Trustworthy AI Framework: GIFTERS

Updated 7 December 2025

GIFTERS is a framework that defines seven key attributes for trustworthy AI, providing clear guidelines for both generative AI and materials science applications.
It employs measurable indicators such as bias ratings, robustness tests, and transparency audits to practically assess and compare AI systems.
The framework integrates human-in-the-loop reviews and lifecycle evaluation to ensure continuous improvement and ethical deployment of AI technologies.

Trustworthy AI frameworks are essential for ensuring that artificial intelligence systems, particularly in high-impact areas such as generative models and scientific discovery, operate consistently with ethical, transparent, and robust methodologies. Two influential lines of research—one focusing on generative AI and the other on materials science—have converged on the acronym GIFTERS to encapsulate key dimensions of trustworthiness. While the precise expansion and content of GIFTERS differ between domains, the shared aim is to support reproducibility, societal acceptability, and responsible deployment of AI/ML systems (Jeong et al., 30 Aug 2025, Amirian et al., 30 Nov 2025).

1. Origins and Definitions of GIFTERS

The GIFTERS framework crystallizes essential attributes for the evaluation of trustworthy AI. In materials discovery, GIFTERS encompasses Generalizability, Interpretability, Fairness, Transparency, Explainability, Robustness, and Stability, emphasizing model performance under data shifts, interpretability, bias control, open reporting, feature attribution, insensitivity to noise, and reproducibility under parameter changes (Amirian et al., 30 Nov 2025). In the context of generative AI, the acronym has been adapted to address distinct ethical and social imperatives, including Governance/Accountability, Integrity (Privacy & IP Protection), Fairness, Transparency, Explainability, Robustness/Reliability, and Source Traceability & Safety (Jeong et al., 30 Aug 2025).

GIFTERS Domain	Principle(s) (Materials Science)	Principle(s) (Generative AI)
G	Generalizability	Governance & Accountability
I	Interpretability	Integrity (Privacy & IP Protection)
F	Fairness	Fairness
T	Transparency	Transparency
E	Explainability	Explainability
R	Robustness	Robustness, Reliability (Accuracy, Consistency)
S	Stability	Safety & Source Traceability

The acronym enables structured assessment across AI/ML applications by organizing evaluation metrics, process controls, and documentation requirements under principled headings.

2. Dimension Definitions and Indicators

Each GIFTERS dimension is defined by explicit criteria, accompanied by measurable indicators and associated quantitative or qualitative evaluation methods.

Generative AI (Jeong et al., 30 Aug 2025):

Fairness: Measured by data, algorithmic, and results bias, as well as accessibility, with group accuracy differentials (ΔAcc = |Accₐ – Acc_b|), 5-point bias ratings, and user surveys.
Transparency: Quantified by documentation completeness, prompt and intermediate processing logs, clarity of data provenance, and user disclosure (e.g., “I am an AI”), via checklist audits and system-log reviews.
Accountability: Documented assignment of responsibilities, remediation mechanisms, audit trails, and governance structures, assessed through policy audits and remediation case studies.
Safety: Assessed by red-team testing (number of successful exploit prompts), filter precision/recall, and harmfulness scales.
Privacy: Evaluated through PII anonymization (k-anonymity, ε-differential privacy), consent audits, data-flow mapping, and penetration tests.
Accuracy: Measured as factual correctness (N_correct/N_total), hallucination rates (h = N_halluc / N_total), and expert-vetted domain relevance.
Consistency: Quantified by repeated-query similarity (Rep = (1/N) Σ Sim(o_run1, o_run2)), BLEU/cosine similarity scores, and compliance with style-guide or dialogue continuity.
Robustness: Tested via input fuzzing (% degradation), adversarial success rates, and stress-test logs.
Explainability: Indicators include provision of explanations, faithfulness to model internals, conciseness, comprehension scores, and reliability—assessed via interviews, expert reviews, and XAI metrics.
Copyright/IP Protection: Metrics include plagiarism rate, source disclosure, and license tracking, with audits and automated scanners.
Source Traceability: % of outputs with explicit references (T = N_referenced / N_total) and link resolvability.

Materials Discovery (Amirian et al., 30 Nov 2025):

Generalizability: Hold-out validations, out-of-distribution detection (EMD, Jensen-Shannon divergence).
Interpretability: Use of decision trees, surrogates, prototype layers.
Fairness: Demographic parity (ΔDP), SMOTE, stratified splits, subgroup AUROC.
Transparency: Code/data/model sharing, adherence to FAIR (Findable, Accessible, Interoperable, Reusable) standards, software quality metrics.
Explainability: SHAP/LIME feature attributions, counterfactual analysis.
Robustness: Noise quantification, adversarial tests, negative result pipelines.
Stability: Hyperparameter sweeps, performance spread (Taylor diagrams, violin plots).

3. Lifecycle Integration and Evaluation

For generative AI, the framework is designed for iterative use through the AI lifecycle:

Development: Define measurable objectives (e.g., bias ΔAcc < 3%), select and weight indicators, prepare data (diversity, anonymization), and conduct pre-deployment safety/robustness tests.
Deployment: Final evaluation using held-out data, expert and small-scale user reviews, compliance checks, and comprehensive reporting.
Monitoring: Continuous interaction logging, periodic reevaluation on fresh data, user surveys and incident investigation, governance updates, and public dashboard disclosure.

In the materials science context, evaluation of each principle uses a binary score (addressed/omitted), yielding a total count per publication (median 5/7). Generalizability and Transparency are most commonly addressed, while Fairness and Stability are least frequently implemented (Amirian et al., 30 Nov 2025). Bayesian protocols often neglect Fairness; non-Bayesian approaches more frequently omit Interpretability.

4. Comparative Policy and Domain Variations

Implementation of GIFTERS dimensions varies by regulatory, scientific, and regional contexts:

South Korea: Principle-based guidelines and self-checks prioritize social consensus. Fairness and Privacy are weighted equally; accountability is often voluntary.
United States: Favors risk-based private self-regulation (e.g., NIST AI RMF); Safety and Transparency prioritized, lower legal enforceability.
EU: The AI Act mandates high-risk system controls (Fairness, Explainability, Source Traceability) with legal enforcement and sanctions.
China: Centralized governance with rapid standard enforcement; Accountability is state-directed, and Privacy is relatively permissive (Jeong et al., 30 Aug 2025).

Materials research demonstrates cross-disciplinary borrowing: climate science uses prototype layers and counterfactuals; healthcare leverages subgroup performance and checklists; NLP incorporates confidence gating and physics-informed vetting (Amirian et al., 30 Nov 2025). This suggests that best practices from regulated or high-trust disciplines can fill gaps in AI trustworthiness protocols.

5. Human-in-the-Loop and Uncertainty Quantification

Integrating human review at critical stages enhances trustworthiness, especially in high-stakes workflows:

Models (Bayesian GPs or ensembles) generate predictions with calibrated uncertainty (σ).
Confidence thresholds (e.g., accept predictions where σ_i < c*) reduce exposure to unreliable predictions.
Human experts audit flagged cases, correcting data or surfacing bias.
This corrected information is incorporated via active learning, improving fairness, robustness, and stability (Amirian et al., 30 Nov 2025).

Such feedback loops directly address deficits observed in both generative and scientific AI implementations.

6. Actionable Recommendations and Best Practices

Comprehensive adoption of GIFTERS requires context-sensitive adaptation:

For Policymakers: Enforce minimum lifecycle evaluation, mandate publicly available evaluation scorecards, harmonize standards internationally, and incentivize transparency and SME adoption.
Developers & Vendors: Integrate GIFTERS metrics into MLOps pipelines, maintain governance review bodies, provide version-controlled documentation, automate bias and safety checks, and implement real-time fact/source-tracking.
Users & Stakeholders: Demand clear disclosure of model capabilities, limitations, and lineage; participate in fairness/privacy assessments; insist on accountability pathways and explainable outputs in safety-critical domains.

By codifying dimensions, selecting measurable indicators, maintaining cross-lifecycle vigilance, and benchmarking against domain-specific best practices, the GIFTERS framework enables the development and operation of AI that meets both technical and ethical standards (Jeong et al., 30 Aug 2025, Amirian et al., 30 Nov 2025).