others that
the participant is positioned to take up, leaving the developmental repertoire
undiminished or enlarged.

The countermodel encodes the constraint the automation literature imposes
(Acemoglu & Restrepo, 2019): displacement and reinstatement are two effects and
their balance is empirical. A framework that could not represent this
configuration would have assumed its conclusion.

A participant’s material is in the corpus and the participant has no
developmental position that the system’s existence worsens.

This is the AI-specific form of the universal-contribution paradox’s dissolution.
Corpus membership is a historical fact and standing is a normative status, and
the countermodel exists to keep the first from being read as the second.

A system operates with no control edge, a participant’s capability falls, and the
system’s own returns rise.

The configuration is undefined under the negative answer to the standing
question, since a node without standing has no return in the relevant sense, and
becomes diagnosable under the positive one with the system in the benefiting
role. It is included to make the dependence on the standing question concrete
rather than rhetorical.

7. The Evidence Gate

This section states what evidence a claim in this area would require. Its
objective is to convert the portfolio’s recorded readiness deficit into an
itemized specification, so that the absence of the evidence is visible rather than
implicit, and so that a subsequent empirical project has a target.

The paper makes none of the claims below and asserts none of the antecedents.
Each gate states a claim, the evidence that would license it, and the reason the
evidence is not substitutable.

Claim: a specific participant’s material entered the corpus from which a
specific system was produced. Required: the assembler’s records of
sources, acquisition, filtering, and retention, or an inference procedure whose
error rate is characterized. Not substitutable by: demonstrations that a
system reproduces material resembling the participant’s, which are consistent with
several origins.

Claim: a participant’s effective generative capability is lower than it
would be under an admissible alternative arrangement. Required: a
declared alternative, a specification of the repertoire, and either identification
or characterized bounds. Not substitutable by: observed reductions in
income, employment, or output, which are neither necessary nor sufficient for a
repertoire contraction.

Claim: a party’s return rose through the reduction rather than alongside
it. Required: a witness path with the counterfactual property that
removing the reduction removes the increase. Not substitutable by:
correlation between a deployment and a return, or by the deploying party’s stated
rationale.

Claim: control of the relevant conditions is concentrated to a degree
that matters for the diagnosis. Required: a specification of which
conditions are necessary for which continuations, and measurement over that
specification. Not substitutable by: market share in adjacent product
markets.

Claim: a deployment displaced participants without reinstating them.
Required: the task-level evidence the automation literature specifies,
plus a separate argument connecting task displacement to repertoire contraction
(Acemoglu & Restrepo, 2019). Not substitutable by: either alone.

Claim: the relevant facts are withheld rather than unavailable in
principle. Required: the distinction between concealment and scale
opacity applied to the specific fact at issue (Burrell, 2016).
Not substitutable by: the observation that an explanation was not
provided.

Two remarks about the gate as a whole. It is demanding, and it is demanding in
the same way for claims in either direction: a claim that no exploitation is
occurring passes through
Gates ?–? exactly as a claim that it is. And it
is not a counsel of despair, because Proposition ? identifies
which gates are opened by disclosure and which are not.

8. Open Questions

This section records what the project leaves unresolved, by register.

8.1 Conceptual and Formal

Proposition ? treats merges of whole nodes. Institutions
overlap partially — shared ownership, interlocking control, joint ventures —
and the criterion for a partial merge is unknown.

The diagnostic requires a benefit-through witness, and nothing here says under
what conditions such a path is identifiable from data rather than assumed. The
causal apparatus for mediation exists (Pearl, 2009); its application
to a predicate rather than an outcome has not been worked out.

Proposition ? assumes a cost function on developmental positions
and exclusive occupation. Which regions of which spaces have these properties is
unknown, and the proposition is empty where they fail.

If most audits return $?$, what a body of such audits supports collectively is
undefined. The reporting discipline for individual quantities is available
(Manski, 1990); a discipline for populations of unresolved diagnoses is
not.

8.2 Political-Economic

Whether the provider–adopter distinction is stable, or whether integration
routinely internalizes the benefit paths that Corollary ?
identifies as diagnostically load-bearing, is an empirical question with direct
consequences for whether the framework can be applied at all.

Whether the returns at issue accrue as rent on control rights
(Birch, 2020), as platform position (Srnicek, 2017),
or as ordinary product revenue determines which witness structures are even
candidates, and the paper does not settle it.

Whether disclosure obligations of the kind now staged in at least one
jurisdiction (Commission, 2026) in fact open Gates ?
and ? is a question about implementation that the regulatory record
does not answer.

8.3 Normative

Whether the position of a party that cannot determine what was done to it, because
the determination depends on records held by another party, constitutes an
epistemic injustice in Fricker’s sense (Fricker, 2007) requires an
argument about the mechanism of the exclusion that this paper does not supply.

Whether responsibility in configurations with no identifiable perpetrator takes
the forward-looking structural form (Young, 2006; Wollner, 2019)
or admits liability depends on facts about anticipation that the framework
deliberately does not require.

Whether the distributive and republican criteria the source discussion invoked
(Rawls, 1999; Pettit, 1997) deliver convergent or divergent
verdicts on the same configuration is untested here and would be worth testing,
since a configuration on which they diverge would be informative about both.

Whether a system’s standing, if established, would place its provider in a
relation the framework already has vocabulary for
(Long et al., 2024; Vrousalis, 2013) is noted and not pursued.

8.4 Empirical

The evidence gate of Section 7 is the empirical programme. Nothing
in it has been attempted here, and the design of a study that would pass even one
gate is itself unresolved.

9. Limits and Revisable Research Program

This section states what the paper contributes, what would defeat it, and what it
has not done.

9.1 Contributions

The conceptual contribution is the replacement of the pairwise question by a
role-indexed one, together with the argument that a training corpus is an
assembled artifact rather than an inheritance, which converts a metaphysical
paradox into an institutional question.

The structural contribution is the aggregation criterion and its two directions:
that merging a controller with a mediator manufactures an attribution, and that
merging a supplier with a deployer destroys a witness and produces a negative
result that resembles a finding.

The formal contribution is four results: marginal underdetermination with its
counting corollary; the aggregation criterion; generic non-resolution under
opacity with disclosure as a decidability condition; and rate foreclosure with the
corollary that holdings equalization does not establish an alternative’s
admissibility.

The methodological contribution is the evidence gate, which states what any claim
in this area requires and applies symmetrically to claims in both directions.

9.2 Defeat Conditions

The conceptual contribution loses support if the four roles coincide in the
configurations of interest, since the pairwise form would then be adequate and the
paper would be solving a problem that does not arise.

The aggregation criterion loses support if diagnostic witnesses are robust to
merges in practice — if, for instance, benefit paths in these settings are
typically short and external, so that internalization is rare.

The opacity result loses support if the resolving facts turn out not to be
privately held, which disclosure regimes could in principle bring about and which
would be a welcome defeat.

The rate result loses support wherever exclusive occupation fails, which is
wherever a development can be independently undertaken by several parties.

The project as a whole loses interest if the evidence gate is never passed, since
a framework specifying conditions no investigation can meet is a description of an
impasse rather than a tool.

9.3 Development Status

The portfolio records this project as an AI-specific comparative and empirical
framework, and records that claims about training traces, capability effects,
concentration, substitution, and opacity require current system-specific evidence
and verified empirical sources.

The status of that deficit is as follows. The verified corpus stands at thirty-one
entries and includes the political economy of AI production, the automation
literature, the opacity and documentation literatures, and a verified regulatory
record. The comparative framework is supplied by the role table and the
countermodels. The empirical component is entirely absent: no system, provider,
or firm is named, no case is examined, and no claim requiring any of the six gates
is made. The paper is therefore a framework paper with an explicit specification
of what it does not have, rather than the comparative and empirical framework the
seed described.

Deferred questions are assigned. The diagnostic itself belongs to the capability
framework. Transmission along chains belongs to the propagation project.
Identification of repertoires belongs to the latent-generativity project. Closed
configurations belong to the cycles project. The commons object belongs to the
commons project. Standing control of expansion conditions belongs to the
political-economic project, and induced dependence to the revocable-infrastructure
project. Recovery and legal relations belong to projects not yet developed.

9.4 The Conservative Position

Stated as conservatively as the arguments permit: the question of who exploits
whom among the parties to an AI-mediated economy cannot be answered in pairwise
form, because the diagnosis quantifies over four roles and a pair fixes two; the
entity list matters, and merging entities can both manufacture and destroy
findings, so each merge must be justified for the question at hand; the facts that
would settle the matter are largely held by the parties whose conduct is at issue,
so most audits will be unresolved, and an unresolved audit is not a negative one;
a system that lets one party move faster than another can foreclose without any
difference in holdings; and a training corpus, whatever else it is, is a bounded
object someone assembled, which makes the question about it institutional rather
than metaphysical.

None of this establishes that any AI-mediated relation is exploitative, and none
of it establishes that none is. The paper’s position is that the second claim is
currently as unsupported as the first, and that saying so is more useful than
choosing.

References

Daron Acemoglu and Pascual Restrepo. Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2):3–30, 2019.

Elizabeth S. Anderson. What is the point of equality? Ethics, 109(2):287–337, 1999.

Kean Birch. Technoscience rent: Toward a theory of rentiership for technoscientific capitalism. Science, Technology, & Human Values, 45(1):3–33, 2020.

Jenna Burrell. How the machine `thinks’: Understanding opacity in machine learning algorithms. Big Data & Society, 3(1), 2016.

G. A. Cohen. The labor theory of value and the concept of exploitation. Philosophy & Public Affairs, 8(4):338–360, 1979.

Nick Couldry and Ulises A. Mejias. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Culture and Economic Life. Stanford University Press, Stanford, CA, 2019.

Kate Crawford. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, New Haven, CT, 2021. The Yale University Press catalogue record consulted lists the paperback, ISBN 9780300264630, and the eBook, ISBN 9780300252392.

European Commission. Regulatory framework for AI. Shaping Europe’s Digital Future, European Commission, 2026. Records that Regulation (EU) 2024/1689 entered into force on 1 August 2024 and became applicable on 2 August 2026, with prohibited-practice rules from 2 February 2025, governance and general-purpose model obligations from 2 August 2025, and obligations for systems used in employment and other listed high-risk areas from 2 December 2027. Consulted 16 August 2026.

Miranda Fricker. Epistemic Injustice: Power and the Ethics of Knowing. Oxford University Press, Oxford, 2007.

Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86–92, 2021.

Lina M. Khan. Amazon’s antitrust paradox. The Yale Law Journal, 126:710–805, 2017.

Robert Long, Jeff Sebo, Patrick Butlin, Kathleen Finlinson, Kyle Fish, Jacqueline Harding, Jacob Pfau, Toni Sims, Jonathan Birch, and David Chalmers. Taking AI welfare seriously, 2024.

Charles F. Manski. Nonparametric bounds on treatment effects. The American Economic Review, 80(2):319–323, May 1990. Papers and Proceedings.

Karl Marx. Grundrisse: Foundations of the Critique of Political Economy. Penguin Books, London, 1973. Translated with a foreword by Martin Nicolaus, in association with New Left Review; ISBN is that of the Penguin Classics reprint of this translation.

Karl Marx. Capital: A Critique of Political Economy, volume 1. Penguin Books, 1992. Translated by Ben Fowkes; original German edition published in 1867.

Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 220–229. Association for Computing Machinery, 2019.

Judea Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, Cambridge, 2 edition, 2009.

Philip Pettit. Republicanism: A Theory of Freedom and Government. Oxford Political Theory. Oxford University Press, Oxford, 1997.

John Rawls. A Theory of Justice. Harvard University Press, Cambridge, MA, revised edition, 1999.

Ingrid Robeyns. The capability approach: A theoretical survey. Journal of Human Development, 6(1):93–117, 2005.

John E. Roemer. A General Theory of Exploitation and Class. Harvard University Press, Cambridge, MA, 1982.

Paul M. Romer. Endogenous technological change. Journal of Political Economy, 98(5, Part 2):S71–S102, 1990.

Ruth J. Sample. Exploitation: What It Is and Why It’s Wrong. Rowman & Littlefield, Lanham, MD, 2003.

Amartya Sen. Capability and well-being. In Martha C. Nussbaum and Amartya Sen, editors, The Quality of Life, pages 30–53. Oxford University Press, Oxford, 1993.

Nick Srnicek. Platform Capitalism. Theory Redux. Polity Press, Cambridge, 2017.

Tiziana Terranova. Free labor: Producing culture for the digital economy. Social Text, 18(2 (63)):33–58, 2000.

Nicholas Vrousalis. Exploitation, vulnerability, and social domination. Philosophy & Public Affairs, 41(2):131–157, 2013.

Alan Wertheimer. Exploitation. Princeton University Press, Princeton, NJ, 1996.

Gabriel Wollner. Anonymous exploitation: Non-individual, non-agential and structural. Review of Social Economy, 77(2):143–162, 2019.

Iris Marion Young. Responsibility and global justice: A social connection model. Social Philosophy and Policy, 23(1):102–130, 2006.

Matt Zwolinski. Structural exploitation. Social Philosophy and Policy, 29(1):154–179, 2012.