INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 361
Human-in-the-Loop AI: Rethinking Automation Ethics in Decision-
Sensitive Domains Case Study of the Education, IT and Non-for-
Profit sectors.
*Nankyer Sarah Joseph, Mohammed Nasiru Yakubu
Department of Information Systems, School of IT and Computing, American University of Nigeria, Yola
Mohammed Nasiru Yakubu Arden University Middlemarch Business Park, Coventry. CV3 4FJ, Nigeria
DOI: https://doi.org/10.51584/IJRIAS.2025.1010000027
Received: 23 Sep 2025; Accepted: 30 Sep 2025; Published: 30 October 2025
ABSTRACT
This study develops and applies the Human-in-the-Loop (HITL) Ethical Assessment Framework (EHAF) to
examine the ethical sufficiency of HITL artificial intelligence (AI) across education, information technology
(IT), and non-profit sectors. The research objective was to evaluate how effectively HITL practices safeguard
human values in decision-sensitive contexts and to identify sector-specific challenges that may compromise
ethical adequacy. Adopting a qualitative thematic approach, we analyzed survey responses from professionals
in the three sectors. Responses were coded against the four diagnostic dimensions of EHAF Impact Severity,
Contextual Ambiguity, Human Agency, and Transparency & Auditing while also allowing for the identification
of emergent themes. Retroductive reasoning was used to move beyond surface patterns to uncover generative
mechanisms shaping HITL practices. Findings demonstrate sectoral variation in how HITL systems are
operationalized and valued. In education, ethical sufficiency is closely tied to human oversight given the high
stakes of student outcomes and the importance of cultural contextualization. In the non-profit sector,
transparency and auditing dominate due to donor accountability pressures and reporting requirements. IT
organizations, by contrast, privilege efficiency and scalability, but often provide weaker safeguards for human
agency and oversight. Across all sectors, emergent themes such as training, trust, infrastructure readiness, and
donor influence were found to condition HITL adequacy. Generative mechanisms identified include
institutional role ambiguity, donor pressure, cultural misalignment, and capacity constraints. The study
concludes by proposing an extension to EHAF that incorporates a fifth dimension, Capacity and Governance
Context to better capture systemic and institutional influences. Conceptually, the paper refines the assessment
of HITL ethics, while practically offering sector-specific recommendations to strengthen oversight,
accountability, and trust in AI-enabled decision-making.
Keywords: Human-in-the-Loop, AI Ethics, Automation, Decision-Sensitive Domains, Responsible AI,
INTRODUCTION
Artificial intelligence (AI) is increasingly shaping decision-making across critical domains such as education,
information technology, and the non-profit sector. While automation promises efficiency and scalability, the
ethical risks of delegating sensitive decisions entirely to algorithms have become more apparent [1]. Concerns
around bias, opacity, and accountability gaps are particularly acute in contexts where outcomes directly affect
human welfare and equity [2] These risks highlight the pressing need to balance automation with human
oversight.
Within the information systems (IS) discipline, scholars emphasize that AI systems are sociotechnical rather
than purely technical artifacts, and thus require governance frameworks that account for both human and
machine roles in decision-making [3]. One promising approach is Human-in-the-Loop (HITL), where human
judgment is embedded within algorithmic processes to provide contextual reasoning, safeguard against
harmful outcomes, and reinforce accountability (Rahwan et al., 2019).
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 362
However, current AI governance models often adopt a techno-centric perspective that overlooks
organizational, cultural, and ethical complexities [4]. While automation is framed as a means of minimizing
human error, little attention has been paid to how HITL practices can reshape ethical responsibility and trust in
decision-sensitive contexts. Moreover, there is limited qualitative empirical research exploring HITL
implementation in diverse organizational domains, particularly within education, IT services, and the non-
profit sector, where decision stakes are high but resources and institutional safeguards may vary [5].
This paper addresses these gaps by examining how HITL approaches can enable more responsible AI in
decision-sensitive domains. Drawing on a qualitative case study design, with data collected from organizations
in IT, education, and the non-profit sector, we investigate how human oversight interacts with automated
decision-making in practice. The study seeks to answer the guiding question:
How can Human-in-the-Loop AI practices enhance ethical responsibility and trust in decision-sensitive
domains?
The contribution of this paper is threefold. First, it extends IS research on sociotechnical governance of AI by
theorizing HITL as an ethical safeguard. Second, it offers empirical insights into the opportunities and
challenges of HITL implementation across multiple domains. Finally, it provides policy and organizational
recommendations for embedding human oversight into AI systems, thereby contributing to ongoing debates on
responsible AI governance.
LITERATURE REVIEW AND THEORETICAL FRAMING
The integration of Artificial Intelligence (AI) in decision-sensitive fields such as information technology (IT),
education, and the non-profit sector has garnered increasing attention, primarily concerning the ethical
implications of Human-in-the-Loop (HITL) systems. HITL approaches, which incorporate human judgment
into AI decision-making processes, are pivotal in ensuring that AI systems operate within ethical boundaries
and align with societal values.
Benefits of HITL Systems
In the education sector, HITL systems facilitate personalized learning experiences by adapting to individual
student needs and providing real-time feedback. This adaptability enhances student engagement and supports
diverse learning styles [6]. Similarly, in the non-profit sector, AI tools enable organizations to streamline
operations, improve donor engagement, and optimize resource allocation, thereby increasing operational
efficiency.
Furthermore, HITL systems contribute to the ethical deployment of AI by allowing human oversight to correct
biases, ensure fairness, and maintain accountability in decision-making processes. This human oversight is
crucial in high-stakes applications where AI decisions can significantly impact individuals' lives.
Challenges Associated with HITL Systems
Despite their advantages, HITL systems present several challenges. In education, the reliance on AI tools may
lead to concerns about data privacy, algorithmic bias, and the potential erosion of teacher-student relationships
[7]. Additionally, the effectiveness of HITL systems is contingent upon the quality of human input; inadequate
or biased human judgment can perpetuate existing inequalities in educational outcomes.
In the non-profit sector, the adoption of AI technologies can be hindered by limited financial resources, lack of
technical expertise, and resistance to change within organizations. Moreover, the ethical implications of using
AI in non-profit settings, such as the potential for exploitation of vulnerable populations and the need for
transparency, require careful consideration.
Applications of HITL Systems in Education and Non-Profit Sectors
In education, HITL systems are applied in adaptive learning platforms, automated grading systems, and virtual
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 363
teaching assistants. These applications aim to enhance learning outcomes by providing personalized support
and reducing administrative burdens on educators [6].
In the non-profit sector, HITL systems are utilized in areas such as donor segmentation, campaign
optimization, and impact assessment. By integrating human expertise with AI capabilities, organizations can
better understand donor behavior, tailor communications, and measure the effectiveness of their initiatives.
Ethical Dilemmas in HITL Systems
The deployment of HITL systems raises several ethical dilemmas. In education, issues related to data privacy,
consent, and the potential for algorithmic bias in student assessments are of paramount concern [7]. The use of
AI tools must be transparent, with clear guidelines on data usage and mechanisms for accountability.
In the non-profit sector, ethical considerations include the equitable access to AI technologies, the potential for
reinforcing existing power imbalances, and the need for inclusive decision-making processes. Organizations
must ensure that AI applications align with their mission and values, promoting social good without exploiting
vulnerable populations.
The integration of HITL systems in education and the non-profit sector offers significant benefits, including
personalized learning experiences and enhanced operational efficiency. However, these advantages must be
weighed against the associated challenges and ethical dilemmas. A balanced approach, incorporating human
oversight and ethical considerations, is essential for the responsible deployment of AI technologies in these
sectors.
Human-in-the-Loop (HITL): definitions & modalities
HITL refers to system architectures that deliberately integrate human judgment/expertise into the lifecycle of
an AI system at critical decision points:
(a) pre-decision (data/model design and feature selection),
(b) real-time intervention (human overrides /interventions during automated operation), and
(c) post-decision oversight (audits, appeals, human review).
HITL spans a spectrum from heavy human control to lightweight human validation; the ethical promise derives
from preserving contextual reasoning and accountability that pure automation lacks [8].
Automation bias and oversight failure
A well-documented risk in HITL contexts is automation bias decision makers tend to over-rely on automated
outputs, reducing vigilance and failing to catch algorithmic errors. Systematic reviews demonstrate automation
bias across domains (healthcare, aviation, administrative systems) and identify mediators (display design,
workload, expertise) and mitigators (confidence displays, training). This means that mere human presence is
not sufficient; humans must be empowered and equipped to challenge machine outputs [9]
Human–ML augmentation & IS perspectives
Recent IS scholarship argues for nuanced typologies of human–ML augmentation (e.g., reactive oversight,
proactive oversight, informed reliance, supervised reliance) [10]
and calls for IS research to rethink classic assumptions in light of ML’s unique properties (data-trained
models, non-deterministic behaviors). Studies have shown how automation fairness failures invite managerial
and design strategies that center human–ML collaboration rather than binary automation/no-automation
choices [11].
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 364
Theoretical lens: sociotechnical systems + normative governance
We frame HITL as a sociotechnical governance mechanism: ethical outcomes arise from the interplay of
algorithmic affordances, human judgment, organizational structures, and regulatory context. This aligns with
IS sensibilities that emphasize technology-organization-environment interactions and the need for institutional
scaffolding to secure accountable outcomes [12] The E U AI Act’s explicit human oversight requirement
further signals policy momentum for operationalizing HITL in high-risk systems.
RESEARCH DESIGN & METHODOLOGY
Research strategy multiple qualitative case study
This study adopts a qualitative multiple case study [12] to compare how HITL is designed, enacted, and
sustained across three decision-sensitive domains: Education, IT, and Non-Profit. This design allows theory-
building about conditions for HITL sufficiency and identification of demi-regularities across contexts while
preserving attention to domain peculiarities.
Case selection & sampling
Purposive sampling of six organizations within the IT, Education and Non-for Profits sectors in Nigeria, (two
per domain) selected for active use of ADS with some human oversight mechanism (formal or informal).
Within each case, we interview 2 participants across roles: designers/engineers, administrators/managers,
frontline human reviewers, and affected stakeholders (teachers, beneficiaries). Total planned interviews: 6.
Documents (policies, decision logs, audit reports) and limited observations were reviewed.
Data collection instruments and Analysis
Semi-structured, open-ended questionnaires were administered to experts in the IT, Education, and Non-profit
sectors. The instrument was designed to capture perceptions of AI oversight, ethical concerns, and practical
experiences with Human-in-the-Loop (HITL) systems. Questions explored decision-making workflows,
accountability practices, and contextual challenges. Responses were documented, compiled, and later
transcribed into structured datasets for analysis. Collected responses from the spreadsheet was loaded into
NVivo for categorization into themes and final analysis [13].
Analytical lens: Ethical HITL Assessment Framework (EHAF) proposed
As Artificial Intelligence (AI) systems become increasingly integrated into critical sectors such as healthcare,
finance, and education, ensuring ethical decision-making is paramount [14]. Human-in-the-Loop (HITL)
systems have emerged as a key approach for embedding human judgment, oversight, and intervention into AI
workflows, helping to mitigate bias and prevent harmful outcomes [15].These systems are particularly
important in domains where decisions carry significant ethical implications, or where AI alone cannot fully
interpret complex, context-sensitive information [16].
To systematically evaluate the ethical adequacy of HITL systems, we propose the Ethical Human-in-the-
Loop Assessment Framework (EHAF), comprising four diagnostic dimensions:
Impact Severity – This dimension assesses the stakes of AI-influenced decisions, emphasizing the need for
human oversight in high-consequence scenarios. HITL systems are especially critical in domains such as
healthcare and student evaluation, where errors may have significant consequences [17].
Contextual Ambiguity – This dimension considers whether AI operates in environments requiring nuanced
human interpretation. AI systems often struggle with contextually ambiguous scenarios, making human
intervention essential for ethical decision-making [18].
Human Agency – This dimension evaluates whether human reviewers are empowered, trained, and authorized
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 365
to override AI decisions. Evidence suggests that ethical outcomes improve when humans are given meaningful
decision-making authority [19].
Transparency & Auditing – This dimension examines whether AI decisions are traceable and auditable. HITL
mechanisms support accountability and facilitate post hoc review of AI decisions, which is crucial for
compliance and ethical assurance (Rahwan, 2018).
The EHAF was employed as an analytic coding frame during cross-case analysis to determine HITL adequacy
and to derive policy and design recommendations [20]
Figure 1:Ethical HITL Assessment Framework (EHAF) Source: Authors work.
RESULTS
Thematic analysis of the survey responses generated insights into how Human-in-the-Loop (HITL) AI is
perceived and operationalized across the Education, IT, and Non-profit sectors. Using a hybrid deductive–
inductive approach, we mapped responses to the four components of the HITL Ethical Assessment Framework
(Impact Severity, Contextual Ambiguity, Human Agency, and Transparency & Auditing) while also identifying
emergent inductive themes such as Training & Capacity, Trust, and Infrastructure & Cost.
Sector-Level HITL Ethical Sufficiency
Sector Impact
Severity
Contextual
Ambiguity
Human
Agency
Transparency
& Auditing
Composite
%
Interpretation
Education High Moderate Moderate Moderate ~70% Moderate–
Strong
IT Low Low Weak Moderate ~45% Weak
Non-
profit
Moderate Moderate Moderate Strong ~60% Moderate
Table 1: Sector-Level HITL Scores: (Derived from coded scores: see sector_scores.xlsx)
Table 1 summarizes sector-level scores derived from coded responses. Education exhibited the highest
sufficiency (≈70%), reflecting strong recognition of decision sensitivity and greater emphasis on human
oversight. Non-profit organizations showed moderate sufficiency (≈60%), driven by donor accountability
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 366
mechanisms but constrained by infrastructural challenges. IT respondents reported the lowest sufficiency
(≈45%), indicating weaker integration of human oversight and contextual adaptation.
Table1. Sector-Level HITL Scores (see above table)
The sectoral contrasts are further visualized in Figure 1, which illustrates the distribution of component-level
scores. Education scored strongest on Impact Severity, while Non-profits were strongest on Transparency &
Auditing. The IT sector consistently underperformed across components, especially in Human Agency.
Figure 2:Sector × Component Heatmap (see above chart)
Deductive Findings: HITL Framework Components Impact Severity.
Education respondents emphasized the irreversible consequences of AI-driven decisions on students: “If the
algorithm sends a child home, that can change their life we cannot let a machine make that decision alone.”
Non-profits echoed this awareness in relation to community programs, whereas IT respondents often described
system errors as “minor and recoverable.” This reflects a divergence in how sectors perceive risk and
consequence.
Contextual Ambiguity
Ambiguity appeared across all sectors but manifested differently. In Education, it reflected cultural nuances
“The AI didn’t recognize the cultural nuance what looks like absenteeism was actually a festival.” In Non-
profits, ambiguity was tied to donor requirements conflicting with local needs, while IT respondents referred to
technical edge cases.
Human Agency
The strongest evidence of human oversight came from Education and Non-profits, where teachers and program
officers retained override authority: “Program officers override AI outputs if they conflict with donor
accountability rules.” By contrast, IT respondents often expressed reliance on automated recommendations:
“We mostly accept what the system suggests.”
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 367
Transparency & Auditing
Transparency was highly salient for Non-profits, where audit trails were maintained for donor accountability:
“We maintain audit trails for donors.” Education institutions relied on dual logging practices, while IT
respondents pointed to opaque audit logs “Logs are there, but no one outside IT understands them.”
These findings are summarized in Table 2, which presents sectoral exemplars for each framework component.
Table 2:Exemplar Quotes by Framework Component (Full exemplar dataset:
exemplar_sentences_by_code_and_sector.xlsx)
Component Education Example IT Example Non-profit Example
Impact Severity “If the algorithm sends a child
home, that can change their life we
cannot let a machine make that
decision alone.”
“Most system errors
are minor and
recoverable.”
“Decisions affect community
access, so we treat them with
caution.”
Contextual
Ambiguity
“The AI didn’t recognize the
cultural nuance what looks like
absenteeism was actually a
festival.”
“System doesn’t
account for unusual
local inputs.”
“Donor requirements often
clash with local realities.”
Human Agency “Teachers must always sign off
before final decisions.”
“We mostly accept
what the system
suggests.”
“Program officers override AI
outputs if they conflict with
donor accountability rules.”
Transparency &
Auditing
“We keep manual logs alongside
AI records.”
“Logs are there, but no
one outside IT
understands them.”
“We maintain audit trails for
donors.”
Table 2. Exemplar Quotes by Framework Component (see above table)
Inductive Findings: Emergent Themes
Beyond the deductive framework, inductive coding revealed themes that expand the ethical discussion of HITL
systems:
Training & Capacity: Respondents across all sectors highlighted skill gaps in interpreting AI outputs, with
Education respondents especially emphasizing staff training.
Trust: Trust in AI recommendations was uneven, with some respondents expressing confidence while others
described skepticism or discomfort.
Infrastructure & Cost: IT and Non-profit respondents frequently pointed to connectivity, affordability, and
sustainability challenges.
Policy & Governance: Non-profit respondents stressed the need for ethical policies and compliance
mechanisms.
Donor/Stakeholder Influence: Donor expectations strongly shaped HITL implementation in Non-profits, often
reinforcing accountability but also constraining flexibility.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 368
Table 3:Top Emergent Inductive Themes: (Counts from code_counts_overall.xlsx)
Table 3 presents these emergent themes with frequency counts and sector presence.
Table 3. Top Emergent Inductive Themes (see above table)
Overall Patterns
Figure 3 highlights the most frequently occurring codes across the dataset, underscoring the centrality of
Human Agency, Transparency, Training, and Trust in shaping perceptions of HITL sufficiency.
Figure 3:Top 10 Codes by Total Mentions (see above chart)
Taken together, the results reveal that while HITL principles are acknowledged across all three sectors, their
Theme Description
Frequency
(mentions)
Sector Presence
Training & Capacity
Calls for more staff training to handle HITL
systems
High All, esp. Education
Trust Mixed trust/distrust in AI recommendations High All sectors
Infrastructure & Cost Connectivity, affordability, system sustainability Moderate IT & Non-profit
Policy & Governance Formal policies, guidelines, ethics frameworks Moderate Non-profit heavy
Donor/Stakeholder
Influence
Donor requirements shape HITL use Moderate Non-profit
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 369
operationalization is uneven. Education emphasizes the gravity of decision-making, Non-profits emphasize
accountability structures, while IT emphasizes efficiency, often at the expense of human agency.
DISCUSSION
Summary of key findings
This study set out to examine how Human-in-the-Loop (HITL) AI is perceived and operationalized across
Education, IT, and Non-profit sectors, and to test the analytic utility of the HITL Ethical Assessment
Framework (Impact Severity; Contextual Ambiguity; Human Agency; Transparency & Auditing). The results
show consistent recognition of HITL principles across sectors, but reveal important differences in how those
principles are enacted. Education respondents showed the strongest concern for impact severity and clearer
human-override practices; Non-profit respondents emphasized transparency and audit trails driven by donor
accountability; IT respondents prioritized efficiency and technical transparency, but routinely showed weaker
human agency and contextual adaptation. Emergent themes especially training and capacity, trust,
infrastructure and cost, and vendor dependence cut across sectors and colored how HITL sufficiency was
experienced in practice.
Below, we interpret these patterns retroductively, proposing the deeper mechanisms that plausibly generate the
observed sectoral differences. We then discuss how these mechanisms confirm, complicate, and extend the
HITL framework, and outline practical and research implications.
Retroductive explanation: generative mechanisms behind sectoral differences
Institutional role ambiguity in IT → weak human agency
Pattern observed: IT respondents frequently reported automated workflows and acceptance of system outputs;
explicit human override or human review procedures were less common.
Retroductive mechanism: In many IT organizations, governance privileges technical uptime, automation, and
rapid response. Organizational roles emphasize system maintenance and scalability rather than deliberative
decision-making. This institutional orientation creates role ambiguity for non-technical staff: either no clear
decision owner exists, or decision authority defaults to technical teams whose priorities favor performance
metrics (uptime, throughput) over deliberative oversight. When role boundaries are unclear and incentives
favor automation, human actors are less empowered to intervene. This reduces real human agency even where
technologies are nominally “human-in-the-loop.”
Evidence link: IT respondents’ exemplar statements about accepting system recommendations and difficulty
interpreting logs suggest humans are present but lack authority or context to act. The mechanism explains low
Human Agency scores despite moderate Transparency measures.
Donor accountability pressures in Non-profits → strong transparency & audit orientation
Pattern observed: Non-profit respondents scored strongly on transparency and auditing; audit trails and
documentation for donors were common.
Retroductive mechanism: Non-profits operate within layered accountability regimes: they must satisfy
beneficiaries, regulators, and most immediately, donors and funders. Donors often require rigorous
documentation and measurable outcomes. This institutional pressure creates a compliance mechanism that
channels organizational behavior toward traceability and auditability. As a result, Non-profits adopt logging
and audit practices not only to support ethical HITL use but also to meet funder requirements. This institutional
logic strengthens Transparency & Auditing scores, even when contextual adaptation or technical capacity
remains partial.
Evidence link: Respondents explicitly referred to audit trails for donor reporting and program-officer approval;
these constraints explain why Non-profits show moderate composite sufficiency despite infrastructural limits.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 370
Decision sensitivity and cultural misalignment in Education → high impact awareness and localized
human checks
Pattern observed: Education respondents emphasized irreversible consequences for learners and described
concrete human override practices.
Retroductive mechanism: educational decisions (e.g., assessment, attendance consequences, disciplinary
actions) affect individual life chances directly and often irreversibly. This stakes-sensitivity creates moral and
social pressure on institutions (teachers, administrators) to retain human deliberation. Moreover, education
operates within rich cultural contexts where local norms, festivals, family arrangements, and linguistic
variation influence how data should be interpreted. When algorithmic outputs fail to capture such nuances,
practitioners deploy local knowledge to correct or contextualize outcomes. Thus, a combined mechanism of
stakes sensitivity + cultural embeddedness produces robust human intervention practices in Education.
Evidence link: Quotes about a student being wrongly identified during a community festival illustrate how
cultural nuance necessitates human judgment. This mechanism aligns with high Impact Severity and
meaningful Human Agency.
Capacity, cost, and vendor-dependence as cross-sector constraints
Pattern observed: Across sectors, training gaps, affordability, and overreliance on third-party providers
emerged as recurrent barriers.
Retroductive mechanism: Limited institutional resources (budget, skills, infrastructure) produce capacity
constraints that limit both the ability to create context-sensitive models and the competence to interpret and
act on AI outputs. Vendor dependence compounds this: when organizations outsource AI functions to third
parties, they often lose visibility and control over model behavior and logging practices. This combination
reduces effective transparency and human agency. Even where audit logs exist, lack of capacity to interpret
them renders transparency nominal rather than practical.
Evidence link: Recurrent references to training needs, costs of subscriptions, and opaque vendor logs show this
mechanism operates across sectors, explaining why some transparency measures do not translate into
improved HITL sufficiency.
How the findings confirm, qualify, and extend the HITL Ethical Assessment Framework
Confirmation: The four framework components capture the key dimensions practitioners consider when
evaluating HITL AI. Impact sensitivity reliably predicts stronger human oversight; transparency and auditing
are central to perceived ethical sufficiency; contextual ambiguity mediates whether automation is appropriate.
Qualification: The framework assumes that stronger presence of each component straightforwardly increases
ethical sufficiency. Our findings qualify that assumption: institutional incentives, resource constraints, and
stakeholder accountability can decouple components from ethical practice. For example, Non-profits can have
strong auditing (high Transparency) yet still experience constrained contextual adaptation because donor
requirements narrow permissible responses.
Extension: Empirically, emergent themes particularly Training & Capacity, Vendor Dependence, and Donor
Influence operate as cross-cutting moderators that affect all four components. We therefore propose extending
the framework to include a fifth layer: Capacity & Governance Context (resources, vendor relationships, and
accountability regimes). This addition helps explain why similar technical measures produce different
outcomes across sectors.
Practical implications
Clarify and formalize human roles. Organizations (especially in IT) should document decision responsi
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 371
-bilities and explicitly empower named roles to override or review AI outputs. Role clarity can be embedded in
standard operating procedures and escalation protocols.
Invest in interpretability + training together. Transparency efforts must be paired with investment in staff
training so that logs and explanations are actionable. Training programs should be sector-tailored (e.g.,
teachers vs. program officers vs. system administrators).
Design for local context. HITL workflows must include mechanisms to surface cultural edge cases and
incorporate local knowledge. In education, for example, systems should have procedures for teachers to flag
and annotate context that informs model retraining.
Audit vendor relationships. Where third-party providers are used, contracts should require explainability,
data access, and logging standards; donors and regulators can insist on these clauses to reduce opaque vendor
dependence.
Align donor incentives with local adaptability. Donors should design accountability frameworks that
encourage contextual adaptation rather than rigid KPIs that incentivize algorithmic standardization.
CONCLUSION
This study examined Human-in-the-Loop (HITL) AI in Education, IT, and Non-profit sectors using the HITL
Ethical Assessment Framework, which we developed to evaluate ethical sufficiency through four components:
Impact Severity, Contextual Ambiguity, Human Agency, and Transparency & Auditing. By applying this
framework to qualitative data, we demonstrated its analytic utility in capturing sectoral differences and
diagnosing ethical strengths and weaknesses.
Our findings reveal that Education emphasizes human oversight due to high decision stakes and cultural
nuance, Non-profits prioritize transparency under donor accountability, while IT organizations privilege
efficiency at the expense of agency. Retroductive analysis identified underlying mechanism’s institutional role
ambiguity, donor pressures, cultural misalignment, and capacity constraints that shape how HITL principles are
enacted.
The study’s contribution is twofold: empirically validating and extending the HITL framework with a Capacity
& Governance Context dimension, and practically offering actionable recommendations for policy and
organizational practice.
REFERENCES
1. Berente, N., et al., Managing Artificial Intelligence. MIS Quarterly, 2021. 45(3): p. 1433-1450.
2. Shrestha, Y.R., S.M. Ben-Menahem, and G. von Krogh, Organizational Decision-Making Structures in
the Age of Artificial Intelligence. California Management Review, 2019. 61(4): p. 66-83.
3. Wamba, S.F., Akter, S., & Guthrie, C., Making big data analytics perform: the mediating effect of big
data analytics dependent organizational agility. . Systèmes d'information & management, , 2020. 25(2),
7-31.
4. Lyytinen, K., J.V. Nickerson, and J.L. King, Metahuman systems = humans + machines that learn.
Journal of Information Technology, 2020. 36(4): p. 427-445.
5. Jarvenpaa, S.L., Sourcing data for data-driven applications: foundational questions. In Research
handbook on artificial intelligence and decision making in organizations. Edward Elgar Publishing.,
2024: p. (pp. 17-37). .
6. Owoc, M.L., Sawicka, A., & Weichbroth, P. , Artificial intelligence technologies in education: benefits,
challenges and strategies of implementation. In IFIP international workshop on artificial intelligence
for knowledge management. 2019: p. (pp. 37-58).
7. Azeem, S. and M. Abbas, Personality correlates of academic use of generative artificial intelligence and
its outcomes: does fairness matter? Education and Information Technologies, 2025. 30(13): p. 18131-
18155.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)
ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025
www.rsisinternational.org
Page 372
8. Amershi, S., et al., Guidelines for Human-AI Interaction, in Proceedings of the 2019 CHI Conference
on Human Factors in Computing Systems. 2019. p. 1-13.
9. Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). , Automation bias: a systematic review of frequency,
effect mediators, and mitigators. Journal of the American Medical Informatics Association, , 2012.
19(1), : p. 121-127.
10. Abdel-Karim, B., et al., How AI-Based Systems Can Induce Reflections: The Case of AI-Augmented
Diagnostic Work. MIS Quarterly, 2023. 47(4): p. 1395-1424.
11. Teodorescu, M.H.M., et al., Failures of Fairness in Automation Require a Deeper Understanding of
Human–ML Augmentation. MIS Quarterly, 2021. 45(3): p. 1483-1500.
12. Yin, R.K., Case study research and applications. Thousand Oaks, CA: Sage., 2018. (Vol. 6). .
13. Danermark, B.A., Interdisciplinary research and critical realism the example of disability research. .
5(1), , 2002. 56-64.
14. Sawyer, B.D., Miller, D. B., Canham, M., & Karwowski, W., Human factors and ergonomics in design
of a 3: automation, autonomy, and artificial intelligence. Handbook of human factors and ergonomics, ,
ed. 1385-1416. 2021.
15. Amershi, S., Cakmak, M., Knox, W. B., & Kulesza, T. , Power to the people: The role of humans in
interactive machine learning. AI magazine, , 2014. 35(4),: p. 105-120.
16. Storey, V.C., Yue, W. T., Zhao, J. L., & Lukyanenko, R. (2025). , Generative artificial intelligence:
Evolving technology, growing societal impact, and opportunities for information systems research. .
Information Systems Frontiers, , 2025: p. 1-22.
17. Doshi-Velez, F., & Kim, B., Towards a rigorous science of interpretable machine learning. . arXiv
preprint arXiv:, 2017. 1702.08608.
18. Jiang, N., Liu, X., Liu, H., Lim, E. T. K., Tan, C. W., & Gu, J., Beyond AI-powered context-aware
services: the role of human–AI collaboration. Industrial Management & Data Systems, , 2023. 123(11),
2771-2802.
19. Bryson, J.J. and A. Theodorou, How Society Can Maintain Human-Centric Artificial Intelligence, in
Human-Centered Digitalization and Services. 2019. p. 305-323.
20. Rahwan, I., Society-in-the-loop: programming the algorithmic social contract. . Ethics and information
technology, , 2018. 20(1), 5-14.
Ethical Considerations and Approval
This study was conducted in full compliance with established ethical standards for research involving human
participants. Ethical approval was obtained from the Institutional Review Board of the American University of
Nigeria; after undergoing the ethical certification review with the certificates number Record ID 3185315 and
Record ID 37009720 and informed consent was secured from all participants prior to data collection.
Conflict of Interest
The authors declare that there are no conflicts of interest, financial or otherwise, that could have influenced the
conduct or outcomes of this research.
Data Set Availability
The datasets generated and analyzed during the current study are available and will be made available when
needed.