Can Responsibility Be Measured? Cultural Mediation and the Operationalisation of the RESC Framework in AI-Mediated Decision-Making

Abstract

Artificial intelligence is increasingly embedded in decision-making processes where human judgement, algorithmic recommendations, organisational procedures, and institutional norms interact. While AI governance has focused on accountability, transparency, oversight, and liability, how human responsibility itself can be operationalised remains comparatively underdeveloped: existing research predominantly examines responsibility attribution after AI-mediated outcomes, causal responsibility, or meaningful human control, with less attention to responsibility as a prospective human capacity – to recognise consequences, exercise agency, reflect critically on AI-supported choices, intervene when necessary, and assume responsibility for resulting decisions. Moreover, emerging cross-cultural evidence suggests perceptions and attribution of responsibility are not culturally invariant. This paper addresses the question: how can human responsibility be operationalised and measured in AI-mediated decision-making when responsibility itself is culturally mediated? Building on the Responsibility Culture (RESC) framework, the paper proposes a conceptual measurement architecture distinguishing a common Responsibility Capacity from its culturally mediated interpretation and behavioural expression. Candidate dimensions include awareness of consequences, agency, reflexivity, control and intervention, responsibility assumption, and responsible action. Rather than proposing a universal responsibility score, the framework adopts a multimethod approach combining self-reported, behavioural, contextual, and culturally situated indicators.

Introduction

Artificial intelligence is altering not only how decisions are made but how responsibility for those decisions is understood, distributed, and exercised. As AI systems increasingly mediate consequential choices, existing accountability frameworks have concentrated on transparency, oversight, and liability, and research has predominantly examined responsibility through retrospective attribution or causal responsibility[1,5], while the prospective human capacity to act responsibly within AI-mediated environments remains comparatively underexamined.

The problem becomes more complex when responsibility is culturally mediated. Agrawal et al.[3] found that OECD participants attributed greater responsibility to human experts than to AI, whereas Indian participants attributed equal responsibility to both – reason enough not to assume responsibility has an identical interpretation across cultural contexts, or a universal score risks turning a socially situated construct into an apparently objective but culturally biased metric. The measurement challenge is therefore not how much responsibility an individual displays, but what it means within a decision context, what capacities enable its exercise, and whether behavioural manifestations retain equivalent meanings across cultural settings.

This paper consequently addresses two related research questions – RQ1. How can human responsibility be operationalised and measured in AI-mediated decision-making when responsibility itself is culturally mediated? RQ2. Which dimensions of human responsibility can constitute a common measurement core, and which require culturally situated interpretation? – by proposing an initial RESC Responsibility Measurement Framework: not a validated psychometric scale, but a conceptual measurement architecture intended to support subsequent empirical testing and cross-cultural validation. The proposed approach makes three contributions: it distinguishes responsibility capacity from retrospective responsibility attribution; it identifies candidate dimensions through which responsibility capacity may become empirically observable, integrating self-reported and behavioural evidence; and it introduces cultural mediation as a constitutive measurement problem rather than an external demographic variable.

From Responsibility Attribution to Responsibility Capacity

Research on responsibility in human–AI interaction has developed along several partially overlapping trajectories – the quantification of causal responsibility, experimental attribution of responsibility or blame, the assessment of meaningful human control, and behavioural measures of reliance and override – collectively showing that responsibility-related phenomena can be empirically investigated, while revealing an important fragmentation: existing measures frequently capture who is perceived as responsible, how causal responsibility is distributed, or whether humans retain control, rather than whether individuals retain and exercise a broader capacity to act responsibly in AI-mediated environments.

One of the most systematic attempts to quantify responsibility is the Responsibility Quantification Model (ResQu) developed by Douer and Meyer[1], which conceptualises responsibility in terms of the relative contribution of human and automated agents to decision outcomes; subsequent experimental work compared theoretical, measured, and subjective responsibility, finding meaningful correspondence between them. ResQu therefore shows that responsibility-related constructs can be operationalised quantitatively, but its object is comparative causal responsibility within a human–automation system, not responsibility as a broader human capacity.

A second strand uses experimental scenarios or vignettes to investigate responsibility attribution, asking participants to allocate responsibility or blame among human actors, AI systems, developers or organisations while manipulating factors such as autonomy, capability, task importance, and outcome success or failure[5]. Such studies illuminate the social attribution of responsibility, but remain predominantly evaluative and retrospective: participants determine who should be considered responsible for an outcome that has already occurred.

A third trajectory concerns Meaningful Human Control (MHC), through subjective, normative and moral control indicators such as perceived control, engagement, intervention, reaction time and system telemetry[2]; this literature is relevant to RESC because it combines subjective and behavioural indicators rather than relying solely on self-reported attitudes. Nevertheless, control and responsibility remain analytically distinct: an actor may formally retain the ability to intervene while failing to recognise an emerging consequence, and responsibility attributed to a human operator should be commensurate with their actual ability and authority to control the system[2] – intervention alone cannot constitute a sufficient behavioural proxy for responsibility.

Perceived agency has similarly been operationalised through capability, autonomy and contribution without becoming a general responsibility-specific instrument, while awareness of consequences, reflexivity and willingness to assume responsibility remain less developed still. The resulting gap is therefore not an absence of responsibility-related measurement, but an absence of an integrated approach to responsibility as a prospective, multidimensional and culturally situated human capacity.

Table 1. Existing Approaches and the Proposed RESC Measurement Perspective.

Approach	Primary construct	Main analytical question
ResQu	Causal responsibility	How much causal responsibility belongs to the human?
Attribution paradigms	Responsibility/blame attribution	Who should be considered responsible?
Meaningful Human Control	Human control over AI-supported action	Does the human retain meaningful control?
Agency/accountability-intention	Agency, willingness to accept accountability	Does the individual accept accountability?
RESC (proposed)	Responsibility Capacity	Can the actor recognise, exercise and assume responsibility?

Each approach nonetheless leaves a gap relative to RESC[1,2]: ResQu does not operationalise broader responsibility capacity; attribution paradigms remain predominantly retrospective; Meaningful Human Control is necessary but does not exhaust responsibility; and agency/accountability-intention constructs remain fragmented.

From Attribution to Capacity

RESC therefore proposes a shift in the unit of analysis: instead of “Who is responsible?”, it asks “What enables an actor to exercise responsibility?” This introduces a prospective dimension – responsibility becomes relevant not only after an outcome requiring explanation, but during the process through which an AI-mediated decision is interpreted, evaluated and enacted.

On this basis, the paper defines Responsibility Capacity provisionally as: the situated capacity of a human actor to recognise the potential consequences of an AI-mediated decision, perceive and exercise agency within it, critically reflect upon available recommendations, intervene or challenge them where appropriate, and remain willing to assume responsibility for the resulting course of action. The definition avoids equating responsibility with control or intervention alone: relevant behaviour may include verification, information seeking, justified acceptance, modification, challenge or override, building on behaviours already employed in automation and supervisory-control paradigms.

However, the transition from definition to measurement introduces another problem: if perceptions and expressions of responsibility differ across cultural contexts, a measurement architecture cannot assume that identical responses or behaviours represent identical underlying responsibility capacities.

Responsibility as a Culturally Mediated Capacity

If responsibility is conceptualised as a human capacity rather than solely as the retrospective attribution of an outcome, its measurement cannot be separated from the social and cultural environments in which that capacity is interpreted and exercised. Responsibility is not enacted in an institutional vacuum: expectations concerning autonomy, authority, obligation, acceptable risk, collective welfare, and the legitimacy of challenging decisions shape how individuals understand both their capacity and their duty to intervene.

This issue is particularly relevant in AI-mediated decision-making. In a comparison between participants from OECD countries and India, Agrawal et al.[3] found that OECD participants attributed greater responsibility to human experts than to AI, whereas Indian participants attributed responsibility equally to human and AI experts. Although a country-group comparison cannot establish the cultural mechanisms responsible for these differences, the finding challenges the assumption that responsibility attribution can automatically be interpreted through a culturally invariant metric. The methodological problem extends beyond responsibility itself: cross-cultural validation of adjacent AI-related constructs remains limited, and the AIAS-4 study identified in the literature review achieved configural and partial metric invariance across Poland, the United Kingdom and the United States, but not scalar invariance[4]. Consequently, even where respondents appear to engage with a common underlying construct, direct comparison of mean scores across cultural groups may not be warranted.

For the measurement of responsibility, this distinction is fundamental. A universal instrument could produce numerically comparable scores while masking culturally different understandings of what responsible action entails: willingness to challenge an AI-supported decision may indicate responsible critical engagement in one setting, but in another may be shaped by stronger expectations concerning institutional authority, hierarchy, collective decision-making, or relational obligation. These possibilities should be treated as hypotheses requiring empirical examination rather than assumed cultural differences.

Culture as Mediation Rather Than Country

The proposed RESC approach therefore distinguishes cultural mediation from simple cross-national comparison. Nationality can provide a sampling variable, but should not itself be treated as a measurement of culture: the relevant analytical question is not whether participants in different countries obtain different responsibility scores, but which culturally embedded meanings, norms and orientations influence the interpretation and enactment of responsibility. This has two methodological implications: future validation of RESC should directly measure theoretically relevant cultural constructs rather than infer them from country of residence, and cultural mediation should be examined in relation to specific components of Responsibility Capacity rather than introduced as a generic explanatory variable. For instance, cultural orientations may plausibly influence how actors understand individual versus collective agency, their willingness to question institutional or technological authority, their interpretation of obligations toward affected others, or the threshold at which intervention becomes perceived as necessary – relationships that remain an empirical task for subsequent research.

Accordingly, RESC does not propose that responsibility itself is entirely culturally relative. Instead, it raises a more specific measurement question: can a common core of human responsibility capacities be identified while allowing their interpretation and behavioural expression to remain culturally situated? This distinction between a common responsibility core and culturally mediated expression provides the conceptual basis for the proposed measurement architecture.

Emerging empirical work lends support to this approach. Cerar et al.[6], studying 172 managers across seven countries, found that cultural tightness shapes reliance on AI in augmented decision-making and interacts with the degree of government endorsement of AI. Chien et al.[7,8] documented cross-cultural differences in trust in automation between participants from the United States, Taiwan and Turkey, and developed an intercultural scale for measuring trust in automation rather than inferring it from nationality alone. More directly, Yuan and Lee[9] found that individualism–collectivism orientations shape how users evaluate AI-generated explanations. Together, these studies indicate that directly measured cultural orientations – not nationality as a proxy – are what predict AI-related trust and reliance, supporting RESC’s proposal to measure cultural orientations rather than treat country membership as a stand-in for culture; whether this extends to Responsibility Capacity specifically remains to be tested.

A Common Core Without a Universal Responsibility Score

The literature reviewed in Sect. 2 suggests several candidate components for such a common core. Human control has already been operationalised through subjective, normative, moral and behavioural indicators, while agency has been investigated through autonomy, capability, contribution and perceived responsibility; awareness of consequences, reflexivity and willingness to assume responsibility remain comparatively underdeveloped as validated constructs. RESC brings these fragmented components together by hypothesising that Responsibility Capacity can be observed through a set of related but analytically distinct dimensions: awareness of consequences; agency; reflexivity; control and intervention; responsibility assumption; and responsible action – candidate dimensions derived from the intersection between RESC’s theoretical premises and existing empirical approaches, not yet a validated scale.

Nor does the framework assume that these dimensions should simply be summed into a universal Responsibility Score, which would obscure three distinctions: between declared and enacted responsibility; between capacity and context; and between common constructs and culturally mediated interpretations. Instead, RESC proposes that responsibility should ultimately be examined through a multimethod profile, combining subjective, behavioural and contextual evidence.

The RESC Responsibility Measurement Framework

The preceding analysis produces a measurement problem with three components: existing approaches capture fragmented aspects of responsibility; several relevant dimensions remain weakly operationalised; and cultural mediation complicates assumptions of universal measurement equivalence. The RESC Responsibility Measurement Framework (RRMF) addresses these problems; its basic structure is: AI-Mediated Decision Context → Responsibility Capacity ↔ Cultural Mediation → Responsible Action.

[FIG. 1 — RESC Responsibility Measurement Framework diagram]

Fig. 1. RESC Responsibility Measurement Framework: awareness of consequences, agency, reflexivity, control/intervention and responsibility assumption feed into culturally mediated, behaviourally observable Responsible Action.

Candidate Dimensions of Responsibility Capacity

Table 2. Candidate Dimensions and Measurement Strategies in the RESC Framework.

Dimension	Operational meaning	Measurement mode
Awareness of consequences	Recognition of potential consequences and affected actors	Self-report + scenario
Agency	Perceived and effective capacity to influence the decision	Self-report + behavioural
Reflexivity	Critical examination of one’s own and AI-supported judgement	Scenario + behavioural
Control & intervention	Effective capacity to affect the AI-mediated process	Behavioural + self-report
Responsibility assumption	Willingness to recognise responsibility for the outcome	Self-report + vignette
Responsible action	Observable enactment of responsible judgement	Behavioural

Responsible Action does not automatically mean “opposing the AI”: accepting a recommendation after critically verifying it can be just as responsible as rejecting it. The paper’s methodological contribution is therefore not to measure opposition to AI, but the quality of the exercise of responsibility in its presence.

Multimethod Measurement Architecture

The operationalisation of Responsibility Capacity requires distinguishing what individuals report, what they actually do, and the cultural and organisational conditions within which decisions occur – since a single mode risks capturing only one manifestation of responsibility, a concern reinforced by evidence that self-reported, measured and behavioural responsibility can diverge and that control indicators alone do not exhaust it[1,2]. RESC therefore proposes three complementary layers: self-reported capacity, behavioural enactment, and cultural-contextual interpretation.

Self-Reported Responsibility Capacity.

The first layer examines actors’ perceptions of their own capacity to exercise responsibility, through items addressing awareness of consequences, agency, reflexivity, capacity to intervene, and willingness to assume responsibility. Self-report is valuable because responsibility includes subjective recognition, but cannot alone demonstrate responsible behaviour; RESC does not treat declared responsibility as equivalent to enacted responsibility.

Behavioural Responsibility.

The second layer examines behaviour in controlled AI-mediated decision scenarios: whether participants verify the recommendation, seek information, identify affected stakeholders, modify or override it, or accept it after critical evaluation. Responsible behaviour is not synonymous with disagreement with AI: the object of measurement is the quality of engagement with the decision process, not the frequency of opposition to automated recommendations.

Cultural and Contextual Interpretation.

The third layer addresses RESC’s central premise: responsibility is socially and culturally situated. Rather than assigning a cultural profile from nationality alone, subsequent work should directly measure orientations concerning individual versus collective agency, authority and hierarchy, uncertainty, and relational and institutional expectations, examining their relationship with specific Responsibility Capacity dimensions[3,6,7,8,9].

From a Responsibility Score to a Responsibility Profile

RESC does not advocate a single universal Responsibility Score; instead, an instrument should generate a Responsibility Profile: Self-reported Capacity × Behavioural Enactment × Cultural/Contextual Interpretation → RESC Responsibility Profile (the × is conceptual, not mathematical; no scoring function is proposed at this stage).

Validation Strategy

The proposed framework requires empirical validation before any claim can be made that Responsibility Capacity has been successfully measured. Table 3 outlines a staged validation programme, from construct development through cross-cultural and cultural-mediation analysis. Cultural differences should not be interpreted before establishing that the instrument functions comparably across groups: measurement-invariance work on an adjacent AI-attitude construct illustrates why this check is necessary, having found only partial invariance across three national samples[4] – a methodological caution, not evidence that RESC itself has been validated.

Table 3. Proposed RESC Validation Roadmap.

Stage	Expected output
1. Construct development (literature synthesis, expert review, cognitive interviews)	Initial RESC instrument
2. Behavioural pilot (AI-mediated scenarios)	Behavioural indicator set
3. Psychometric validation (factor analysis, reliability/validity testing)	Validated dimensional model
4. Cross-cultural validation (multi-group invariance testing)	Comparable/non-comparable dimensions
5. Cultural mediation (moderation/mediation analysis)	Empirical cultural mediation model

Discussion, Governance Implications and Limitations

Responsibility cannot be adequately understood solely through the allocation of blame or causal contribution after an outcome has occurred. RESC changes the analytical question: rather than asking exclusively who is responsible, it asks whether human actors possess and exercise the capacities through which responsibility becomes possible – particularly relevant where human oversight exists formally but may become substantively weak, echoing evidence that declared, measured and behavioural responsibility can diverge and that meaningful control requires more than a formal ability to intervene[1,2]. Accountability mechanisms remain necessary, but predominantly explain and attribute decisions after the fact; RESC introduces a prospective question about the conditions that enable responsible judgement beforehand, treating cultural variation – for which directly measured orientations, not nationality, appear to matter[3,6,7,8,9] – as theoretically meaningful, not statistical noise.

Implications for AI Governance

If subsequently validated, the RESC approach could contribute to AI governance at three levels: individually, by evidencing whether users recognise consequences, perceive agency and critically engage with AI-supported decisions; organisationally, by helping identify environments where formal oversight exists but substantive engagement is weak; and at the governance level, by complementing compliance mechanisms with empirical attention to the human capacities meaningful responsibility requires. RESC should not be interpreted as transferring responsibility from organisations onto individual users: measurement should assess both human capacity and the conditions under which it can be exercised.

Limitations and Research Agenda

The present contribution is conceptual. The proposed dimensions have not yet been psychometrically validated, and the framework does not establish causal relationships between cultural orientations and Responsibility Capacity. Several candidate dimensions overlap with constructs already studied in human factors, moral psychology and human–AI interaction; future research must demonstrate discriminant validity rather than simply relabelling existing measures as responsibility. Cultural constructs are themselves contested and should not be treated as fixed characteristics of national populations. Finally, the relationship between the candidate dimensions remains open and should be resolved empirically rather than imposed conceptually.

Conclusion

As artificial intelligence becomes increasingly embedded in human decision-making, the question of responsibility cannot be limited to determining who should answer for an outcome after it occurs. This paper has proposed the RESC Responsibility Measurement Framework as an initial architecture for operationalising responsibility as a prospective conception: rather than a universal Responsibility Score, RESC distinguishes between Responsibility Capacity, behavioural enactment and cultural mediation, examined through complementary self-reported, behavioural and contextual measures.

The framework therefore reframes the measurement question: not simply “Who is responsible?” but “Under what conditions can human actors recognise, exercise and assume responsibility in AI-mediated decision-making?” Its contribution is deliberately preliminary, yet these questions matter for AI governance: a system capable of assigning accountability without preserving the human capacity to exercise responsibility may succeed procedurally while remaining insufficient socially.

The authors have no competing interests to declare that are relevant to the content of this article.

1. Douer, N., Meyer, J.: Theoretical, Measured, and Subjective Responsibility in Aided Decision Making. ACM Trans. Interact. Intell. Syst. 11(1) (2021). https://doi.org/10.1145/3425732

2. Cavalcante Siebert, L., Lupetti, M.L., Aizenberg, E., Beckers, N., Zgonnikov, A., Veluwenkamp, H., Abbink, D., Giaccardi, E., Houben, G.-J., Jonker, C.M., van den Hoven, J., Forster, D., Lagendijk, R.L.: Meaningful human control: actionable properties for AI system development. AI Ethics 3(1), 241–255 (2023). https://doi.org/10.1007/s43681-022-00167-3

3. Agrawal, V., Kandul, S., Kneer, M., Christen, M.: From OECD to India: Exploring cross-cultural differences in perceived trust, responsibility and reliance of AI and human experts (2023)

4. Talik, W., Talik, E.B., Grassini, S.: Measurement invariance of the artificial intelligence attitude scale (AIAS-4): cross-cultural studies in Poland, the USA, and the UK. Curr. Psychol. 44(19), 15758–15766 (2025). https://doi.org/10.1007/s12144-025-08348-z

5. Tsumura, T., Yamada, S.: Effects of knowledge and importance on responsibility in human-AI decision making. Sci. Rep. 16, 2670 (2026). https://doi.org/10.1038/s41598-025-32513-w

6. Cerar, J., Minbaeva, D., Nell, P.C.: Reliance on AI in augmented strategic decision-making: Navigating cultural and national dynamics. Int. Bus. Rev. 35(4), 102600 (2026). https://doi.org/10.1016/j.ibusrev.2026.102600

7. Chien, S.-Y., Lewis, M.J., Sycara, K.P., Liu, J.S., Kumru, A.: The Effect of Culture on Trust in Automation: Reliability and Workload (2018)

8. Chien, S.-Y., Semnani-Azad, Z., Lewis, M., Sycara, K.: Towards the Development of an Inter-Cultural Scale to Measure Trust in Automation. In: Proc. 6th Int. Conf. Cross-Cultural Design, pp. 35–46 (2014)

9. Yuan, C.W.(T.), Lee, Y.-H.: How cultural cognition affects trust and perceived quality of AI explanations. Comput. Hum. Behav. Rep. 22, 101021 (2026). https://doi.org/10.1016/j.chbr.2026.101021