Generative AI and the Allocation of Administrative Burden in Public Employment Services for Older Jobseekers

1. Introduction

Public services can be formally available yet practically inaccessible when securing them requires knowledge, time and confidence that applicants cannot readily mobilize (Halling & Baekgaard, 2024; Moynihan et al., 2015). Consider a hypothetical older jobseeker who asks an official generative artificial intelligence (AI) assistant about vocational training. After tailoring its instructions to the applicant’s work history, the assistant advises them to verify the answer independently. Confirming that history and checking the institution’s interpretation of eligibility rules are different tasks. The reasonableness of participation depends on the purpose of the task and the information available to the person expected to perform it (Busuioc, 2021; Doughty & Baehler, 2020). When may a public institution reasonably ask someone seeking guidance to establish whether that guidance is correct?

Generative guidance sharpens this question because it assembles an individualized answer rather than presenting a fixed instruction. Retrieval-augmented generation can connect responses to external sources, but its retrieval and generation stages remain fallible (Zhao et al., 2026). Source attribution alone cannot establish that the relevant rule has been selected or fully explained for an individual case. Relying on applicants to detect these failures assumes that they have the necessary resources and expertise. Yet resources for navigating administration are unequal (Christensen et al., 2020), and information asymmetries complicate algorithmic oversight (Busuioc, 2021). Evidence that human-AI collaboration varies with task characteristics and relative capabilities makes the choice of verifier consequential (Vaccaro et al., 2024). A warning identifies a reason for caution; it does not determine who should make the answer dependable.

Administrative burden scholarship provides a starting point for evaluating this distribution of work. Its distinction between learning, compliance and psychological costs concerns how people experience interactions with the state (Moynihan et al., 2015). Digital self-service can simplify access while transferring professional casework to applicants (Madsen et al., 2022). Normative work distinguishes reasonable administrative demands from excessive ones (Doughty & Baehler, 2020), and recent research examines how generative AI reduces or intensifies burdens among street-level officials (Huang et al., 2026). The allocation of verification work to applicants requires a further judgment. We ask which tasks they may reasonably be expected to perform, given the purpose of the task, the parties’ capacities and the consequences of error.

Older jobseekers make the stakes of this allocation particularly clear. We use the term for people seeking employment support late in their working lives, including those returning after retirement. Meta-analytic evidence associates older age with slower and less successful reemployment after job loss (Wanberg et al., 2016), yet digital skills vary considerably among older US Internet users (Hargittai et al., 2019). Verification demands must therefore be assessed against applicants’ actual resources (Christensen et al., 2020). Research with older US food-assistance applicants also identifies financial, health and life crises that complicate administrative navigation (Negoita et al., 2024). These circumstances direct attention to the conditions under which guidance is sought. In the training example, an incorrect explanation of eligibility could discourage an application altogether. Responsibility for guidance therefore matters from the initial enquiry, before formal application and review procedures begin.

Here we develop a normative conceptual account of verification in guidance supplied by public employment agencies and their commissioned providers for jobseeker registration and vocational training applications. Private recruitment platforms, employer hiring decisions and automated eligibility adjudication fall outside this scope. Following a theory-adaptation approach (Jaakkola, 2020), we connect administrative reasonableness (Doughty & Baehler, 2020), unequal capacities to navigate public services (Christensen et al., 2020) and algorithmic accountability (Busuioc, 2021). We distinguish three tasks: confirming personal information, checking official interpretations and diagnosing system errors. Their allocation depends on purpose and necessity, relevant information and control, practical capacity and assistance, and the severity and reversibility of error. Our central argument is that reasonable applicant participation must be matched by institutional responsibilities for intelligible explanation, accessible assistance and effective correction. Establishing those responsibilities requires distinguishing the work assigned to applicants from the costs they experience.

2. Theoretical Foundations and Normative Approach

2.1 Administrative Tasks and Experienced Burden

A task is work that an organization expects an applicant to perform; burden concerns the costs experienced in meeting that expectation (Baekgaard & Tankink, 2022; Moynihan et al., 2015). Identical instructions can impose different costs because people differ in the cognitive and practical resources available to interpret and act on them (Christensen et al., 2020). Digital self-service makes this distinction visible: work once performed by a caseworker may become a condition of applicants completing the process themselves (Madsen et al., 2022). Our unit of normative assessment is therefore a task allocated within a service relationship. We ask what justifies assigning it to a particular applicant under particular conditions. Its likely costs inform that judgment, but their existence alone does not make the demand unreasonable.

Verification is an activity that can generate learning, compliance and psychological costs (Halling & Baekgaard, 2024; Moynihan et al., 2015). In the training example, locating a rule and understanding its relevance involve learning. Obtaining documentation to resolve a discrepancy involves compliance, while uncertainty about proceeding correctly may carry psychological costs. These difficulties can coexist despite a simplified digital interface (Madsen et al., 2022). Keeping the activity separate from its costs clarifies both what has been assigned and how it affects the applicant (Baekgaard & Tankink, 2022). Some effort serves legitimate administrative purposes (Doughty & Baehler, 2020). Evaluating verification therefore requires reasons for the demand, alongside an assessment of its costs.

2.2 Equal Access, Autonomy, and Public Accountability

We assess those reasons against three normative commitments: practical access to public services, autonomous participation and public accountability. Equal access means an effective opportunity to seek services for which one is eligible (Doughty & Baehler, 2020). This concerns access to the process, rather than a guarantee of favourable decisions or identical outcomes. Personal circumstances, available resources and digital skills shape the difficulty of navigating institutional arrangements (Christensen et al., 2020; Chudnovsky & Peeters, 2021; Hargittai et al., 2019). Formally uniform requirements must therefore be assessed against the resources needed to meet them. Requiring applicants to master the rules before they can obtain guidance may make access depend on expertise the service exists to supply.

Autonomy concerns applicants’ control over their participation, including their use of assistance. Self-determination theory distinguishes volition from independence, allowing support and autonomous action to coexist (Ryan & Deci, 2000). In AI ethics, people’s control over whether and how to delegate tasks is likewise central (Floridi et al., 2018). These distinctions support assistance chosen by the applicant and matched to the task and their skills (Hargittai et al., 2019). Transferring casework to citizens can increase their workload without increasing meaningful control (Madsen et al., 2022). Applicants should be able to choose support without surrendering decisions they can and wish to make themselves.

Public accountability makes responsibility for this support attributable and open to challenge. It requires an actor to explain and justify conduct before a forum able to question, assess and attach consequences to it (Bovens, 2007). In algorithmic settings, such scrutiny depends on expertise and authority as well as access to technical information (Busuioc, 2021). We apply this principle to institutional explanation of guidance, competent oversight and responses to error. These responsibilities must be coordinated with reasonable participation (Doughty & Baehler, 2020) and applicants’ control over delegation (Floridi et al., 2018). The normative task is to connect these commitments to specific work that applicants and providers can perform.

2.3 Theory Adaptation and Normative Argumentation

We combine conceptual analysis, theory adaptation and normative argumentation. Administrative burden scholarship defines applicant costs (Moynihan et al., 2015). Following Jaakkola’s (2020) account of theory adaptation, we use AI accountability to examine information asymmetries and institutional control (Busuioc, 2021), and technical reliability research to identify retrieval and interpretation failures (Zhao et al., 2026). The adaptation introduces task-specific responsibility into the justification of burden allocation. We select peer-reviewed literature purposively for conceptual clarification, normative reasoning and evidence about relevant technical and practical conditions. Empirical findings inform assumptions about users and systems; judgments about who ought to perform a task also require explicit normative premises.

Hypothetical contrasts, counterexamples and objections test the consistency of allocation judgments (Jaakkola, 2020). We compare personal and institution-controlled information, then vary applicants’ resources to examine practical capacity (Christensen et al., 2020). A capable applicant’s preference for independent use tests whether assistance preserves autonomy (Ryan & Deci, 2000). Resource constraints and legitimate cooperation demands test institutional feasibility (Doughty & Baehler, 2020). We also vary error consequences and correction opportunities. All service scenarios are hypothetical. The analysis uses existing studies rather than original data collection or model testing. Its conclusions concern justified arrangements; empirical effects and jurisdiction-specific legal duties require separate assessment. We begin by distinguishing the work each verification task requires.

3. Verification Tasks in AI-Assisted Public Employment Services

3.1 Personal Information, Official Guidance, and System Errors

Confirming personal information asks an applicant to check facts such as their contact details or work history. The justification depends on what that check requires (Doughty & Baehler, 2020). In the training example, confirming a current telephone number draws on knowledge the applicant may possess more readily than the agency. Reconstructing years of intermittent employment from missing records requires more work. Cognitive resources and the demands of digitally transferred casework affect the difficulty of both tasks (Christensen et al., 2020; Madsen et al., 2022). Both concern personal facts, but the latter can entail substantial learning and compliance costs (Moynihan et al., 2015). Informational advantage supports participation only within the limits of what the applicant can reasonably establish.

Verifying official guidance changes the object of scrutiny from personal facts to institutional interpretations. Eligibility conditions, document requirements and application deadlines depend on rules whose meaning may not be apparent from everyday experience. Administrative literacy concerns understanding and navigating such demands (Döring & Madsen, 2022). Transferring this work can require applicants to acquire knowledge previously supplied by professionals (Madsen et al., 2022). Comparing a stated date with a clearly applicable notice may be straightforward. Establishing whether an exception applies or a document list omits a requirement calls for interpretive judgment. Thus, confirming when a person worked differs from determining whether that work satisfies a programme’s eligibility definition. Accountability for the latter rests on the capacity and authority to clarify the governing rules (Bovens, 2007; Busuioc, 2021).

Diagnosing system errors concerns how an answer was produced. A knowledge base may omit a relevant rule, retrieval may select an unsuitable passage, or generation may distort the material retrieved (Ji et al., 2023; Zhao et al., 2026). The same incorrect deadline could result from any of these failures. Noticing the discrepancy would not identify its cause. Information and correction powers may also be divided: the agency may maintain rule content while a contractor controls retrieval settings or software updates. Algorithmic accountability scholarship highlights these asymmetries (Busuioc, 2021), while meaningful scrutiny requires the capacity to question explanations and assess conduct (Bovens, 2007). Reporting a suspect answer, tracing its origin and correcting the underlying process are therefore separate contributions. An applicant may make the first without the access, expertise or authority needed for the others.

3.2 Personalized Generation and the Limits of Source Grounding

Personalized generation can combine all three tasks within a single answer. A retrieval-augmented assistant can use an applicant’s question to obtain external material and base its response on the retrieved passages (Zhao et al., 2026). For the training enquiry, it might organize instructions around work history rather than reproduce a general notice. Yet fluent generation can remain inconsistent with its source material (Ji et al., 2023). In controlled question-answering and retrieval tasks, the models studied by Liu et al. (2024) used relevant information less reliably when it appeared in the middle of long inputs. Access to information and effective use of it are thus separate technical conditions. Combined with information asymmetries between users and providers (Busuioc, 2021), this distinction explains why inspecting the supplied materials may leave the accuracy of an individualized answer unresolved.

Source grounding therefore raises four questions: whether an answer has a genuine source, whether that source applies, whether the explanation is complete, and whether the individualized advice is correct. Research on hallucination and retrieval-augmented generation distinguishes source fidelity from answer quality (Ji et al., 2023; Zhao et al., 2026). Findings on context use further show that relevant material can be available yet inadequately incorporated (Liu et al., 2024). Suppose the assistant cites an authentic training notice but overlooks an exception relevant to intermittent employment. Opening the link could establish the notice’s authenticity while leaving the exception and its implications unexplained. Recognizing the omission requires understanding the programme’s rules, bringing administrative literacy into the verification task (Döring & Madsen, 2022). The unresolved question is whether the advice correctly applies those rules to the applicant.

Official knowledge bases, source links and risk warnings address different parts of this problem. A curated knowledge base constrains the materials available to retrieval; the appropriate use of those materials still depends on retrieval and generation (Zhao et al., 2026). A link enables inspection, while an explanation must establish why the rule applies (Busuioc, 2021). A general warning communicates possible fallibility but leaves the doubtful proposition and the work needed to check it unspecified. The reasonableness of assigning that work depends on what remains unresolved and who can resolve it (Doughty & Baehler, 2020). Review must accordingly match the uncertainty with the checker’s capabilities (Vaccaro et al., 2024). This makes applicants’ actual verification conditions central to the allocation question.

3.3 Older Jobseekers and Unequal Conditions of Verification

Those conditions include digital literacy, understanding of administrative rules, available help and time to act. Among older US Internet users, digital skills vary with socioeconomic circumstances and autonomy of use (Hargittai et al., 2019). Administrative literacy concerns the separate ability to understand institutional demands (Döring & Madsen, 2022). An applicant might navigate a chat interface confidently yet lack the knowledge needed to question its interpretation of eligibility. Cognitive-resource research and accounts of unequal burden distribution locate such difficulties in the interaction between demands and personal circumstances (Christensen et al., 2020; Chudnovsky & Peeters, 2021). Age alone therefore provides an inadequate basis for assessing whether a particular verification task is feasible.

Help is useful only when it addresses the difficulty in time. A friend who can navigate the portal may be unable to resolve a disputed eligibility rule. An authoritative response may resolve that dispute but arrive too late to inform an application. Research on administrative literacy and social capital distinguishes knowledge of administration from the social resources available to navigate it (Döring & Madsen, 2022). Negoita et al. (2024), examining older adults’ experiences of the US Supplemental Nutrition Assistance Program (SNAP), identify how financial, health and other life crises complicate information seeking, recall and planning. These findings identify dimensions of difficulty relevant to assessing assistance. Cognitive resources and unequal access to support also matter to whether help can resolve a specific uncertainty within the time available (Christensen et al., 2020; Chudnovsky & Peeters, 2021).

Unresolved uncertainty can also prevent an application from being submitted. Suppose the jobseeker accepts an incorrect explanation that their employment history excludes them from training and never applies. A review confined to filed applications would miss this route to exclusion. Administrative exclusion research examines non-participation arising from the claiming process (Brodkin & Majmundar, 2010), while learning costs include discovering whether and how to seek a service (Moynihan et al., 2015). The applicant in this example would need the knowledge sought through official guidance to recognize that the guidance was wrong. Institutional answerability and administrative reasonableness therefore extend to the initial service encounter (Busuioc, 2021; Doughty & Baehler, 2020). Justifying a verification demand requires attention to the work assigned, the conditions for completing it and the opportunity at risk.

4. Justifying the Allocation of Administrative Burden

4.1 Legitimate Purpose and Necessity

The first question is whether a check serves a clear and legitimate administrative purpose. Reasonableness requires examining both the function of a demand and when it becomes excessive (Doughty & Baehler, 2020). Confirming employment dates may supply facts needed to identify a relevant training programme; checking whether a registration already exists may prevent duplicate processing. Asking the applicant to discover why the assistant used an obsolete rule instead recruits them into maintaining the service. Each purpose can justify seeking reliable information, but assigning the work to the applicant requires further reasons. Learning and compliance requirements have costs beyond institutional processing time (Moynihan et al., 2015), and digital self-service can relocate professional work without removing it (Madsen et al., 2022). Algorithmic accountability also depends on who can explain and scrutinize system operation (Busuioc, 2021). The institution should therefore identify the uncertainty, its relevance to the service and the contribution sought from the applicant. Only then can the need for that contribution be assessed.

Necessity requires comparison with practicable arrangements that could achieve the same purpose with less applicant effort (Doughty & Baehler, 2020). Suppose the training agency can lawfully consult a reliable employment record already available to it. Asking the applicant to confirm a displayed discrepancy may be easier than requiring a fresh reconstruction of the entire record. Where uncertainty concerns an exception in the training rules, an authoritative explanation may address it more directly than repeated requests for documents. Digital arrangements can support such alternatives through information provision and data transfers (Madsen et al., 2022). The comparison must account for unequal resources (Christensen et al., 2020). In US disability programmes, increased application costs following local office closures reduced benefit receipt, particularly among some less-educated applicants (Deshpande & Li, 2019). Saving staff time is therefore only part of the assessment. The institution must explain why a less demanding alternative is impracticable or materially less effective, and why the remaining effort is proportionate to its purpose.

4.2 Information, Control, and Responsibility

Once a check is necessary, information and control help determine who should perform it. Knowing that an answer is wrong differs from identifying its cause or authorizing its correction. These capacities depend on access, expertise and authority (Busuioc, 2021). A source link alone supplies little insight into the system that selected and interpreted the source (Ananny & Crawford, 2018). In the training example, the applicant knows their employment dates, the agency interprets the programme rules, and the technical provider can inspect retrieval or update a defective component. Different stages of retrieval-augmented generation can require different corrective actions (Zhao et al., 2026). Drawing on prospective responsibility for preventing and addressing failures (Santoni de Sio & Mecacci, 2021), we propose a rebuttable presumption: the party with greater relevant knowledge and control should take primary responsibility for the corresponding quality assurance and correction. This presumption assigns identifiable tasks to those equipped to perform them. Transferring those tasks to an applicant requires additional reasons.

Responsibility under this presumption follows actual spheres of control. Technical opacity may limit a developer’s explanation of an output, and organizational arrangements may divide relevant powers among several actors (Busuioc, 2021; Santoni de Sio & Mecacci, 2021). Greater control supports responsibility for investigation, limits on reliance and correction within that sphere. Applicants can in turn clarify facts that institutional records cannot establish. Such requests must account for the resources needed to retrieve missing documents or resolve inconsistencies (Christensen et al., 2020). Exceptional competence also has limits as a reason for allocation. A jobseeker who understands software may voluntarily identify a retrieval failure, but that skill does not make diagnostic work a necessary condition of assistance (Doughty & Baehler, 2020). The contribution must remain connected to obtaining the service. Where institutional control is divided, providers need to coordinate their investigation rather than assign their internal coordination problem to the applicant.

4.3 Practical Capacity and Accessible Assistance

An allocation based on informational advantage must also be feasible in practice. The assessment concerns the time, skills and assistance available for the particular task. Task requirements must be assessed against both digital skills and administrative literacy (Döring & Madsen, 2022; Hargittai et al., 2019). Personal circumstances also affect whether skills can be used to meet a requirement (Christensen et al., 2020; Chudnovsky & Peeters, 2021). We therefore ask whether the applicant can obtain the relevant information, understand the requested comparison and communicate a response within the available period. This tests the feasibility of the provisional allocation. Publishing a source is insufficient where understanding it requires specialist interpretation. Likewise, a task that could be completed with more time may be unreasonable near a deadline. In either case, the institution must change the demand or provide the support needed to meet it.

Assistance can make a task feasible when it supplies the missing capability at the point of need. In a randomized field experiment among older US adults likely eligible for SNAP, information combined with application assistance produced greater enrolment than information alone (Finkelstein & Notowidigdo, 2019). The study distinguishes disclosure from practical help in that setting. Administrative literacy and social resources also make different contributions to navigating public programmes (Döring & Madsen, 2022). Assessing assistance therefore requires asking whether applicants can reach someone able to resolve the uncertainty in time. The learning and compliance costs of obtaining help are part of this assessment (Moynihan et al., 2015). Additional registration or a demand to prove difficulty may reproduce the obstacle assistance is intended to address. Support must also fit the applicant’s skills (Hargittai et al., 2019). Assistance changes the feasibility judgment only when the applicant can use it to complete the task.

4.4 Consequences of Error and Opportunities for Correction

Even a feasible task may be unreasonable if an overlooked error has serious, irreversible consequences. Administrative exclusion research examines opportunities lost through the claiming process, including losses outside formal eligibility decisions (Brodkin & Majmundar, 2010). Evidence on application costs likewise connects procedural conditions with actual benefit receipt (Deshpande & Li, 2019). Consider four possible consequences of inaccurate guidance: a minor informational discrepancy, unnecessary preparation of documents, a mistaken assessment of eligibility, or a missed application deadline. An erroneous explanation of ineligibility may prevent the process from starting. Severity also depends on applicants’ circumstances (Christensen et al., 2020). Preparing another document may be manageable for one person but consume the time another needs to apply. We therefore extend the reasonableness inquiry to the severity and reversibility of error (Doughty & Baehler, 2020). As consequences become more serious and harder to reverse, reliance on applicant checking requires stronger justification. Correction must address the opportunity at stake: replacing an answer after a training intake closes does not restore the chance to apply.

This requires examining the full route from recognizing an error to obtaining timely correction. Warnings, review and corrective action have distinct roles along that route. Information enables scrutiny only under suitable conditions (Ananny & Crawford, 2018), and reviewers’ judgments also need examination. Dutch experiments found no greater adherence to algorithmic advice than to equivalent human advice. One experiment nevertheless found preferential adherence to stereotype-consistent advice from either source (Alon-Barkat & Busuioc, 2023). A wider meta-analysis identifies substantial variation in human-AI performance (Vaccaro et al., 2024). Review therefore depends on the reviewer’s competence, information and authority. Accountability also requires institutional responses that can have consequences (Bovens, 2007). We assess the combined protection offered by these arrangements, including gaps between a warning, access to review and effective correction. Prospective responsibility extends to preventing avoidable harm (Santoni de Sio & Mecacci, 2021). Where timely recovery is doubtful, competent verification before consequential reliance becomes especially important.

4.5 Conditional Allocation of Verification Tasks

These criteria produce different allocations for the three verification tasks. Applicants can reasonably confirm readily known personal information when the purpose is clear, the effort modest and mistakes easy to correct. This retains legitimate cooperation (Doughty & Baehler, 2020) while accounting for the resources needed to complete it (Christensen et al., 2020). Complex rule application requires a more specific division of work. The applicant may confirm relevant events while the institution explains their administrative significance, with assistance addressing any remaining difficulty. Administrative literacy helps distinguish these contributions (Döring & Madsen, 2022). Where checking depends on institution-held rules, internal system information or correction powers, primary responsibility rests with the relevant provider (Busuioc, 2021). Serious and difficult-to-reverse consequences strengthen that allocation. Applicants may still supply information or report discrepancies, but shared responsibility must specify who resolves each uncertainty. It cannot leave both parties nominally responsible for checking the whole answer.

The criteria form a sequence of judgments rather than a numerical score. Purpose and necessity identify the demand to be justified. Information and control establish a provisional allocation, which practical capacity and opportunities for correction then test. This sequence connects administrative reasonableness (Doughty & Baehler, 2020) with task-specific responsibility (Santoni de Sio & Mecacci, 2021). Each condition has independent force: a legitimate purpose cannot make an impossible task feasible, and personal competence cannot confer institutional authority. Because burden depends on circumstances, the same task may warrant different assistance as a deadline approaches or support becomes unavailable (Chudnovsky & Peeters, 2021). The framework makes disagreements open to explanation and challenge through identifiable reasons and alternatives (Bovens, 2007). Its application therefore depends on how the service organizes explanation, assistance and correction. These arrangements are part of the justification for assigning work, not additions made after the allocation has been settled.

5. Assistance and Correction in Public Employment Services

5.1 Reliable Guidance and Intelligible Explanations

Quality assurance provides the institutional basis for reasonable applicant verification. Retrieval depends on accessible source materials, while generation can misrepresent even relevant information (Ji et al., 2023; Zhao et al., 2026). Agencies should therefore assign responsibility for maintaining programme rules, recording effective dates and withdrawing superseded instructions. This makes responsibility for the guidance identifiable (Busuioc, 2021). Explanations should connect the relevant rule to the circumstances supplied, distinguishing confirmed facts from assumptions and unresolved questions. Their adequacy depends on how applicants understand and evaluate them (Miller, 2019). In the training example, this means identifying the programme, intake and treatment of intermittent employment. The explanation should make clear how those details support the advice.

That explanation should also state what the applicant needs to do. It should distinguish facts to be confirmed from questions of rule application requiring clarification (Döring & Madsen, 2022). A reasonable request identifies the particular information to check and allows straightforward correction; an unresolved exception calls for competent interpretation (Doughty & Baehler, 2020). Source links can support this exchange by making the basis of the explanation available for inspection (Ananny & Crawford, 2018). Warnings should identify unresolved matters and a route to clarification. This defines the work still needed, instead of treating acceptance of a disclaimer as acceptance of all verification tasks. The distinction between prospective responsibility and liability clarifies the purpose of this arrangement: to assign responsibility for resolving uncertainty before applicants act on the guidance (Santoni de Sio & Mecacci, 2021).

5.2 Human Assistance and Meaningful Service Choice

Unresolved questions need an effective route to human assistance. Contestability depends on organizational capacity and authority to intervene (Alfrink et al., 2023). Referral should follow unsubstantiated answers, disputed rule applicability or difficulty completing a required step, with response times tied to the remaining application period. Transferring the enquiry and unresolved issue to staff would avoid requiring applicants to reconstruct digital casework (Madsen et al., 2022). Staff need authoritative rules and powers to correct guidance or obtain clarification (Busuioc, 2021). They also need to reassess the disputed answer independently, as the value of human-AI collaboration depends on the capabilities brought to the task (Vaccaro et al., 2024). Referral is useful when it enables resolution, not simply when it adds another contact.

Applicants should retain control over how they use this assistance. Autonomy permits chosen support (Ryan & Deci, 2000), and varied digital skills among older adults call for more than one route through the service (Hargittai et al., 2019). Applicants may continue independently, seek help with a particular step or move to an assisted channel. Support should follow the difficulty encountered and the resources available to address it (Christensen et al., 2020). Additional registration, age-based identity checks or proof of digital disadvantage would create demands requiring their own justification (Doughty & Baehler, 2020). Meaningful service choice therefore requires an accessible alternative when self-service becomes difficult, while keeping independent use available to those who prefer it.

5.3 Correction and the Preservation of Application Opportunities

Assistance addresses uncertainty; correction addresses errors and their consequences for the application. Administrative exclusion can arise during the claiming process (Brodkin & Majmundar, 2010), and learning costs arise before a formal application exists (Moynihan et al., 2015). Corrective review should therefore be available from the initial enquiry, without requiring an application number. In the training case, an applicant who prepared the wrong documents may need a corrected list, time to supply missing material and renewed verification of the relevant facts. Someone discouraged from applying needs clarification of any remaining route to participation. These responses give practical effect to contestability (Alfrink et al., 2023). Accountability further requires an explanation of what was investigated, corrected and left unresolved (Bovens, 2007). The response should specify next steps and any limits to restoring the application opportunity.

Initiating review should require an account proportionate to the applicant’s knowledge (Busuioc, 2021; Doughty & Baehler, 2020). The applicant might describe the enquiry, approximate date, advice recalled and action taken. The institution should examine the retained answer, applicable rule version and subsequent exchanges. Where no screenshot is available, it should help reconstruct the encounter and assess the evidence that can be recovered. This supports reasoned investigation (Bovens, 2007) while keeping correction distinct from compensation (Santoni de Sio & Mecacci, 2021). Managerial responsibility concerns investigating and correcting guidance. Extending a deadline or reopening an application requires the relevant procedural authority, while compensation raises a separate question of liability. The governing jurisdiction and procedure determine which remedies are available. Review should explain both the corrective action taken and the basis for any further remedy.

5.4 Organizational Responsibility and Technical Provision

Delivering that response requires clear organizational responsibility. Accountability depends on attributable conduct and answerability (Bovens, 2007), whereas algorithmic systems can disperse information and expertise across organizations (Busuioc, 2021). The public agency should maintain authoritative rules, organize assistance and explain the service’s response. Technical providers should investigate retrieval, generation, updating and operational failures within their actual spheres of control (Zhao et al., 2026). Commissioning arrangements should specify access to evidence, responsibility for investigation and authority to implement corrections. This assigns responsibilities before failures occur (Santoni de Sio & Mecacci, 2021). The agency and provider may need to establish internally which component failed. Applicants should have a clear point of contact throughout that investigation, rather than having to identify the responsible provider before obtaining help.

Record retention should support this coordination while limiting personal data collection. Privacy engineering distinguishes controls over data use from architectures that reduce identifiable data collection (Spiekermann & Cranor, 2009). Combined with contestability’s need for usable documentation (Alfrink et al., 2023), this supports retaining the answer, timing, relevant source versions and only the personal context necessary to investigate. Identifiers should be separated from technical records where feasible, access limited by role and retention periods tied to review needs. Unneeded records should be deleted or appropriately de-identified, subject to applicable requirements. Effective scrutiny depends on usable information rather than archive size (Ananny & Crawford, 2018). Requests for extensive new employment histories merely to report a suspected error would add burdens requiring justification (Doughty & Baehler, 2020).

6. Discussion

6.1 Resource Constraints and Reasonable Citizen Participation

These arrangements raise a practical question: how much verification can an agency sustain without delaying access to assistance? Administrative burdens arise within established organizational structures and practices (Peeters, 2020), and reasonableness requires attention to practicable alternatives (Doughty & Baehler, 2020). Continuing quality management and review of consequential uncertainty can be distinguished from manual inspection of every exchange. The former organize the capacity to scrutinize and respond to system operation (Busuioc, 2021). The latter is neither necessary for every routine enquiry nor an assurance of correctness, given variation in human-AI performance (Vaccaro et al., 2024). Where reliable individualized interpretation exceeds available capacity, the agency may need to narrow the guidance offered and arrange authoritative clarification. Resource constraints shape the level and organization of safeguards. They do not supply the expertise applicants would need to resolve the institution’s uncertainties.

This approach still leaves room for reasonable citizen participation. Applicants can supply facts unavailable to the agency and confirm readily understood details when those contributions serve the application (Doughty & Baehler, 2020). The resource assessment must nevertheless include work transferred through digital self-service (Madsen et al., 2022). Staff time saved may be accompanied by learning, compliance or psychological costs for applicants (Moynihan et al., 2015), with different effects according to their resources (Christensen et al., 2020). Efficiency should therefore be assessed across the service encounter, including the effort required to correct avoidable misunderstandings. A resource-based defence of applicant checking must identify whose work is saved, whose work is increased and why the resulting allocation remains reasonable.

6.2 Autonomy and Age-Sensitive Assistance

Assistance also raises a concern about paternalism: support may presume dependence that recipients reject. In a study of a health-robot prototype, older test users resisted identification with an imagined dependent user (Neven, 2010). Support must accommodate recipients whose skills differ from those assumed in the design (Hargittai et al., 2019). Assistance should therefore preserve control over participation and delegation (Ryan & Deci, 2000; Floridi et al., 2018). A request for clarification of one rule should leave the applicant in control of the remaining application. Equally, a preference for independent use should leave institutional responsibilities intact. Choice concerns how the applicant participates, not who controls the rules or the system that explains them.

The test of a needs-based service is how easily applicants can obtain help with a specific obstacle. Administrative literacy and social resources shape navigation (Döring & Madsen, 2022), while cognitive-resource research locates difficulty in the interaction between demands and circumstances (Christensen et al., 2020). Requiring proof of general incapacity would shift attention from resolving the obstacle to classifying the applicant. That additional demand would itself need justification (Doughty & Baehler, 2020). A simple request or an unresolved service issue should be sufficient to seek assistance. This reasoning also covers younger applicants facing unfamiliar rules or unavailable support, since unequal burdens extend across age groups (Chudnovsky & Peeters, 2021). Older jobseekers remain the focal context, but the grounds for support are task-specific need and the applicant’s chosen participation.

6.3 Theoretical Contribution and Scope Conditions

The resulting account extends the analysis of how administrative work is allocated. Existing research locates burdens in organizational arrangements (Peeters, 2020) and documents transfers of casework through digital self-service (Madsen et al., 2022). Human advice and static instructions can also contain errors that require verification. Generative guidance combines an enquiry, retrieved materials and generated interpretation into an individualized explanation (Zhao et al., 2026), while information and control may be divided across providers (Busuioc, 2021). Distinguishing personal confirmation, interpretive verification and technical diagnosis reveals who can resolve each uncertainty. The contribution is to connect administrative reasonableness with task type, control and timely correction (Doughty & Baehler, 2020). This evaluates responsibility for the work that generates burden while keeping assigned demands distinct from experienced costs (Baekgaard & Tankink, 2022).

Applying this account requires attention to organizational capacity, the service sought and the governing rules. Theory adaptation supplies reasons for evaluating those arrangements (Jaakkola, 2020); empirical evaluation must examine their effects on burden, reemployment and fairness. Variation in applicants’ resources calls for assessing experiences across settings and groups (Chudnovsky & Peeters, 2021). Evaluation should also include encounters that never become applications, given the possibility of organizationally induced non-participation (Brodkin & Majmundar, 2010). Implementation requires matching prospective responsibilities to the institution’s actual powers to investigate, intervene and correct (Santoni de Sio & Mecacci, 2021). The central question is whether the proposed arrangements let applicants make a reasonable contribution while preserving an effective opportunity to proceed.

7. Conclusion

Verification in AI-assisted public employment services should be allocated according to task purpose and necessity, relevant information and control, practical capacity and available assistance, and the consequences and reversibility of error. These conditions distinguish reasonable participation from the transfer of institutional work to applicants. Confirming readily known personal facts can remain an applicant task when the effort is proportionate and mistakes are easy to correct. Checking official rule interpretations and diagnosing system failures depend on different information, expertise and powers. Primary responsibility for that work should remain with the actors who hold them. Where responsibilities are shared, the service must identify the uncertainty each party is expected to resolve. A general warning about AI fallibility leaves that allocation unexplained.

For older jobseekers, the same standards require attention to actual resources and support needs. Assistance should enable applicants to complete particular tasks while preserving their choice to proceed independently or seek help. That choice concerns participation; it does not transfer control over official interpretations or system operation. Protection must also extend to initial enquiries, where incorrect guidance can prevent an application from being made. Correction should address the resulting obstacle and the opportunity to proceed, as well as the inaccurate answer. Applicants should retain control over their participation, and providers should remain answerable for intelligible guidance, accessible assistance and timely correction within their respective spheres of responsibility. The governing principle is a clear division of work that applicants can reasonably perform and institutions can explain and discharge.