What Kind of AI-Mediated Society Are We Building? Responsibility and the Human Choices Behind Automation

Congress: International Conference on the Social Impact of AI
Accepted for Presentation

Artificial intelligence increasingly mediates and replaces functions previously performed by humans. This conceptual paper examines responsibility for the decision to delegate, distinguishing technical substitutability from social delegability. Building on task-delegability research, meaningful human control, responsibility gaps, responsible innovation, and impact assessment, it develops a Responsibility Culture based framework comprising Purpose, Function, Displacement, Conditions, and Trajectory. Its contribution is an integrative decision procedure that connects the justification of delegation with the preservation of human capabilities and the monitoring of cumulative change. The procedure compares human, assisted, hybrid, and substituted configurations, screens for non-compensable constraints, documents contextual priorities and disagreement, and specifies review and suspension conditions. A hypothetical recruitment example applies all five dimensions and illustrates why efficiency gains alone do not justify automatic exclusion. Trajectory assessment uses scenarios, baseline indicators, repeated reviews, and escalation across organisational and sectoral levels rather than deterministic prediction. The framework is proposed as a deliberative heuristic, not a validated scoring instrument. It supports conditional authorisation, redesign, or non-automation while making the assumptions and responsibilities behind each choice available for scrutiny.

Keywords: Responsibility Culture · AI delegation · Human agency · Social delegability · Responsible innovation

1 Introduction

Artificial intelligence is increasingly embedded in the infrastructures through which contemporary societies work, communicate, organise services, access information, and make decisions. Its role is no longer limited to providing computational assistance. AI systems recommend, rank, allocate, evaluate, communicate, predict, and increasingly perform functions that were previously associated with human workers, professionals, managers, and intermediaries.

This transformation is already visible in the organisation of work. Algorithmic management systems can organise, assign, monitor, supervise, and evaluate labour, while workers may increasingly interact with a system rather than with a human manager [1]. Similar developments are occurring in customer service, banking, logistics, healthcare, recruitment, education, and public administration. The distinction between AI assistance, AI mediation, AI delegation, and AI substitution is consequently becoming socially significant.

These developments offer substantial potential benefits. AI may increase accessibility, improve consistency, reduce repetitive work, support human decision-making, increase organisational capacity, and make some services available at scales and times that would otherwise be impossible. The purpose of responsible AI governance cannot therefore be to preserve human performance of every task regardless of technological progress.

At the same time, organisational decisions concerning automation are frequently driven by measurable criteria such as performance, scalability, productivity, availability, and cost. These are legitimate considerations, but they do not exhaust the social meaning of replacing a human function.

A technically replaceable task may also embody judgement, discretion, care, reciprocity, professional knowledge, social recognition, responsibility, or the possibility of human contestation. Removing the human actor may therefore alter more than the production process. It may change the conditions under which agency is exercised, decisions are challenged, responsibility is assumed, and social relationships are maintained.

Technical substitutability does not imply social delegability.

The capacity of an AI system to perform a function does not, by itself, determine whether delegating that function is socially desirable or responsible.

Existing scholarship provides substantial resources for examining this problem. Research on automation bias demonstrates that decision-support systems may generate over-reliance under identifiable conditions [2, 3]. The responsibility-gap literature examines how AI complicates culpability, moral accountability, public accountability, and active responsibility [4]. Meaningful Human Control (MHC) asks whether human actors retain genuine rather than merely nominal capacities to understand, intervene, and respond [4]. Task-delegability research asks what humans are willing to delegate and identifies motivation, difficulty, risk, and trust as relevant factors [5]. AI impact-assessment frameworks increasingly evaluate effects on individuals, groups, and society throughout the system lifecycle [6].

These approaches already address the justification and governance of automation. The present paper proposes an integrated procedure focused on the human and social functions changed by delegation, while making its overlaps with earlier work explicit.

This paper proposes an additional perspective: the human choice to delegate a function to AI should itself become an explicit object of responsibility.

The central research question is therefore:

How can responsibility guide decisions about whether, why, and under what conditions human functions should be delegated to AI, particularly when the cumulative societal consequences of such choices remain uncertain?

A related question concerns the contribution of Responsibility Culture (RESC):

How can Responsibility Culture extend existing approaches to AI governance by connecting responsibility in AI use, responsibility in AI deployment, and responsibility for the cumulative societal trajectory produced by automation choices?

To address these questions, the paper develops a conceptual framework for responsibility-informed AI delegation based on five dimensions: Purpose, Function, Displacement, Conditions, and Trajectory. The objective is not to prescribe universally which functions should remain human. It is to expand the criteria through which automation decisions are evaluated and to make visible the responsibility involved in shaping an increasingly AI-mediated society.

2 From AI Assistance to an AI-Mediated Society

AI adoption is often described through the language of automation. Yet automation encompasses different relationships between humans and technological systems.

At one end, AI can assist human activity by providing information, predictions, summaries, or recommendations while leaving the substantive task and decision with the human actor. AI can also mediate an activity by structuring the information, interaction, options, or workflow through which human action occurs. A further step involves delegation, in which a system performs part of a function that would otherwise require human judgement or action. Finally, substitution occurs when the system performs a function with limited or no routine human participation.

The workplace illustrates this progression. The International Labour Organization defines algorithmic management as the use of algorithmic systems to organise, assign, monitor, supervise, and evaluate work [1]. Importantly, under such arrangements workers may interact with a system rather than a human manager. Algorithmic management therefore does not simply automate isolated calculations. It can modify the relational architecture of work.

Algorithmic management is not limited to one sector. Conceptual analysis across platform work and conventional workplaces examines how task allocation, monitoring, and evaluation can change work organisation and conditions, with implications that depend on the institutional context [15].

The consequences of AI delegation likewise extend beyond productivity. Existing evidence identifies effects and potential effects across agency and responsibility, social relationships, legitimacy and public values, inequality and political cohesion, and the distribution of human skills. However, the evidentiary strength is uneven. Immediate effects on behaviour and perception are better established than long-term societal transformation.

AI-mediated interactions can generate real benefits. Experimental research on AI companionship suggests that such systems may reduce loneliness under some conditions, with feeling heard contributing to the reported benefit [7]. Such findings make a simplistic opposition between “human interaction” and “artificial interaction” untenable.

At the same time, short-term functional equivalence does not necessarily imply social equivalence. A system that successfully performs one observable aspect of a human function may not reproduce all the downstream relational, institutional, or social consequences associated with that function.

The same problem appears in public decision-making. In high-stakes domains such as taxation, justice, or child protection, technically accurate or explainable systems must still operate within institutional environments characterised by discretion, professional knowledge, accountability relationships, and opportunities for contestation [8]. The relevant question is therefore not simply whether an algorithm can generate a prediction. It is also who can decide, explain, challenge, correct, and ultimately answer for the consequences.

This suggests that AI substitution should not be conceptualised exclusively as the replacement of labour input. It can also represent the replacement or transformation of social functions.

A customer-service employee, for example, does more than transmit standard information. Depending on the context, the function may include recognising exceptional circumstances, interpreting ambiguity, responding to vulnerability, exercising discretion, escalating a problem, or providing an identifiable human point of responsibility. An AI system may perform many routine elements of this role extremely effectively. The responsible question is therefore not whether customer service should remain human in general, but which elements can appropriately be automated, which human functions may be displaced, and what mechanisms must remain when the automated interaction is insufficient.

The emergence of an AI-mediated society should consequently be understood not as a single technological event but as the cumulative result of many decisions about where AI assistance becomes mediation, where mediation becomes delegation, and where delegation becomes substitution.

3 Human Agency and Responsibility Under AI Delegation

Concerns about AI delegation often assume that greater automation necessarily entails less human agency. Existing evidence does not justify such a simple causal claim.

The strongest evidence concerns automation bias and over-reliance. Automation bias occurs when users rely excessively on automated decision support, reducing vigilance in information seeking and verification [2]. Reviews identify factors including trust, confidence, workload, task complexity, time pressure, and user experience as mediators of this effect [2, 3]. The reviewed evidence supports attention to training, verification demands, and organisational context rather than assuming that a human reviewer will reliably correct automation errors [2, 3].

The important implication is not that AI assistance inevitably weakens human judgement. Rather, the design and organisational context of AI use can alter the conditions under which judgement is exercised.

A person may remain formally “in the loop” while exercising very little substantive judgement. If an employee lacks sufficient expertise, time, authority, evidence, or organisational permission to reject an AI recommendation, the existence of an approval button provides only nominal control.

This observation connects automation-bias research with the literature on responsibility gaps and Meaningful Human Control.

Santoni de Sio and Mecacci distinguish at least four interconnected responsibility gaps associated with AI: culpability, moral accountability, public accountability, and active responsibility [4]. Their analysis rejects the assumption that responsibility problems can be solved through a single legal or technical intervention and instead situates responsibility within sociotechnical systems designed for meaningful human control.

This distinction is particularly relevant to AI delegation. The governance objective should not be merely to identify a human actor to whom accountability can be attributed after an adverse outcome. It should also ensure that relevant human actors possessed the actual capacity to understand, contest, intervene, and assume responsibility.

This concern also intersects with Responsibility Culture (RESC). RESC conceptualises responsibility not solely as retrospective attribution but as a capacity that must be exercisable prospectively. In this perspective, awareness of consequences, agency, reflexivity, ability to intervene, responsibility assumption, and responsible action form part of the conditions through which responsibility becomes socially meaningful.

This framing resonates with recent work on responsible agency under conditions of AI-induced vulnerability, which argues that responsibility cannot be assessed independently of the material and epistemic conditions — access to information, time, standing — under which an agent is asked to act responsibly [14]. A capacity-based account of responsibility, in other words, must attend not only to whether an actor is formally positioned to respond, but to whether the surrounding conditions make responsible action genuinely available to them.

Importantly, the current empirical literature does not establish that repeated AI use produces a general and irreversible decline in human responsibility or competence. Evidence for durable deskilling is narrower and task-specific. Studies of information automation and ML decision support examine task-specific skill retention and inhibited learning [9, 10]. Macnamara et al. [11] provide a theoretical perspective identifying research questions rather than new experimental evidence.

Accordingly, claims that AI is producing a generally “deskilled” or “irresponsible” society would exceed current evidence.

The more defensible conclusion is:

AI delegation does not automatically remove human agency; it changes the conditions under which agency and responsibility are exercised.

This distinction has important governance consequences. If the long-term effects of AI delegation on skills, judgement, relationships, and responsibility remain partly uncertain, uncertainty cannot justify ignoring those dimensions. Instead, uncertainty itself becomes relevant to responsible decision-making.

4 Positioning the Contribution

This paper uses conceptual synthesis rather than a systematic review or an empirical study. It selects adjacent approaches for their relevance to delegation, agency, responsibility, and societal consequences, and develops a normative decision procedure from their intersection. The five dimensions are analytical prompts proposed by the author; their distinctness, usability, and effects require empirical evaluation.

The contribution is integrative rather than a claim that existing scholarship ignores the decision whether to automate. Lubars and Tan explicitly examine what should be delegated [5]. Meaningful Human Control connects autonomous action to relevant human reasons and responsible human actors through tracking and tracing conditions [16]. AI impact assessment includes broader lifecycle effects [6], while algorithmic impact assessment already asks whether public-sector automation is acceptable [18]. Responsible innovation also addresses the direction of technological change through anticipation, reflexivity, inclusion, and responsiveness [17].

The present proposal brings these concerns into one sequence focused on the task-plus-human-function being delegated. Purpose records the reason for changing an existing arrangement; Function identifies its less measurable human roles; Displacement examines what changes across alternative configurations; Conditions tests whether responsibility remains exercisable; and Trajectory requires repeated assessment of cumulative effects. None of these questions is claimed to be unprecedented in isolation.

Task-delegability research helps explain delegation preferences and perceived suitability [5, 13]. Here, those preferences inform deliberation but do not settle whose interests should govern a decision or which losses are acceptable. Meaningful Human Control provides a substantive basis for Conditions [16]; the present framework additionally asks whether a controlled arrangement is justified relative to alternative distributions of the function. Sociotechnical assessment supplies the analysis of institutional context [8], while the contribution here is to organise that analysis around explicit configuration choices and their revision.

Responsibility Culture (RESC) is the author’s developing conceptual research programme, not an established or empirically validated consensus model [19]. It treats responsibility as a human capability shaped by cultural and institutional conditions. The five delegation dimensions developed here are a domain-specific extension, distinct from the four dimensions of the broader RESC intercultural model. Its relationship to responsible innovation concerns the institutional practice of anticipating and revising consequences [17]; its relationship to responsible agency concerns whether people can actually exercise the responsibility assigned to them [14]. RESC supplies the interpretive perspective, while these established approaches supply important theoretical constraints.

Social delegability therefore means the contextual justifiability of delegation after considering performance, displaced functions, exercisable responsibility, and plausible cumulative effects. It is not a universal numerical property of a task. Sections 6 and 7 translate this concept into a proposed procedure and an illustrative application.

5 A RESC-Based Framework for Responsibility-Informed AI Delegation

Responsibility Culture provides a useful basis for this inquiry because it shifts attention from responsibility as a formal endpoint toward the conditions through which responsibility is recognised, exercised, and sustained [19].

The proposed Responsibility-Informed AI Delegation Framework comprises five dimensions: Purpose, Function, Displacement, Conditions, and Trajectory.

5.1 Purpose Why Are We Delegating?

The first dimension concerns the reason for automation.

Organisations may delegate functions to AI for many legitimate reasons: greater accuracy, safety, accessibility, continuity, scalability, speed, reduction of repetitive work, shortage of qualified personnel, improved consistency, or lower cost.

The responsibility question concerns whether the rationale for delegation adequately reflects the range of consequences associated with the decision.

A decision justified solely by labour-cost reduction may overlook effects on service quality, vulnerability, accessibility, institutional knowledge, professional judgement, or opportunities for contestation. Conversely, refusing automation solely to preserve existing roles may overlook improvements in safety, accessibility, or human welfare.

Purpose therefore requires explicit articulation of why delegation is being considered and whose interests the expected benefits serve.

5.2 Function What Is Actually Being Delegated?

Automation decisions are frequently described at the level of tasks. Responsibility requires an additional question: what human function is embedded within the task?

A task may carry functions that are not captured by its most measurable output.

A teacher provides information but may also interpret misunderstanding, encourage participation, recognise vulnerability, and model intellectual judgement.

A healthcare professional may interpret clinical information while also contextualising uncertainty and communicating consequences.

A manager may allocate tasks while also negotiating conflict, recognising exceptional circumstances, and assuming responsibility.

A customer-service worker may provide information while also exercising discretion and offering an accountable human point of escalation.

This does not mean that these functions cannot be technologically supported or partially automated. It means that responsible delegation requires identifying them before determining what is actually being replaced.

The unit of analysis should therefore move from task to task-plus-human-function.

5.3 Displacement What Is Gained and What May Be Lost?

Delegation creates benefits and displacements simultaneously.

The relevant comparison should therefore not be framed simply as:

human performance versus AI performance.

It should examine:

benefits generated versus functions displaced.

Possible benefits include efficiency, accessibility, consistency, safety, scalability, reduced workload, and improved decision support.

Possible displacement may involve human judgement, discretion, care, reciprocity, professional learning, opportunities for explanation, social interaction, responsibility, or institutional knowledge.

The existence of displacement does not automatically make automation undesirable. The question is whether the displaced element has a value that remains socially necessary and, if so, whether that value can be preserved through hybrid design or alternative arrangements.

This dimension is particularly important because current evidence on long-term social effects remains incomplete. Responsibility therefore requires distinguishing between established effects, plausible risks, and speculative concerns rather than treating them as epistemically equivalent.

A brief illustration clarifies the distinction. Consider the delegation of routine clinical triage to an AI system that prioritises patients based on reported symptoms and risk indicators. The measurable benefit is significant: faster initial assessment, consistent application of criteria, and expanded capacity in resource-constrained settings. The displacement, however, is not confined to the triage task itself. It may include the clinician’s opportunity to notice an atypical presentation that does not fit standard risk indicators, the patient’s opportunity to be heard by a person capable of adjusting course in real time, and the informal channel through which clinical staff historically identified emerging patterns — an early cluster of unusual symptoms, for instance — before they were formally recognised. None of these displaced elements is necessarily irreplaceable; some can be reconstructed through alternative mechanisms, such as structured escalation protocols or periodic clinical review of triage outcomes. The point of the Displacement dimension is not to conclude that triage automation is undesirable, but to make these specific losses visible and open to deliberate compensation, rather than allowing them to disappear silently behind an aggregate efficiency gain. Related concerns arise wherever consequential decisions are delegated to autonomous systems with limited provision for reconsideration [13].

5.4 Conditions Under What Conditions Is Delegation Responsible?

Once delegation is considered appropriate, existing research provides substantial guidance on the conditions required to preserve agency and accountability.

These include meaningful rather than ceremonial human control; sufficient knowledge, authority, and time to intervene; contestability; independent verification in high-stakes settings; transparency regarding uncertainty and limitations; clear escalation mechanisms; responsibility allocation across designers, deployers, managers, operators, and institutions; and preservation of relevant human capabilities.

The objective is not maximum human intervention or maximum automation. It is an appropriate distribution of functions in which AI is used where it provides genuine benefits while human capability and authority are retained where consequences involve vulnerability, contested values, irreversible outcomes, or public legitimacy.

Responsible delegation therefore differs from merely retaining a nominal human reviewer.

5.5 Trajectory What Happens When This Choice Scales and Persists?

The fifth dimension moves the analysis from the organisation to society.

A single organisation may rationally conclude that automating a function increases efficiency while maintaining acceptable service quality. Yet many individually rational automation decisions can aggregate into social outcomes that no individual organisation intended.

Responsibility Culture adds a forward-looking question:

What social trajectory is being produced by repeated decisions of this kind?

The relevant effects may include changes in distributed expertise, expectations of human availability, opportunities for interpersonal contact, labour-market structure, professional formation, institutional discretion, public contestability, and the social distribution of responsibility.

Not all such effects are currently measurable, and many remain empirical research questions. Trajectory assessment should therefore not be treated as prediction of a deterministic future.

Rather, it introduces responsibility toward cumulative transformation.

It asks organisations and institutions to consider whether individually advantageous delegation decisions may collectively alter social capacities or institutions in ways that deserve explicit scrutiny.

6 A Procedure for Responsible Delegation

The framework becomes operational through a recorded comparison of configurations. The following procedure is a normative proposal for decision support, not an empirically calibrated instrument. Its output is a reasoned authorisation record, including conditions and unresolved disagreements, rather than a single delegability score.

6.1 Define the Decision and Compare Alternatives

The organisation first defines the function, affected groups, decision authority, geographical scope, and period of authorisation. A named assessment owner gathers evidence with domain professionals, system specialists, and representatives of affected workers or users. An accountable decision body, distinct from the supplier’s commercial interests where feasible, decides whether the configuration is justified. Participation requires accessible information and a record of whose views were absent.

The comparison should include the current arrangement, a feasible improved human arrangement, AI assistance, a hybrid division of functions, and fuller substitution. Comparing automation only with an under-resourced baseline can make substitution appear necessary by construction. Each alternative is assessed across all five dimensions, recording expected benefits, displaced capabilities, available evidence, uncertainty, and the distribution of consequences.

Evidence is labelled as observed in the deployment context, supported by relevant external studies, scenario-based, or unknown. These labels describe the basis of a claim rather than supplying a statistical confidence score. The team records disagreements about values separately from gaps in evidence: collecting more data cannot by itself resolve disagreement over the importance of human contact or discretion.

6.2 Screen Constraints Before Weighing Benefits

The procedure does not assign universal weights to Purpose, Function, Displacement, Conditions, and Trajectory. They are interdependent questions, not commensurable quantities. A high efficiency benefit must not automatically cancel an unacceptable loss of contestability. Before balancing benefits, the decision body identifies applicable legal and professional requirements and specifies minimum conditions for the responsible exercise of the function. Legal compliance is a necessary baseline, not a complete ethical justification.

A configuration fails the proposed responsibility screen when it leaves serious foreseeable harm without an effective response, assigns consequential accountability to actors without the authority or competence to intervene, or removes a socially necessary function without a credible alternative. In consequential decisions, inaccessible correction and review mechanisms are grounds for redesign or non-authorisation. These are normative screening principles; their concrete application requires domain-specific justification.

Displacement becomes unacceptable when the lost function is necessary to meet an identified protection or public purpose, the loss is serious for affected people, and no feasible compensating arrangement preserves that purpose. Severity, reversibility, vulnerability, distribution, and uncertainty must be considered explicitly. For example, fewer human interactions may be acceptable for routine information retrieval but unacceptable where this also removes the only effective route for an exceptional case to be heard. Human presence itself is not presumed sufficient or intrinsically superior.

6.3 Resolve Tradeoffs and Record the Decision

For alternatives that pass the screen, the decision body agrees context-specific priorities with affected participants before selecting an option. The rationale states which benefits justify which residual losses, who receives the benefits, who bears the losses, and why a less substitutive arrangement would or would not achieve the same legitimate purpose. A structured qualitative comparison is preferable to a numerical total unless locally validated measures and justified weights exist.

Sensitivity analysis asks whether the preferred option changes if a contested assumption is altered: for example, if available review time is lower than planned, or accessibility receives greater priority than throughput. If modest changes reverse the decision, the authorisation should remain limited, seek additional evidence, or select the more reversible configuration. Persistent disagreement about a major protection is escalated to independent or higher-level review rather than concealed in an averaged score.

The possible outcomes are authorisation, conditional pilot, redesign, deferral, or non-automation. The record identifies an accountable owner, safeguards, resources for implementation, indicators, review dates, and suspension triggers. A pilot is appropriate only when its potential harms can be contained; uncertainty about serious irreversible consequences cannot be resolved simply by exposing people to them. The ability to resume a human process must be resourced and tested, not merely promised.

6.4 Assess Trajectory Across Repeated Decisions

Trajectory assessment begins with an inventory of related delegation decisions rather than treating each deployment in isolation. The team maps which capabilities, escalation routes, learning opportunities, and providers are shared across services. It then develops contrasting scenarios: limited augmentation, progressive substitution, and widespread adoption across organisations. Scenarios identify dependencies and possible feedback loops; they do not predict a necessary future. Hila’s delegation dilemma provides a theoretical account of how individual strategies may generate collective lock-in [12].

Before deployment, the organisation records baseline indicators relevant to the function. These may include access to human review, correction times, independently assessed staff capability, training opportunities, complaint patterns, and dependence on a single provider. Reviews compare changes over time and across affected groups, examine alternative explanations, and document uncertainty. Where feasible, phased deployment or comparison groups can strengthen inference; descriptive trends alone do not establish that automation caused a change.

An organisation can observe only part of the societal trajectory. Sectoral associations, worker and user representatives, researchers, and public institutions may therefore need to aggregate appropriately protected evidence about common exclusion patterns, shrinking expertise, or shared infrastructure dependence. Organisational authorisation should include a route for reporting such signals. This is a proposed allocation of assessment responsibilities, not a claim that these arrangements already exist.

Review triggers are set before authorisation and linked to consequences. Failure to provide promised human review, repeated uncorrected exclusions, or inability to operate the fallback process requires immediate reassessment and possible suspension. Gradual deterioration in retained expertise prompts a capability review and adjustment of the division of work. Indicators and thresholds must be locally justified and revised as evidence improves; no universal numerical cutoff is asserted.

7 Worked Example of AI Assisted Recruitment

The following hypothetical example illustrates application of all five dimensions. It reports no field study, measured outcomes, or validated threshold. Consider an employer receiving large numbers of applications for administrative roles. It proposes AI ranking followed by automatic rejection below a cutoff. The alternatives are manual review, AI summaries with recruiter-led assessment, and ranking with automatic exclusion. The decision concerns initial screening, not appointment, and includes candidates with unconventional career histories.

7.1 Purpose and Function

Purpose. The stated objectives are to reduce processing delays and repetitive work while preserving fair access to consideration. The assessment asks for baseline processing times, recruiter workload, and the frequency of incomplete or incorrect screening. Cost savings are recorded alongside candidate interests. A faster process is beneficial only if its handling of relevant applications remains defensible; speed is not treated as a substitute for that assessment.

Function. Screening includes checking minimum requirements, interpreting evidence of transferable skills, recognising contextual explanations, and deciding when an ambiguous record merits further attention. Candidate representatives and recruiters identify which elements can be standardised and which require contextual interpretation. The unit of analysis is therefore the application-plus-context, rather than the production of a ranked list. AI may assist the extraction of information without being authorised to make an exclusion final.

7.2 Displacement and Conditions

Displacement. Automatic exclusion could save time but remove the opportunity to recognise relevant experience expressed in an unexpected form. It could also reduce junior recruiters’ exposure to varied applications. These are plausible effects to investigate, not demonstrated findings about the proposed tool. The assessment therefore compares errors across relevant candidate profiles, examines ambiguous applications independently of rankings, and tests whether staff can still evaluate cases without automated summaries.

Conditions. A recruiter must have access to the original application, sufficient review time, and authority to reject the recommendation. Candidates need a practicable route to correct factual errors and obtain human reconsideration. The employer remains responsible for the screening policy, while supplier responsibilities for system evidence and incident response are specified. Supplier assurances do not replace local assessment. Recruiter approval cannot serve as a safeguard if workload makes it a routine endorsement.

Under the proposed screen, automatic rejection is not authorised if affected candidates cannot obtain effective reconsideration or if the employer cannot establish an adequate basis for excluding atypical but eligible applications. A candidate’s failure to fit a standard profile does not itself demonstrate ineligibility. The same screen also applies to the human baseline: existing arbitrary exclusions must be addressed rather than treated as an acceptable standard.

7.3 Trajectory and the Resulting Decision

Trajectory. The employer considers what happens if the same screening logic becomes standard across its departments and if similar tools are adopted widely. Possible cumulative effects include shared blind spots, dependence on the same provider, and fewer opportunities for recruiters to develop contextual judgement. A single employer cannot establish a labour-market effect, but it can monitor its own contribution and communicate recurring problems through appropriate sectoral channels.

The assessment records baseline and follow-up measures of time to substantive human review, errors found in independently reviewed applications, review requests and corrections, and staff performance on unaided sample cases. Group comparisons are made only with an appropriate basis for handling the relevant data and an explicit account of their limitations. Changes in application volume, staffing, and role requirements are recorded as possible alternative explanations.

The illustrative decision is a conditional pilot of AI summaries with recruiter-led screening. Fuller substitution is deferred because the proposed automatic rejection process has not demonstrated that the displaced contextual function and correction route are preserved. This decision follows from the example’s assumptions; the framework does not imply that every recruitment system must receive the same result.

The authorisation record names a recruitment lead responsible for the pilot, reserves staff time for review, retains the original application as the evidential source, and provides an accessible correction channel. An independent sample is reviewed without AI summaries to identify overlooked information. The decision body reviews the pilot after a predefined recruitment cycle and before any extension to other departments.

If faster processing is achieved while effective review remains available, the pilot may continue with documented conditions. If promised reconsideration cannot be delivered, the employer suspends the automated screening step and resumes its maintained human process. If gains depend on removing the safeguards, the arrangement returns for a fresh decision rather than receiving automatic expansion. Evidence of broader recurring exclusion prompts escalation beyond the local team.

This example demonstrates the framework’s use without reducing its dimensions to a total score. Purpose favours improved throughput; Function and Displacement identify what must be preserved; Conditions limits admissible configurations; and Trajectory determines what must remain observable as the choice scales. The result is an explicit and revisable allocation of functions, not a general preference for either human or automated screening.

8 Discussion

The central distinction remains that technical substitutability does not imply social delegability. Technical capability establishes an available option; social delegability concerns its justification against realistic alternatives and consequences. The added procedure makes that justification inspectable by recording constraints, priorities, evidence, dissent, and conditions of revision.

The contribution is narrower than a new general theory of AI governance. Task-delegability research, Meaningful Human Control, impact assessment, and responsible innovation already address substantial parts of the problem. This framework connects them through a common decision object: changing the distribution of a human function. The five dimensions organise the assessment, while the screening and review procedure provides a proposed way to act on it.

Responsibility Culture contributes an emphasis on responsibility as a practice that must remain possible under the resulting arrangement. An individual cannot meaningfully assume a decision if organisational design removes the time, knowledge, or authority needed to question it. Conversely, keeping a nominal human approver does not establish that the underlying delegation is justified. Individual responsibility and organisational authorisation must be assessed together.

The trajectory dimension links this assessment to collective effects. Individually advantageous choices may alter the conditions under which later actors can choose, for example by making human alternatives costly to restore or narrowing opportunities to acquire expertise. Such pathways are hypotheses requiring investigation. Their inclusion makes cumulative change a continuing responsibility of governance rather than an assumed consequence of technical progress.

The public-good implication is that evaluation should consider the capabilities and relationships through which people can participate in decisions, alongside the efficiency of services. The appropriate outcome may be assistance, selective substitution, or fuller automation with effective safeguards. Non-automation remains legitimate when an identified necessary function cannot otherwise be maintained. None of these outcomes should become exempt from subsequent scrutiny.

9 Limitations and Research Agenda

The framework and procedure are conceptual and unvalidated. The worked example demonstrates internal application, not effectiveness, inter-rater reliability, stakeholder acceptance, or improved outcomes. The selected literature supports conceptual positioning rather than a comprehensive mapping of the field. Future work should test whether assessors interpret the dimensions consistently and whether the procedure changes decisions or reveals previously neglected consequences.

The proposed screening principles leave substantive judgement to contextual deliberation. They do not resolve all conflicts among rights, welfare, professional obligations, and resource constraints. Organisations may also manipulate baselines, narrow participation, or present safeguards that cannot be delivered. Independent scrutiny and transparent records can expose some failures but do not eliminate asymmetries of power.

Human functions and acceptable displacement cannot be assumed culturally universal. Comparative studies should examine how affected groups understand discretion, care, autonomy, and institutional responsibility without treating nationality as a sufficient proxy. Participation may reveal conflicting priorities that remain unresolved; these conflicts should be recorded as findings rather than treated as defective data.

Trajectory assessment faces attribution and coordination problems. Organisational indicators cannot establish society-wide effects, and sectoral evidence sharing may be limited by confidentiality and institutional capacity. Longitudinal studies, comparative deployments, and mixed methods are needed to distinguish transient adaptation from persistent changes in expertise, access, and dependence. Research should also test whether safeguards survive scale and workload pressure.

A feasible validation programme would compare assessments of the same cases by different teams, interview affected groups about omitted functions, and follow conditional deployments over time. It should evaluate decision quality, burden of participation, correction effectiveness, distributional outcomes, and the practical ability to reverse delegation. Universal weights or thresholds should not be introduced without evidence and an explicit normative justification.

10 Conclusion

AI-mediated societies are shaped through repeated human choices about the distribution of work, judgement, interaction, and responsibility. This paper proposes that these choices be assessed as decisions about social delegability rather than inferred from technical substitutability.

The five dimensions of Purpose, Function, Displacement, Conditions, and Trajectory connect the justification of delegation with the functions it changes and the consequences it may accumulate. A proposed decision procedure compares configurations, screens non-compensable losses, records contextual tradeoffs, and makes authorisation conditional on monitoring and revision. The recruitment example illustrates how this can support useful assistance while withholding unjustified automatic exclusion.

The proposal does not establish universal limits to automation. It makes the responsibility for deciding, preserving necessary capabilities, and revisiting consequences explicit. The capacity to automate does not remove the responsibility to decide whether, how, and under what conditions to do so.

Disclosure of Interests. The author declares that she has no competing financial or non-financial interests relevant to the content of this paper.

References

1. International Labour Organization: Algorithmic management in the workplace. https://www.ilo.org/algorithmic-management-workplace, last accessed 2026/10/18

2. Lyell, D., Coiera, E.: Automation bias and verification complexity: a systematic review. Journal of the American Medical Informatics Association 24(2), 423–431 (2017). doi: 10.1093/jamia/ocw105

3. Goddard, K., Roudsari, A., Wyatt, J.C.: Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association 19(1), 121–127 (2012). doi: 10.1136/amiajnl-2011-000089

4. Santoni de Sio, F., Mecacci, G.: Four Responsibility Gaps with Artificial Intelligence: Why they Matter and How to Address them. Philosophy & Technology 34, 1057–1084 (2021). doi: 10.1007/s13347-021-00450-x

5. Lubars, B., Tan, C.: Ask Not What AI Can Do, But What AI Should Do: Towards a Framework of Task Delegability. In: Advances in Neural Information Processing Systems, vol. 32 (2019).

6. International Organization for Standardization: ISO/IEC 42005:2025. Information technology — Artificial intelligence — AI system impact assessment. ISO, Geneva (2025).

7. De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A., Puntoni, S.: AI companions reduce loneliness. Journal of Consumer Research 52(6), 1126–1148 (2026). https://doi.org/10.1093/jcr/ucaf040

8. Veale, M., Van Kleek, M., Binns, R.: Fairness and Accountability Design Needs for Algorithmic Support in High-Stakes Public Sector Decision-Making. In: Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. ACM (2018). doi: 10.1145/3173574.3174014

9. Volz, K., Yang, E., Dudley, R., Lynch, E., Dropps, M., Dorneich, M.C.: An evaluation of cognitive skill degradation in information automation. Proceedings of the Human Factors and Ergonomics Society Annual Meeting (2016). https://doi.org/10.1177/1541931213601043

10. Bauer, K., Nofer, M., Abdel-Karim, B.M., Hinz, O.: The effects of discontinuing machine learning decision support. SAFE Working Paper No. 370 (2022). https://ssrn.com/abstract=4299664

11. Macnamara, B.N., et al.: Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers’ awareness? Cognitive Research: Principles and Implications 9, 46 (2024). https://doi.org/10.1186/s41235-024-00572-8

12. Hila, A.: The Human-AI Delegation Dilemma: Individual Strategies, Collective Equilibria and Sociotechnical Lock-in. arXiv preprint arXiv:2605.21351 (2026). https://arxiv.org/abs/2605.21351

13. Candrian, C., Scherer, A.: Rise of the machines: Delegating decisions to autonomous AI. Computers in Human Behavior 134, 107308 (2022). doi: 10.1016/j.chb.2022.107308

14. Vallor, S., Vierkant, T.: Find the gap: AI, responsible agency and vulnerability. Minds and Machines 34, 20 (2024). https://doi.org/10.1007/s11023-024-09674-0

15. Baiocco, S., Fernández-Macías, E., Rani, U., Pesole, A.: The algorithmic management of work and its implications in different contexts. JRC Working Papers Series on Labour, Education and Technology 2022/02. European Commission, Seville (2022). https://publications.jrc.ec.europa.eu/repository/handle/JRC129749

16. Santoni de Sio, F., van den Hoven, J.: Meaningful human control over autonomous systems: a philosophical account. Frontiers in Robotics and AI 5, 15 (2018). https://doi.org/10.3389/frobt.2018.00015

17. Stilgoe, J., Owen, R., Macnaghten, P.: Developing a framework for responsible innovation. Research Policy 42(9), 1568–1580 (2013). https://doi.org/10.1016/j.respol.2013.05.008

18. Reisman, D., Schultz, J., Crawford, K., Whittaker, M.: Algorithmic Impact Assessments: A Practical Framework for Public Agency Accountability. AI Now Institute (2018). https://ainowinstitute.org/publications/algorithmic-impact-assessments-report-2

19. Caloisi, I.: The Responsibility Culture Framework; Intercultural Model of Responsibility. RESC research programme (2026). https://resc-responsibilityculture.com/the-framework/; https://resc-responsibilityculture.com/the-model/, last accessed 2026/10/18

Resources

Recent