"Ensuring outputs are clinically and ethically appropriate for ageing populations, reflecting societal values of equity, inclusivity, and wellbeing."
— 50-Responsible_Generative_AI_RAIDT_V10.docx, Section 3.2

Evaluation Criteria

Research Findings

Method Performance Analysis

RAG

4.9/5.0

High thematic fidelity with trusted retrieval sources

  • Best for policy and legal domains
  • Strong source attribution
  • Paired with CCRA3, ECtHR corpora

PEFT

4.8/5.0

Domain-specific training reduces drift by 34%

  • Reduced domain drift in finance/healthcare
  • Adapter versioning support
  • Strong RAID-T compliance

RLHF

4.3/5.0

Tone and value alignment

  • Risk surfacing in clinical settings
  • Reward-linked summaries
  • Effective for retail and finance tone

Prompting

3.0/5.0

Lightweight but limited

  • Effective for constrained tasks
  • Fails in high-risk settings
  • Needs moderation/scaffolding
"Influence methods such as RLHF and PEFT consistently supported domain-aligned outputs, especially when trained on domain-specific datasets. In finance and healthcare, PEFT helped reduce domain drift by 34% compared to prompt engineering alone." — 02-Generative_AI_V10.docx, Section 6.2.1

Governance Standards Alignment

EU AI Act Article 9

Requires risk management and human oversight for high-risk systems, including healthcare

European Commission, 2024

ISO/IEC 42001

Emphasizes organizational accountability and ethical governance structures

ISO/IEC, 2025

Philosophical Foundations

Scholars such as Floridi et al. (2018) and Mittelstadt et al. (2019) have laid a philosophical foundation for RAI, articulating the need for principles such as explicability, fairness, and respect for human dignity to be embedded throughout the AI lifecycle.

AI Governance Mechanisms

  • Ethics review boards and cross-disciplinary advisory panels
  • Model documentation protocols (e.g., model cards, data statements)
  • Algorithmic risk assessments and socio-technical impact evaluations
  • Technical standards and certifications (e.g., ISO/IEC 42001, NIST AI RMF)
  • Public regulations, such as the EU Artificial Intelligence Act and UK regulatory proposals
"Responsible AI demands a shift from techno-solutionism to socio-technical systems thinking, recognising that algorithmic impacts are always contextual and political." — AI Now Institute, 2021

AI Lifecycle Management Framework

  • Data phase: Ethical data collection, anonymisation, fairness in representation
  • Model phase: Bias detection, robustness checks, interpretable architecture design
  • Deployment phase: Monitoring, red-teaming, feedback loops, retraining protocols
  • Decommissioning phase: Sunset procedures, data retention audits, traceability

A Responsible AI ecosystem encompasses a set of principles and practices designed to ensure that AI systems are aligned with ethical values, social norms, and legal expectations.

AI Taxonomy & Classification

▼

Establishes classification systems for understanding and managing different forms of AI behaviour, intent, and capabilities

Barredo Arrieta et al., 2020

Explainable AI (XAI)

▼

Aims to make black-box AI models interpretable by humans, enhancing transparency and accountability. XAI enhances human oversight, enables fault diagnosis, and supports regulatory compliance.

"Explainable AI (XAI) is an emerging class of machine learning techniques that produce more interpretable models while maintaining predictive performance." — Rai et al., 2019

AI Governance & Oversight

▼

Covers institutional mechanisms for managing AI development and deployment, including oversight, accountability structures, and regulation

Human-AI Collaboration

▼

Focuses on designing AI systems that augment, rather than replace, human decision-making. This is particularly critical in high-stakes contexts such as clinical diagnostics, judicial sentencing, and financial risk management.

"Designing for hybrid agency enhances resilience, adaptability, and ethical grounding in socio-technical systems." — Vaia et al., 2022

Ethical AI Design

▼

Embeds values such as fairness, inclusivity, and non-maleficence into AI algorithms and datasets

Petratos, 2021

Model Transparency & Trust

▼

Making the inner workings, decisions, and data dependencies of AI systems visible and understandable. The overarching goal is to develop trustworthy AI—a concept formalised in the European Commission's Ethics Guidelines (2019).

AI Risk Management

▼

Identifying, assessing, and mitigating risks such as data drift, misuse, adversarial attacks, and systemic bias

AI Lifecycle Management

▼

Frameworks to manage AI from data collection and model training to monitoring, updating, and decommissioning

AI Auditing & Compliance

▼

Tools and methodologies for verifying AI system behaviour, data integrity, and regulatory compliance

Decision Support Systems

▼

Systems that leverage AI to inform or support, but not fully automate, human decision-making

Hybrid AI Systems

▼

Combine symbolic reasoning with deep learning or integrate human feedback loops, increasing reliability and adaptability

"Healthcare was chosen as a critical test domain because clinical decisions are time-sensitive, data-dense, and often life-critical. Generative AI used in this space must be evaluated not only on fluency, but on factual consistency, medical relevance, and accountability readiness." — 02-Generative_AI_V10.docx, Section 6.3.1

Key Healthcare Results

  • Clinical accuracy scores: 4.6-4.8/5.0 across methods
  • Red flag surfacing significantly improved with RLHF
  • Domain-specific medical alignment strong with PEFT
  • RAG + PEFT combination best for hallucination reduction
Source: 21-Healthcare_Domain_V8.docx, Reviewer Checkpoints
"By embedding RAID-T into empirical evaluation, we operationalise abstract governance principles into measurable criteria. Whereas frameworks such as the EU AI Act and ISO/IEC 42001 articulate requirements at regulatory or organisational levels, RAID-T translates these into task-level dimensions that can be applied directly to AI outputs." — 50-Responsible_Generative_AI_RAIDT_V10.docx, Section 6.1

The study advances theorisation on influence methods by demonstrating that stacked influence methods deliver superior RAID-T alignment, suggesting that multi-method orchestration represents a stronger model of responsible AI. This supports emerging debates on "AI controllability" and "steerability," positioning influence methods as governance mechanisms rather than mere optimisation techniques.

Research Proposition

"P1: Responsibility. Stacked influence methods (e.g., LoRA + RAG + prompting) will generate outputs with higher clinical and ethical appropriateness than prompt-only baselines." — 50-Responsible_Generative_AI_RAIDT_V10.docx, Section 3.8