2.1 International AI Ethics Norms and Principles

The AI Principles established by the OECD in 2019 defined human-centered values, fairness, transparency, safety, and accountability as core components of AI ethics and laid the foundation for international discussions [1]. In the same year, the EU published its Ethics Guidelines for Trustworthy AI, specifying requirements for “trustworthy AI” [9], and UNESCO strengthened AI ethics to a global normative level through its 2021 Recommendation [10]. However, Floridi and Cowls argue that although international guidelines use similar language, they embed heterogeneous ethical concepts and require a unified metaframework [2]. Jobin et al. found that many global AI ethics guidelines repeat the same values but lack concrete mechanisms for implementation and evaluation [4]. Thus, while international norms are abundant, their application through empirical or operational models remains limited [4].

2.2 Responsibility Gap

Accountability is one of the oldest and most fundamental issues in AI ethics. Matthias defined the situation in which learning AI systems autonomously make decisions that no longer allow clear attribution of responsibility as the “responsibility gap” [5]. The responsibility gap has since become a central challenge in norms, law, and policy, particularly in domains such as medical AI misdiagnosis, autonomous vehicle accidents, and automated credit scoring [2], [5]. However, research on the responsibility gap remains largely rooted in moral philosophy, creating significant gaps with real-world operational contexts, and lacks clear legal or institutional criteria for responsibility allocation [2], [5].

2.3 Fairness and Bias

Fairness is among the ethical values most strongly emphasized by the OECD, EU, and UNESCO [1], [9, 10]. Numerous studies highlight that AI systems, which rely on data-driven learning, may internalize or reinforce societal biases [4]. Fairness research has expanded across three layers—data-level bias, algorithmic structural bias, and outcome bias [1, 2], [12]—but much of the work remains measurement-oriented, with insufficient development of actionable mitigation frameworks [2], [4].

2.4 Human-Centered Design (HCD/HCAI)

Recent AI ethics discourse has shifted from technical rule-based approaches toward human-centered perspectives [11]. Schmager et al. identified key components of human-centered AI design, including human controllability, explainability, predictability, trustworthiness, ethical safety, and social acceptability [12]. Bevilacqua defined human-centered AI as a design philosophy that ensures human involvement, control, psychological safety, and societal trust [13]. Nevertheless, Wilkens and colleagues found that much of the human-centered AI literature remains conceptual or declarative, with limited empirical validation [14, 15].

2.5 Generative AI and the Re-establishment of Research Ethics

Since the rise of generative AI, “Generative AI + Research Ethics” has become one of the most rapidly expanding areas of inquiry [3], [6]. Nature warned that generative AI could destabilize or collapse scholarly production systems and argued for stronger researcher-led oversight mechanisms [3]. The ERA Forum’s guidelines established explicit obligations for researchers, including AI usage disclosure, verification responsibilities, provenance transparency, privacy considerations, risk evaluation, and misuse prevention [6]. Yet these guidelines remain mostly principle-driven, leaving a persistent gap between ethical norms and practical implementation [7], [17].

2.6 Limitations of Existing Research and Contributions of This Study Despite the richness of international AI ethics principles, existing research shares several limitations:

(1) lack of executable evaluation models,

(2) fragmented treatment of fairness, accountability, and HCD,

(3) insufficient incorporation of risks associated with post–generative AI environments,

(4) lack of empirical or practice-based validation, and

(5) limited reflection of sector-specific characteristics, particularly in the content industry [1], [3, 4], [1 3, 14].

To address these gaps, this study integrates the three core ethical components and incorporates AI intelligence levels (AIL-5) to propose a practical and quantifiable evaluation methodology for AI research ethics.

3.1 Research Ethics from the Perspective of Fairness

Fairness is a foundational value that determines the ethical acceptability of AI technologies [1]. Because AI systems learn and make decisions based on data, insufficient representativeness or embedded historical bias can result in structural disadvantages for particular groups [4]. As AI increasingly mediates resource allocation, evaluation, and recommendation processes within digital transformation environments, fairness concerns evolve from technical imperfections to systemic social risks [18]. Accordingly, fairness must be institutionalized as a lifecycle-based procedure encompassing (1) minimizing sampling bias at the data collection stage, (2) bias detection and mitigation at the algorithmic design stage, and (3) verifying that outcomes do not reproduce or exacerbate inequality at the deployment stage [1], [12], [19].

3.2 Research Ethics from the Perspective of Accountability

Accountability is the most fundamental concern in AI research ethics, and the responsibility gap remains an area without complete institutional solutions [5]. In digital transformation environments, accountability extends beyond responsibility for automated decisions to include obligations such as the researcher’s duty to verify AI-generated outputs, responsibility for data governance, and mechanisms for preventing research misconduct [2]. In particular, the rise of generative AI has strengthened norms requiring researchers to (1) take ultimate responsibility for verifying AI-generated results and (2) disclose the extent of AI involvement in their work [3], [6, 7]. Thus, accountability must be re-established as a proactive ethical element integrated into research design—shifting from post-hoc regulation to anticipatory responsibility frameworks that identify risks and specify responsibility structures in advance [7], [17].

3.3 Research Ethics from the Perspective of Human-Centered Design (HCD)

Human-centered design reflects the ethical demand that AI systems must not violate but rather enhance human values, rights, autonomy, and welfare [10, 11], [13]. Key components of HCD include human controllability, explainability, risk awareness, and psychological and social acceptability [12, 13]. If HCD is not institutionalized during the research and development process, fairness and accountability risk remaining at the level of technical declarations rather than becoming actionable ethical practices [14]. Therefore, research ethics must integrate HCD as the foundational philosophical frame that supports fairness and accountability and continuously monitor how well human values are embedded throughout the AI lifecycle [13], [20].

3.4 Ethical Issues of Generative AI from an HCI Perspective and Corresponding Responses

Generative AI intervenes in problem-solving, information acquisition, and content creation processes, fundamentally altering human cognition and interaction. From an HCI perspective, the central ethical issues can be summarized as:

(1) weakened user control,

(2) instability of trust experiences,

(3) increased cognitive dependency,

(4) erosion of authorship and creative agency, and

(5) deterioration of human-to-human interaction.

To address these issues, the following strategies are required:

— Human-in-the-Loop design,

— construction of Explainable UX,

— designing AI for human cognitive augmentation rather than replacement,

— transparent attribution and provenance systems,

— and interface designs that preserve human-human interaction [9-14].

These strategies provide concrete methods for restructuring generative AI systems within a human-centered design paradigm.

4.1 Design of the FAH-75

The FAH-75 model quantitatively evaluates compliance with AI research ethics based on the three pillars of fairness (F), accountability (A), and human-centered design (H). It draws on the OECD AI Principles [1], the EU Trustworthy AI Guidelines [9], UNESCO’s Recommendation [10], Floridi’s ethical design principles [2], [19], and the responsibility gap literature [5]. FAH-75 comprises 15 items (5 per pillar), each scored on a 1–5 scale, yielding a maximum of 75 points. The model aims to move beyond principle-based ethical discourse toward actionable and measurable procedures [4], [10, 11]. It also incorporates a three-stage evaluation process—self-assessment, expert review, and stakeholder verification—in line with the heightened research accountability norms of the generative AI era [3], [6, 7].

4.2 Structure and Evaluation Criteria of the FAH-75

The FAH-75 (Fairness–Accountability–Human-Centered Design) evaluation framework proposed in this study is designed to quantitatively assess key ethical dimensions of AI research. The framework consists of a total of 15 evaluation items, each scored on a five-point scale ranging from 1 (very insufficient) to 5 (excellent), yielding a maximum score of 75 points. The items are organized into three core ethical dimensions: Fairness (F), Accountability (A), and Human-Centered Design (H).

(1) Fairness (25 points)

The fairness dimension assesses whether an AI system is designed and operated in a manner that prevents the reproduction or amplification of social biases. The evaluation focuses on the following five criteria:

Data Representativeness: Whether training data adequately reflect demographic and contextual diversity.

Bias Detection and Mitigation Mechanisms: The existence of systematic procedures to identify and reduce algorithmic bias.

Disparate Impact Analysis: Assessment of whether algorithmic outcomes produce discriminatory effects across social groups.

Pre-assessment of Social Inequality Risks: Evaluation of potential risks related to reinforcing structural inequalities.

Continuous Fairness Monitoring: Presence of mechanisms for ongoing fairness evaluation after deployment.

Each item is scored based on the transparency, consistency, and verifiability of fairness-related controls and processes.

(2) Accountability (25 points)

The accountability dimension evaluates whether responsibility for AI system behavior is clearly defined and enforceable across the system lifecycle.

Attribution of Responsibility: Clarity regarding who is accountable (developers, operators, institutions).

Disclosure of AI Use and Intervention Scope: Transparency regarding when and how AI systems influence decisions.

Verification and Error-Correction Mechanisms: Existence of procedures for validating outputs and correcting errors.

Data Governance and Security Responsibility: Structures for managing data integrity, privacy, and security.

Risk Response and Liability Mechanisms: Defined protocols for managing failures, harm, and accountability allocation.

Scores are assigned based on the clarity of responsibility structures, documentation practices, and the feasibility of external auditing.

(3) Human-Centered Design (HCD) (25 points)

The human-centered design dimension evaluates whether AI systems enhance human autonomy, trust, and well-being rather than undermine them.

Human-in-the-Loop Control: The extent to which humans can intervene, override, or supervise system decisions.

Explainability and Interpretability: The degree to which system behavior can be understood by users.

Psychological and Cognitive Safety: Consideration of user well-being and avoidance of manipulation or overreliance.

Human Capability Augmentation: Whether AI is designed to support and enhance human judgment rather than replace it.

Social Acceptability and Ethical Impact Awareness: Assessment of broader societal implications and ethical acceptability.

(4) Evaluation emphasizes user experience, social consequences, and long-term ethical implications rather than purely technical performance. Scoring Methodology (1–5 Scale)

Each evaluation item is assessed according to the following criteria:

These criteria are designed not to eliminate qualitative judgment, but to ensure consistency and reproducibility across evaluators by providing a semi-quantitative assessment framework.

(5) Multi-Stage Evaluation Structure and Conflict Resolution Mechanism

The FAH-75 framework adopts a three-stage evaluation structure:

1) Self-Assessment Developers or researchers conduct an internal assessment to identify potential ethical risks and compliance gaps at an early stage.

2) Expert Review Domain experts in AI ethics, law and policy, and human–computer interaction (HCI) review and validate the assessment results, providing professional judgment and corrective feedback.

3) Stakeholder Verification Stakeholders—including end users, civil society representatives, and domain specialists—evaluate the system from a societal impact perspective.

When discrepancies arise among the three evaluation stages, the precautionary principle is applied: the lowest score is adopted as the final value, and the rationale for this decision must be explicitly documented. This mechanism is designed to prevent the underestimation of ethical risks and to ensure conservative, responsibility-oriented decision-making.

4.3 The AI Intelligence Level Framework (AIL-5)

Ethical requirements for AI must intensify as the “intelligence level” of AI systems increases. Reflecting advancements in cognitive science, HCI, and AGI research, this study defines the AIL-5 (AI Intelligence Level-5) framework as follows:

AIL-1: Reactive intelligence (rule-based or simple statistical systems)

AIL-2: Pattern-learning intelligence (ML/DL prediction models)

AIL-3: Language and reasoning intelligence (LLM-based inference and generation)

AIL-4: Autonomous agent intelligence (goal formulation, planning, and action)

AIL-5: General adaptive intelligence (AGI-oriented, highest-risk systems)

AIL-5 is not a mere performance taxonomy.

It functions as an ethical-risk indicator that integrates autonomy, generality, societal impact, high-risk decision involvement, and sensitivity to data. Therefore, as AIL levels increase, AI systems require proportionally stronger safeguards in fairness, accountability, and human-centered design.

3.7 FAH-75 × AIL-5 Integrated Evaluation Methodology (IFAEF)

FAH-75 operates in conjunction with the AIL-5 intelligence level framework. As AI systems exhibit higher levels of autonomy, generality, and social impact, the required ethical threshold increases accordingly. Thus, FAH scores are not evaluated in isolation but are interpreted relative to the system’s AIL level.

This design ensures that:

Low-risk systems are not over-regulated, and High-risk or agentic systems are subject to proportionally stronger ethical requirements.

Existing AI ethics assessments remain largely at the level of declarative principles and fail to adequately reflect varying levels of AI risk. To address these limitations, this study proposes an integrated evaluation model—IFAEF (Integrated FAH–AIL Evaluation Framework).

The framework consists of four components:

1. FAH-75 Ethical Compliance Assessment (0–75 points)

Evaluates fairness (F), accountability (A), and human-centered design (H) across 15 items, each rated on a 1–5 scale.

2. AIL-5 Intelligence and Risk Level Assessment (5–25 points)

Scores the AI system based on autonomy, generality, societal impact, high-risk decision involvement, and data sensitivity.

3. Level-Based Ethical Threshold Adjustment

Ethical requirements are strengthened proportionally to the AI’s intelligence/risk score (L).

A higher L score requires a higher minimum FAH score to pass the evaluation.

4. Final Score Calculation

ESS = FAH + L (Total: 100 points)

The system is deemed ethically acceptable only when the FAH score meets or exceeds the threshold corresponding to its L score.

The characteristics of the IFAEF integrated evaluation model are distinguished from existing AI ethics assessment frameworks in four major aspects.

First, the model adopts a level-adjusted ethical threshold, meaning that the required levels of fairness, accountability, and human-centered design automatically increase as the AI’s intelligence level and associated risks rise. This transforms the abstract guidance of international AI ethics principles into concrete, risk-responsive regulatory criteria that correspond to the actual capabilities of AI models [1, 2], [1 2].

Second, IFAEF enables a quantified and operational ethics assessment, moving beyond the declarative and principle-driven evaluations that dominate existing approaches. By assigning numerical values (25 points each) to fairness (F), accountability (A), and human-centered design (H), and by evaluating autonomy, generality, and high-risk action potential through the AIL index, the framework allows researchers and institutions to diagnose the ethical status of AI systems quantitatively. This advances the field by overcoming the ambiguity of traditional ethical discourse and providing a comparable and verifiable evaluation methodology [4], [16].

Third, the model has multi-dimensional applicability across research, development, and policy domains. FAH-75 covers both research ethics (RCR) and practical development ethics, while AIL-5 is designed to support policy-oriented risk and impact assessments. Accordingly, IFAEF is not only a tool for researchers but can also be applied to service operation standards, administrative decision-making systems, and public-sector AI impact assessment frameworks [3], [6], [9].

Fourth, IFAEF surpasses existing models by providing an ethical computation framework capable of addressing agentic AI and AGI-level systems. AI at AIL-4 to AIL-5 possesses high autonomy—such as goal-setting, action execution, and tool use—and consequently introduces dramatically heightened risks, including responsibility gaps, weakened human control, and amplified societal impacts. By incorporating strengthened accountability checks and HCD requirements for these high-risk systems, the framework offers an ethical preparedness model suitable for the AGI-oriented technological era [3], [19].

In conclusion, IFAEF quantitatively reflects risk levels associated with AI development stages and incorporates an adaptive structure that automatically adjusts ethical requirements. Thus, it can be regarded as a practical, predictive, and integrated evaluation framework that supports AI research ethics in the digital transformation era.