PaperlayerView exact revision on arXiv

Source-linked AI summary

Transparency as Answerability: Role-Dependent Information Flow Modeling

Mattias Brännström, Themis Dimitra Xanthopoulou, Lili Jiang

arXiv:2610.16887v1cs.AI

TL;DR

AI transparency is widely required, but existing system- or component-focused approaches do not provide a structural account of stakeholder questions across organizational boundaries. The paper formalizes transparency as answerability over information flow networks, then applies the framework to a municipal recruitment-tool assessment. Its analysis shows that deployment composition changes what roles can know, ask, and act on, while the framework does not itself resolve the practical challenge of defining adequate roles.

  • Problem

    Existing approaches lack a structural account of how stakeholder questions can be answered across AI components, deployment contexts, and Provider–Deployer boundaries.

    Source: §1 · p. 1, §6 · p. 17, §6.4 · p. 20

  • Method

    The paper models transparency as answerability over information flow networks, including role-dependent access and questions evaluated across composed systems.

    Source: §7 · p. 23, §7.2 · p. 25, §9 · p. 28

  • Results

    The case study shows that the same Provider system can yield different distributions of epistemic access, decision weight, and accountability across deployment scenarios.

    Source: §8 · pp. 25–26

  • Takeaways & Limitations

    The analysis identifies answerable, partially answerable, or structurally blocked questions from system structure, revealing features component-level assessment cannot show.

    Source: §8 · p. 25

    The framework does not resolve the practical challenge of defining roles that are formally adequate and grounded in organizational reality, which requires domain expertise and AI literacy.

    Source: §8.3 · pp. 27–28

Abstract

from arXiv · show

Transparency in AI-supported decision systems is widely required but poorly defined. Existing approaches treat transparency as a property of systems or components, producing characteristic failures: assessments that do not transfer across organizational boundaries, stakeholder questions that cannot be formulated within the available frame of analysis, and component-level findings that say nothing about system-level behavior. We propose treating transparency as a relation: a system is transparent to a stakeholder to the degree that their legitimate questions are answerable from its structure. We formalize this using information flow networks, in which transparency questions are expressed in terms of network structure and answerability is the condition that the system settles them. The framework supports three dimensions of analysis: structural answerability from the network alone, compositional answerability across organizational boundaries for deployed systems, and role-dependent answerability given a stakeholder's accessible sub-network. We demonstrate the framework on a real pre-integration assessment of an AI recruitment tool in a public sector municipality. The assessment uses participatory design and interviews to model the decision system structure. We show how structural and role-dependent transparency questions can be formally stated and analyzed across the Provider and Deployer boundary where the AI system is composed together with the deployer's decision process.

1 Introduction

AI transparency is widely required, but existing approaches lack a structural account of what must be visible, to whom, and under what conditions. The paper treats transparency as a relation between a system and the legitimate questions of a stakeholder role.

1 Introduction

  • Existing component-level assessments can mistake model properties for system-level compliance, while Provider–Deployer boundaries obscure how the system relates to deployment and stakeholders.The paper identifies this as the framing trap and notes that questions such as why a candidate was rejected concern a socio-technical process, not an algorithm alone.Source: §1 · pp. 1–2
  • Transparency is a relational property: a system is transparent to a role insofar as it can answer that role’s legitimate questions.Roles are stakeholders with particular responsibilities and accountability relations, and their legitimate questions express what they need to know to meet those responsibilities.Source: §1 · pp. 1–2
  • The paper develops a formal account of answerability over information flow networks that model decision systems as information sites and transformation channels.The account builds on the Information Flow Methodology (IFM).Source: §1 · p. 2
  • The framework is grounded in a real pre-integration assessment of a high-risk AI recruitment tool with an external auditor at a public-sector municipality.Source: §1 · p. 2
  • The paper asks whether transparency can be characterized relationally, how questions spanning organizational boundaries can be answered, and how role-dependent access shapes answerability.Source: §1 · p. 2

2 Background

The background frames transparency as an information-demand and system-supply relation that must account for organizational boundaries and stakeholder roles. It distinguishes this structural question from related work on accountability, explanations, and recipient understanding.

2 Background

  • The EU AI Act distributes obligations between Providers and Deployers, making transparency analysis across their boundary necessary in multi-actor systems.In practice, the division may be blurred by post-delivery adaptation, organizations acting as both developer and user, and documentation produced from one actor’s perspective for another.Source: §2 · pp. 2–3

2.1 Related Work

  • Prior socio-technical accountability work rejects inferring system-level properties from component evaluation, while existing governance tools often address one side of organizational boundaries at a time.The framework proposes a shared information-flow representation connecting stakeholder demands with information generated, transformed, and accessed across technical and organizational processes.Source: §2 · p. 3, §2.1 · p. 4
  • XAI methods can produce explanations about model behavior, but do not determine whether a system’s structure can answer a stakeholder’s question.The paper’s answerability question precedes explanation extraction and delivery: it concerns whether information capable of settling the question exists in the structure.Source: §2 · p. 3, §2.2 · p. 5

2.2 Transparency, Answerability, and Explainability

  • The framework treats stakeholder information demands as legitimate questions and asks whether system structure can settle them, rather than which available answer is explanatorily best.Questions are understood through inquisitive semantics, where questions are characterized by the information states that settle them.Source: §2 · pp. 3–4, §2.2 · p. 5
  • Answerability is a structural precondition for meaningful explanations, but the framework does not assess whether recipients understand the information provided.The authors note that role-specific demands narrow the relevant information space, but do not guarantee that every answer or representation will be understood.Source: §2.2 · pp. 5–6

3 Transparency as Answerability

The formal framework represents decision processes as information networks and expresses questions as propositions over possible network structures. It defines answerability by whether a system proposition conforms to and settles the alternatives a question leaves open.

3 Transparency as Answerability

  • A system proposition conforms to a question when it satisfies the question’s shared presupposition; partial settling excludes at least one alternative, while full settling selects exactly one.These conditions make answerability a precise relation between the system proposition and the question’s alternatives.Source: §3 · pp. 6, 9
  • Information networks represent decision processes through sites holding information and directed channels connecting input sites to output sites.A refinement order captures when one network is more specific than another, providing a formal basis for comparing structures.Source: §3 · pp. 6–7
  • The network language and refinement order provide the structural objects used to represent what is known about a system and which more specific networks remain possible.An information state is a set of networks consistent with current knowledge; propositions over these states provide the framework’s question and answer objects.Source: §3 · pp. 6–7
  • Descriptions attach propositions about site values and channel transformations to a network without changing its underlying structure.The framework allows qualitative descriptions, run-specific values, probability distributions, and implementation transformation functions.Source: §3 · pp. 8–9

4 The Use Case

The case study models a municipality’s pre-integration recruitment process and a Provider’s AI tool separately, making their boundary visible before composition. It uses the models to examine role-based information needs and the shared interview template.

4 The Use Case

  • The pre-integration assessment used participatory interviews and document analysis to model the systems, keeping the Provider/Deployer boundary visible before integration.Stakeholder interviews and document analysis informed models that were iteratively constructed and validated with stakeholders.Source: §4 · p. 10

4.1 The Deployer Process

  • The illustrated senior recruitment process uses a requirement profile, but initial scoring and shortlist selection are informal and not formally documented.The four-grade initial scores and assessors’ broader impressions inform selection, complicating decision traceability and AI-tool integration.Source: §4.1 · p. 12
  • Different roles carry distinct responsibilities, and their legitimate questions span process structure, individual runs, past records, and patterns across cases.The illustrative role obligations use interviewed stakeholders’ professional responsibilities as a proxy and are not an exhaustive account of legal duties.Source: §4.1.1 · pp. 12–13

4.2 The Provider System

  • The Provider tool conducts template-governed AI chat interviews and returns ranked candidates with per-question scores; automated and manual scoring are mutually exclusive, but the output omits which path was used.The jointly configured interview template determines questions and scoring; recruiters receive the full Scored Template, while candidates see the final total score.Source: §4.2 · pp. 13–14, Figure 2 · p. 14

4.3 The Template as Shared Configuration

  • The shared interview template operationalizes the Deployer’s requirement profile as Provider-side questions and scoring criteria, forming the basis for composing the systems.Template construction is described as a preparatory flow orthogonal to the candidate flow.Source: §4.3 · pp. 14–15

4.4 Analytical Scope

  • The analysis examines answerability from each system’s structure, across the composed systems, and relative to stakeholder access.These dimensions are taken up in Sections 5, 6, and 7, respectively.Source: §4.4 · p. 15

5 Structural Answerability

Structural answerability asks what a network alone settles, while trace questions concern specific runs of an implementation. Applied to the Provider system, the analysis distinguishes what its structure determines from what requires access to traces or implementation details.

5 Structural Answerability

  • Structural questions are answerable from the Provider or Deployer network alone, and the network’s structure describes what a system does and how.The section defines structural questions as those determined by the structure of P or D alone.Source: §5 · p. 15, §5.1 · p. 15

5.1 What and How

  • A network channel answers a “what” question, while expanding it into sub-channels answers progressively more specific “how” questions at the network’s supported granularity.For P, candidates are filtered, interviewed, and then their answers are scored.Source: §5.1 · p. 15

5.2 Structural Questions over P

  • The Provider structure excludes CVs from scoring, permits four alternative scoring paths, and leaves the free-text scoring channel internally unspecified.The four alternatives are three automatic methods and one human fallback; the structure alone does not determine which applies to a candidate.Source: §5.2 · pp. 15–16

5.3 Trace Questions

  • An implementation trace settles values at the answer sites for a completed interview, though whether a stakeholder can access that trace is a separate question.The structural analysis finds no technical hindrance to those answer values being present in a trace; stakeholder access is addressed separately.Source: §5.3 · pp. 16–17
  • Counterfactual questions can be answered by searching settled traces, but full search is likely practically infeasible; trying alternatives requires access to the implementation and its sub-channels.The paper connects such questions to the right to meaningful explanation under Article 86, while noting that experimentation depends on structure and operational access.Source: §5.3 · p. 17

6 Composition and Answerability

This section examines fairness questions that cannot be answered from the Provider or Deployer alone, analyzing how their composition shapes decisions, information flows, and the meaning of outcomes. It also considers a conjecture that treats systematic deviations from a model as structural alternatives, with constructing an alternative serving as a diagnostic exercise into the deviation’s mechanism.

6.1 Composition

  • Composition situates P within D by preserving an unchanged sub-network, adapting D to accommodate P, and incorporating P into the resulting composed network.Questions that concern the recruitment process as a whole, such as whether candidates were treated fairly, require this composed system rather than P or D alone.Source: §6 · p. 17, §6.1 · p. 18

6.2 Integration Scenarios

  • MinD uses P for intake before human assessment, Assessment integration replaces human pre-scoring with P’s ranking and scored template, and Full integration also removes the selection meeting and adds a threshold.In Full integration, only top x proceed to personality testing and competence interview; the rest are excluded directly.Source: §6.2 · pp. 18–19

6.4 Effective Input, Impact and the Framing Trap

  • Decision attribution can be read structurally from upstream paths, and the same AI component can have different stakeholder effects because its output path varies by integration scenario.The model’s input path is consistent across scenarios, but its outputs have different downstream consequences; candidate answers are shaped by Template-governed questioning and may act as demographic proxies.Source: §6.3 · p. 19, §6.4 · p. 21

6.3 Decision Attribution

  • The composed system reveals how decision authority shifts across integration scenarios: MinD has no AI-driven rest cohort, AltD has an AI-determined rest cohort up to interview, and MaxD’s ranking-stage exclusions receive no human review.In MaxD, later human judgment applies only to the AI-selected top x, so its scope is bounded by the ranking threshold.Source: §6.3 · p. 20

6.4 Effective Input, Impact and the Framing Trap

  • Component-level assessment can miss consequential context: the framework asks where the AI channel sits, how outputs reach outcomes, and how inputs originate, including data outside the modeled system.The candidates enter through a third-party ATS fed by public job adverts, while language-dependent or non-inclusive interview interactions could make answers demographic proxies.Source: §6.4 · pp. 20–21

6.5 Decision Semantics

  • An outcome’s meaning depends on its full upstream path, and systematic deviations from the model can be represented as alternative sub-networks that produce outcomes by different paths.The paper conjectures that such an alternative network can approximate an implementation trace the original model does not sanction.Source: §6.5 · pp. 21–22
  • Constructing the alternative reveals which sites and channels are required to produce the deviation, making it a diagnostic exercise rather than merely a formal representation.The conjecture frames systematic deviation as expressible through an alternative network, but the structural requirements of that alternative are the substantive analytical question.Source: §6.5 · p. 22

7 Role-Dependent Answerability

Role-dependent answerability asks whether a stakeholder’s accessible, partially described view settles their questions. The case study shows how access boundaries and missing records constrain what different roles can establish.

7 Role-Dependent Answerability

  • Structural knowledge captures the sites and channels a role can reason about; operational access captures values it can read in a running system.The role model represents these as partially described sub-networks of the system and its implementation.Source: §7 · p. 22
  • When a described dependency also has an undescribed path, the maximal descriptive network cannot retain the described site as an intermediate, so ambiguous interior structure collapses to a weaker dependency.The collapse prevents visible fragments from licensing structural conclusions about channels they may participate in.Source: §7 · p. 23
  • Role-dependent answerability tests whether access suffices to settle a particular question, while role-adequate transparency evaluates the role’s position across its full information demand.The framework distinguishes structural access from access to settled values in a running implementation.Source: §7 · p. 23

7.2 Role-dependent answerability across the scenarios

  • Table 5 compares answerability before integration and across three integration scenarios, marking questions answerable, partial, or blocked and separating structural from operational judgments where they diverge.Some information demands are deferred to another role, while auditability questions are evaluated over the logged view of the composed process.Source: Table 5 · p. 24, §7.2 · p. 25

7.1 Examples of Role-Dependent Perspectives

  • The rejected candidate’s sparse structural knowledge supports a plausible but wrong interpretation of the interview, leaving them unable to formulate the questions needed to challenge it.The actual process separates information gathering from independent asynchronous scoring.Source: §7.1 · p. 24
  • An appeals assessor can identify unexplained low scores but cannot resolve them because their operational access omits the scoring path and internal logic.Candidate and assessor access is complementary and non-overlapping; contestability also depends on records covering the process structure.Source: §7.1 · pp. 24–25

7.2 Role-dependent answerability across the scenarios

  • Undisclosed human fallback and informal, unlogged Deployer deliberations were the two especially limiting constraints on answerability across the case study.The table marks these limitations with † and ‡, respectively.Source: Table 5 · p. 24, §7.2 · p. 25

8 Discussion

The case study shows that transparency can be analyzed as a relation between system structure and role-specific questions, while deployment composition shapes what each role can know, ask, and do. It also identifies structural deficiencies, policy implications, and limits of the framework.

8 Discussion

  • The same Provider system produces different distributions of epistemic access, decision weight, and accountability across deployment scenarios; composition, not the tool alone, determines what roles can know, ask, and act on.The case study treats relational transparency as an analytical instrument for determining whether questions are answerable, partially answerable, or structurally blocked.Source: §8 · pp. 25–26

8.1 Reflections on the Use-Case and Practical Application

  • Logging informal decisions and disclosing human-fallback use addressed deficiencies affecting multiple roles; changing one information-flow relation can improve several transparency relations.The Deployer introduced logging routines, and the Provider was required to identify candidates scored through the human fallback.Source: §8.1 · p. 26
  • The framework identifies where role-specific information is available, access-dependent, partial, or blocked, but does not determine which roles or information demands are normatively appropriate.It supports comparing alternative designs or compositions and locating changes to information flow, access, or functionality that could improve transparency.Source: §8.1 · p. 26

8.2 Policy Implications

  • Transparency obligations should connect to roles’ existing responsibilities: Provider obligations should enable Deployers to meet theirs, shifting compliance attention from disclosure to whether roles can do their work.The passage frames transparency as a precondition for exercising rights and fulfilling obligations, including obligations that precede AI use.Source: §8.2 · p. 27
  • Human oversight requires more than access to information: a role without a mandate or practical ability to act cannot exercise genuine oversight.The passage notes that delegated responsibility can separate the immediate user from the organizational level where accountability resides.Source: §8.2 · p. 27
  • The paper argues that deployment context, rather than the software product alone, is the more reasonable basis for high-risk classification, though this differs from current AI Act structuring.The same system can have different risk profiles depending on deployment context.Source: §8.2 · p. 27

8.3 Limitations & Challenges

  • Role definitions must reflect organizational reality and require expertise the framework does not supply; the framework is a modeling substrate that augments rather than replaces practical assessment tools.Many Deployers struggle to articulate transparency obligations and accountability practices in the presence of AI systems.Source: §8.3 · pp. 27–28

8.4 Further Directions

  • Answerability is not yet effective transparency: the answer must afford the role’s required action or judgment, and extending the framework to agency and inquirer understanding remains open.The passage also identifies understandability and simulability by the receiver as requirements for effective transparency.Source: §8.4 · p. 28

9 Conclusion

The paper formalizes transparency as answerability between a system and role-specific questions, using information-flow networks to analyze structural, compositional, and role-dependent answerability. It illustrates the framework in a real pre-integration assessment and identifies effective transparency—answerability that affords action—as a key open direction.

9 Conclusion

  • The paper develops transparency as answerability between a system and the questions a specific role needs answered, and analyzes it structurally, compositionally, and by role-dependent access.The framework is illustrated in a real pre-integration assessment.Source: §9 · p. 28
  • Information-flow networks provide a shared structure for representing and querying technical components, human processes, organizational roles, and their compositions.Source: §9 · p. 28
  • Effective transparency remains an open direction: answerability must afford the action a role requires, not merely provide an answer.Source: §9 · p. 28

Related papers

Loading 2610.16887v1…