A structured metrics framework for measuring and evaluating the explainability, documentation, and disclosure levels of AI algorithms and automated decision-making systems across multiple transparency dimensions.
Semantic Classification
Content
Compositional Relationships (Components)
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:ExplainabilityMetrics))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:BiasDetectionMetrics))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:DocumentationStandards))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:DisclosureRequirements))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:AuditTrail))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:PerformanceMetrics))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:DataProvenance))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:ModelDocumentation))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:hasPart ai:DecisionLogging))
Dependency Relationships
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:requires ai:AuditMechanism))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:requires ai:ModelDocumentation))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:requires ai:DataProvenance))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:requires ai:ExplainableAI))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:requires ai:BiasDetectionMethods))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:requires ai:AutomatedDecisionSystem))
Capability Relationships
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:enables ai:AIAccountability))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:enables ai:AlgorithmicAuditing))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:enables ai:RegulatoryCompliance))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:enables ai:StakeholderTrust))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:enables ai:AlgorithmicTransparency))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:enables ai:TrustworthyAI))
Implementation Relationships
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:implements ai:IEEE7001-2021))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:implements ai:EUAIAct))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:implements ai:NISTAIRiskManagementFramework))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:implements ai:ISOiEC42001))
Reduction Relationships
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:reducesTo ai:PerformanceMetrics))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:reducesTo ai:ExplainabilityMetrics))
SubClassOf(ai:AlgorithmicTransparencyIndex
ObjectSomeValuesFrom(ai:reducesTo ai:GovernanceInstrument))
About
The Algorithmic Transparency Index (ATI) is a structured measurement framework that translates the broad normative commitment to algorithmic transparency into scored, auditable dimensions applicable to any Machine Learning Model or Automated Decision System. Its conceptual lineage runs through three converging traditions: information law’s requirement that public decisions be reasoned and reviewable (traceable to administrative law principles dating to the 1940s); the AI ethics literature’s operationalisation of transparency as a technical requirement rather than a political preference (Floridi et al. 2018; Jobin et al. 2019); and the governance standardisation agenda that crystallised in IEEE 7001-2021, the NIST AI RMF (2023), and IEC 42001 (2023). The ATI provides the measurement layer that sits between these normative commitments and the compliance evidence that organisations must produce for regulators and stakeholders.
The most influential external ATI in the current landscape is the Stanford Foundation Model Transparency Index (FMTI), first published in 2023 and updated annually. The 2025 FMTI edition evaluated 14 major foundation-model developers — including OpenAI, Google, Meta, Anthropic, Mistral, xAI, and Alibaba — across 100 transparency indicators spanning upstream information (training data, compute, labour), model characteristics (capabilities, limitations, risks), and downstream information (deployment policies, usage data, impact). The 2025 average score of 40 out of 100 (down from 58 in 2024, reflecting more demanding indicators) revealed persistent opacity in training data disclosure and post-deployment usage tracking. IBM scored highest at 95; xAI and Midjourney scored lowest at 14. The FMTI explicitly maps its indicators to the General-Purpose AI Code of Practice requirements under EU AI Act Articles 53 and 55, making it a practical compliance evaluation instrument as well as an external accountability benchmark.
The IEEE 7001-2021 standard provides a complementary ATI architecture specifically designed for autonomous systems. It defines six transparency levels (0–5) applied across five stakeholder categories: the operating organisation, users, affected third parties, reviewing bodies, and accident investigators. Level 0 represents no transparency; Level 5 represents full insight into all decision processes in human-interpretable form. The standard requires that the level target be specified at system design time and verified by measurable, testable criteria, making it the most technically precise ATI specification available. The AI Standards Hub (BSI/NPL/Turing Institute) provides UK-specific implementation guidance for IEEE 7001-2021 adoption.
Internally deployed ATIs typically integrate with an organisation’s AI Governance Framework as a scored dashboard updated at each model release cycle, each annual audit, and upon significant model updates. Common implementation architectures combine automated measurement (SHAP fidelity scores, bias gap statistics, audit log completeness checks) with human-curated documentation scores (model card completeness, dataset datasheet coverage, stakeholder notification records). The composite ATI score provides executives and audit committees with a single governance maturity signal while the sub-dimension breakdown directs engineering and compliance teams toward specific remediation priorities.
Components / Architecture
Process Transparency Dimension
Process transparency measures the degree to which the system’s development lifecycle is documented and accessible. Indicators include: existence and completeness of a model development specification; version control log coverage linking each deployed model to training artifacts; red-teaming exercise documentation (scope, findings, remediation); ethics board review records for high-risk systems; and change management logs for post-deployment model updates. Scoring typically assigns 0–3 points per indicator, with partial credit for documentation that exists but is not publicly accessible. Process transparency is a prerequisite for independent Algorithmic Auditing because auditors require evidence that documented procedures were actually followed.
Data Transparency Dimension
Data transparency covers Data Provenance disclosure quality. Indicators assess whether a datasheet for the training dataset exists (per Gebru et al. 2021 specification); whether demographic coverage statistics are reported; whether consent documentation and licensing records are maintained; whether bias survey results covering protected characteristics are available; and whether training data filtering and preprocessing steps are documented. The EU AI Act’s GPAI transparency obligations (Article 53(1)(d)), which became applicable from 2 August 2025, require providers to make training data summaries publicly available, creating a legal compliance anchor for this dimension. The 2025 FMTI found training data disclosure to be the most opaque dimension across all evaluated developers, with the majority scoring 0 on data transparency indicators.
Model Documentation Dimension
Model documentation scores the completeness and accessibility of model cards, system cards, and factsheets. A fully scored model card (per Mitchell et al. 2019, updated for EU AI Act alignment in 2024) covers: intended use, out-of-scope uses, training factors and decisions, evaluation data description, performance metrics across subpopulations, ethical considerations, known limitations, and caveats and recommendations. Model cards are now required artefacts for EU AI Act Annex III high-risk system technical documentation (Article 11) and are the primary transparency instrument under the UK ATRS for government AI deployments.
Decision Transparency Dimension
Decision transparency assesses the degree to which individual algorithmic decisions are explained to affected parties and logged for audit purposes. This dimension directly maps to GDPR Article 22(3) (meaningful information about the logic involved in automated decisions), EU AI Act Article 13 (transparency for users of high-risk AI), and the UK’s right to explanation under the Data Protection Act 2018. Indicators include: whether a natural-language explanation of decisions is generated and delivered to affected parties; whether a per-decision audit record capturing inputs, outputs, model version, and timestamp is maintained; whether a contestation mechanism exists and its response rate is documented; and whether Decision Logging architecture enables post-hoc reconstruction of any past decision.
Explainability Metrics Dimension
Explainability Metrics quantify the faithfulness and stability of interpretability methods applied to model outputs. SHAP (SHapley Additive exPlanations) fidelity scores measure how accurately SHAP attributions reconstruct model predictions when given access only to the attributed features. LIME stability scores measure whether repeated application of LIME to the same input produces consistent feature importance rankings. Attention attribution faithfulness (for transformer-based models) assesses whether high-attention tokens are genuinely causally important to the output. The OpenXAI benchmark (2023) and M-4 benchmark (2024) provide standardised evaluation protocols. A well-formed ATI records the specific explainability tooling used, the model classes to which it is applied, the evaluation metrics and thresholds, and the cadence of re-evaluation.
Bias Detection Metrics Dimension
Bias Detection Metrics score the comprehensiveness and rigour of fairness evaluation. Canonical fairness metrics include: statistical parity difference (difference in positive prediction rates across groups); equal opportunity difference (difference in true-positive rates); average odds difference; disparate impact ratio; and calibration error disaggregated by group. A fully scored ATI documents: which fairness metrics are measured; for which protected characteristics; using which evaluation dataset; at what threshold values adverse findings trigger remediation; and what the measured values were at last evaluation. IBM AI Fairness 360, Microsoft Fairlearn, and Google’s What-If Tool are the principal open-source toolkits used for bias metric computation, each of which exposes machine-readable metric outputs suitable for automated ATI scoring.
Audit Trail Completeness Dimension
Audit Trail completeness scores the degree to which tamper-evident logging covers all significant model decisions and governance events. A complete audit trail captures: decision inputs (hashed, not raw sensitive data), decision outputs, model version identifier, decision timestamp, the identity of the deploying system and operator, and any human override events. For EU AI Act Annex III systems, post-market surveillance logs must be retained for ten years and must be accessible to national market surveillance authorities on request. The ATI scores audit trail completeness against these retention requirements, the integrity mechanism (digital signature, hash chain, or immutable log service), and the accessibility and query performance of the log store. Blockchain-anchored audit logs are explored as a mechanism for providing external verifiability of log integrity without requiring trust in the deploying organisation’s internal infrastructure.
Disclosure Requirements Compliance Dimension
Disclosure Requirements compliance scores the degree to which the system satisfies applicable mandatory transparency obligations. For EU-deployed systems, this includes: EU AI Act Article 50 (informing users when interacting with AI chatbots and deepfakes); Article 13 (transparency for users of high-risk AI systems); Article 53 GPAI transparency requirements; and DSA Article 27 (transparency reporting for Recommendation Systems on very large online platforms). For UK government systems, it includes ATRS compliance (59 records published as of mid-2025, with mandatory coverage expanding). For financial services, it includes FCA consumer duty requirements for explainability of AI-driven credit decisions. The ATI aggregates compliance status across applicable instruments into a single compliance coverage percentage.
Use Cases / Major Families
Foundation Model Benchmarking: The Stanford FMTI exemplifies external ATI as a civil-society accountability instrument. By scoring 14 major foundation-model developers across 100 standardised indicators, it enables comparative transparency evaluation, tracks industry trends (the 2024–2025 score decline from 58 to 40 revealed increasing opacity despite regulatory pressure), and provides a credible external reference that regulators, investors, and civil society organisations use to evaluate AI developer conduct. The 2025 edition maps indicators to the EU AI Act GPAI Code of Practice, providing a practical pre-compliance assessment tool for developers negotiating GPAI obligations.
Regulatory Compliance Evidence: National market surveillance authorities under the EU AI Act require evidence that high-risk AI system providers have satisfied Article 9 (risk management), Article 11 (technical documentation), and Article 13 (transparency to users) obligations before granting market access or CE marking. A well-documented internal ATI provides the structured evidence base for conformity assessments, both for self-assessment and for third-party notified body review. ISO/IEC 42001 certification auditors increasingly treat ATI scores as proxy indicators of governance maturity.
Government AI Deployment: The UK Algorithmic Transparency Recording Standard (ATRS), made mandatory for all central government departments and Arms Length Bodies in 2025, functions as a government-sector ATI. Each ATRS record documents the tool’s purpose, data inputs, decision significance, affected populations, oversight mechanisms, and accuracy. The 59 published ATRS records as of mid-2025 represent a public ATI dataset for government algorithmic tools, enabling civil society and parliamentary scrutiny. The OECD’s 2025 report on algorithmic transparency in the public sector drew on ATRS and analogous registers in France (INVENTAIRE), Spain (the AESIA algorithm register), and Canada (DAID) to compare government ATI implementation approaches.
Platform Algorithmic Accountability: Recommendation Systems and Content Moderation Systems operated by very large online platforms (VLOPs) under the Digital Services Act must provide researchers and the European Centre for Algorithmic Transparency (ECAT) with system access sufficient for independent audits. ECAT, launched in April 2023 as part of the Joint Research Centre, applies ATI-equivalent scoring to VLOP algorithmic systems in the context of DSA Article 34 risk assessment reviews. The DSA biannual transparency reports published by Meta, Google, TikTok, and X provide a mandatory ATI-adjacent public disclosure dataset for Content Moderation System performance and Recommendation System parameters.
Healthcare AI: NHS England’s AI and Digital Regulations Safety Scheme (ADSS) applies ATI concepts to clinical AI tools, requiring model cards, bias evaluations disaggregated by NHS-relevant demographic groups (ethnicity, deprivation quintile, sex, age), and evidence of post-deployment monitoring. The MHRA’s Good Machine Learning Practice (GMLP) guidelines for AI as a medical device require software lifecycle transparency documentation equivalent to an ATI process transparency score.
Academic Context
The ATI concept draws on several intellectual traditions. The foundational transparency-as-accountability argument was articulated by Diakopoulos and Koliska (2017), who distinguished between process transparency (how algorithmic systems are built), data transparency (what data they consume), and algorithmic transparency (what logic they apply). Crawford and Schultz (2014) in “Big Data and Due Process” established the legal accountability grounding in constitutional due process norms. Mittelstadt et al. (2016) in “The Ethics of Algorithms” provided the most cited taxonomy of algorithmic transparency failures, identifying harms arising from inconclusive evidence, inscrutable evidence, misguided evidence, unfair outcomes, and transformative effects. Wachter, Mittelstadt, and Russell (2017) proposed counterfactual explanations as the legally minimal form of algorithmic transparency consistent with GDPR Article 22.
The fairness metrics dimension draws on the landmark ACM FAccT (Fairness, Accountability, and Transparency) research programme, launched in 2018, which has produced the principal technical vocabulary for bias measurement: Hardt, Price, and Srebro (2016) on equality of opportunity; Chouldechova (2017) on the mathematical impossibility of simultaneously satisfying competing fairness criteria; and Barocas, Hardt, and Narayanan’s textbook “Fairness and Machine Learning” (2019, updated 2023) as the field’s canonical reference. The SHAP explainability framework (Lundberg and Lee 2017) and LIME (Ribeiro, Singh, and Guestrin 2016) provide the principal technical instruments for the explainability metrics dimension.
The IEEE 7001-2021 standard emerged from the P7001 working group chaired by Alan Winfield (University of the West of England) — notably including significant UK academic leadership — and represents the most technically complete formalisation of the ATI concept to date, providing the six-level transparency scale that distinguishes the standard from less structured governance instruments. The Foundation Model Transparency Index (Bommasani et al. 2023, 2024, 2025) has rapidly become the dominant external ATI reference for large model providers.
Current Landscape (2026)
The period 2024–2026 has produced a convergence of mandatory and voluntary ATI obligations that is transforming industry practice. The EU AI Act’s transparency provisions entered full applicability on 2 August 2026, following a phased implementation that began with prohibited practices (February 2025) and GPAI obligations (August 2025). Annex III high-risk system providers must now demonstrate Article 13 transparency compliance, Article 11 technical documentation completeness, and post-market surveillance log retention as part of conformity assessments — all of which map directly to ATI sub-dimensions. The European AI Office, established within DG CONNECT in 2024, oversees GPAI model transparency compliance and can commission evaluations equivalent to external ATI scoring.
The GPAI Code of Practice, finalised on 10 July 2025, introduced a Model Documentation Form — a semi-standardised transparency template for general-purpose AI model providers that aligns with FMTI indicators and constitutes the first binding-adjacent transparency standard for frontier models. Amazon, Anthropic, Google, IBM, Mistral, OpenAI, Writer, and xAI were among signatories as of publication. The 2025 FMTI update explicitly maps its 100 indicators to Code of Practice requirements, enabling developers to use the FMTI as a pre-compliance self-assessment.
In the UK, the ATRS mandatory expansion to all central government from 2025 represents the most institutionally significant government ATI implementation globally. The 125 ATRS records published by mid-2025 cover a growing range of government algorithmic tools, from benefit eligibility assessors to road network management systems. The OECD’s 2025 comparative report on algorithmic transparency in the public sector benchmarked ATRS against analogous registers in France, Spain, Canada, and the Netherlands, identifying the ATRS’s structured template and centralised public repository as international best practice. UK government bodies including DSIT, GDS, and the Centre for Data Ethics and Innovation are co-developing an updated ATRS v2 for release in 2026 that incorporates generative AI disclosure provisions not present in the 2021 original.
In civil society, Access Now’s Transparency Reporting Index tracks voluntary platform transparency disclosure across 37 major technology companies and 11 thematic dimensions, providing a regularly updated external ATI equivalent for the global platform ecosystem. AlgorithmWatch’s ADM Watch project monitors automated decision-making deployments in Europe, collecting transparency information beyond what platforms formally disclose, using the ATI concept as its analytical frame.
UK Context
The United Kingdom has a particularly rich ATI ecosystem spanning government, civil society, academia, and standards bodies. The Centre for Data Ethics and Innovation (CDEI), now within DSIT’s AI Policy Directorate, developed the ATRS from inception in 2021 and has been the primary driver of its mandatory extension. The Government Digital Service (GDS) maintains the public ATRS repository on GOV.UK, which functions as the national government ATI database. The UK AI Standards Hub (BSI, NPL, Alan Turing Institute) provides implementation guidance for IEEE 7001-2021 adoption in UK contexts and represents UK interests in ISO/IEC JTC 1/SC 42 standardisation work on AI transparency metrics.
In academic research, Edinburgh’s Bayes Centre and the University of Edinburgh’s School of Informatics have contributed to explainability metric research underpinning ATI, particularly through work by Amos Storkey’s group on uncertainty quantification and out-of-distribution detection — both relevant to honest performance disclosure. The Alan Turing Institute’s Fairness, Transparency, and Privacy interest group, involving researchers from institutions including Oxford, Cambridge, UCL, and Manchester, has produced applied guidance on operationalising transparency metrics for public-sector AI, directly informing ATRS design. UCL’s Responsible Technology Institute has focused on ATI implementation challenges in healthcare AI, producing guidance used by NHS England’s ADSS. Imperial College London’s Data Science Institute has applied ATI concepts to financial services AI, contributing to FCA guidance on explainability for algorithmic credit decisions.
In Northern England, Manchester plays a particularly significant role. The University of Manchester’s Alliance Manchester Business School has studied ATI implementation challenges in enterprise settings, finding (in a 2024 study of 45 UK organisations) that organisations with structured ATI processes reported 31% fewer post-deployment bias incidents than those relying on ad-hoc disclosure practices. Manchester Metropolitan University’s Centre for Digital Ethics has applied ATI concepts in social housing and local government contexts, examining how transparency obligations interact with algorithmic decision rights for vulnerable populations. Leeds City Council’s data analytics programme was one of the early ATRS adopters and has published five ATRS records covering tools used in housing needs assessments and environmental enforcement. Sheffield Hallam University’s Advanced Wellbeing Research Centre has applied ATI to healthcare triage AI, examining the intersection of NHS obligations and ATRS requirements.
Future Directions (2026–2030)
Several trajectories will shape ATI development through 2030. First, dynamic and real-time ATI scoring is emerging to replace the current predominantly static, audit-point-in-time approach. As model behaviour can shift between audit cycles through distribution drift, data pipeline changes, or adversarial inputs, continuous ATI monitoring — with automated alerts when sub-dimension scores fall below threshold — is becoming a design requirement for high-risk system governance. Second, interoperability between ATI frameworks — FMTI, IEEE 7001, ATRS, DSA audit frameworks — remains poor, creating compliance overhead for multinational operators who must satisfy multiple overlapping transparency requirements. The OECD’s 2025 comparative report calls for development of a Common Transparency Evaluation Framework (CTEF) that maps equivalent indicators across instruments, a project expected to advance through 2027. Third, generative AI transparency creates novel ATI challenges: the probabilistic, context-sensitive nature of LLM outputs makes decision logging more complex than for traditional classifiers; training data disclosure obligations conflict with copyright and privacy constraints; and the GPAI Code of Practice’s transparency provisions are still being operationalised. Fourth, ATI automation using AI to evaluate AI transparency — processing model cards, documentation, and log samples with language models to produce structured ATI scores — is a research frontier explored by the 2024 AI Transparency Atlas project (Zafar et al. 2024), which demonstrated automated model-card evaluation with 87% agreement with human expert ratings.
Research & Literature
- Bommasani, R., et al. (2023). “The Foundation Model Transparency Index.” arXiv:2310.12941. https://arxiv.org/abs/2310.12941
- Bommasani, R., et al. (2024). “The 2024 Foundation Model Transparency Index.” arXiv:2407.12929. https://arxiv.org/abs/2407.12929
- Wan, A., et al. (2025). “The 2025 Foundation Model Transparency Index.” arXiv:2512.10169. https://arxiv.org/abs/2512.10169
- IEEE (2021). IEEE 7001-2021: Standard for Transparency of Autonomous Systems. IEEE. https://ieeexplore.ieee.org/document/9726144
- Mitchell, M., et al. (2019). “Model Cards for Model Reporting.” Proceedings of the 2019 ACM FAccT, 220–229. https://doi.org/10.1145/3287560.3287596
- Gebru, T., et al. (2021). “Datasheets for Datasets.” Communications of the ACM, 64(12), 86–92. https://doi.org/10.1145/3458723
- Diakopoulos, N., & Koliska, M. (2017). “Algorithmic Transparency in the News Media.” Digital Journalism, 5(7), 809–828. https://doi.org/10.1080/21670811.2016.1208053
- Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). “The Ethics of Algorithms: Mapping the Debate.” Big Data & Society, 3(2). https://doi.org/10.1177/2053951716679679
- Wachter, S., Mittelstadt, B., & Russell, C. (2017). “Counterfactual Explanations without Opening the Black Box.” Harvard Journal of Law and Technology, 31(2), 841–887.
- Lundberg, S. M., & Lee, S.-I. (2017). “A Unified Approach to Interpreting Model Predictions.” Advances in NeurIPS, 30.
- Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You? Explaining the Predictions of Any Classifier.” Proceedings of KDD 2016, 1135–1144.
- Hardt, M., Price, E., & Srebro, N. (2016). “Equality of Opportunity in Supervised Learning.” Advances in NeurIPS, 29.
- Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press. https://fairmlbook.org
- Chouldechova, A. (2017). “Fair Prediction with Disparate Impact.” Big Data, 5(2), 153–163.
- Floridi, L., Cowls, J., Beltrametti, M., et al. (2018). “AI4People — An Ethical Framework for a Good AI Society.” Minds and Machines, 28(4), 689–707.
- Jobin, A., Ienca, M., & Vayena, E. (2019). “The Global Landscape of AI Ethics Guidelines.” Nature Machine Intelligence, 1(9), 389–399.
- Crawford, K., & Schultz, J. (2014). “Big Data and Due Process.” Boston College Law Review, 55(1), 93–128.
- European Parliament (2024). EU AI Act. Regulation (EU) 2024/1689. Official Journal of the EU. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689
- NIST (2023). AI Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1
- ISO/IEC JTC 1/SC 42 (2023). ISO/IEC 42001:2023 — AI Management System. https://www.iso.org/standard/81234.html
- European Commission (2025). General-Purpose AI Code of Practice. EC AI Office. https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
- CDEI/GDS (2021, updated 2025). Algorithmic Transparency Recording Standard. GOV.UK. https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub
- Zafar, M. B., et al. (2024). “AI Transparency Atlas: Framework, Scoring, and Real-Time Model Card Evaluation Pipeline.” arXiv:2512.12443. https://arxiv.org/pdf/2512.12443
- OECD (2025). Algorithmic Transparency in the Public Sector: A State-of-the-Art Report. OECD AI Policy Observatory. https://wp.oecd.ai/app/uploads/2025/05/algorithmic-transparency-in-the-public-sector.pdf
- Winfield, A. F. T., & Jirotka, M. (2018). “Ethical Governance Is Essential to Building Trust in Robotics and Artificial Intelligence.” Philosophical Transactions of the Royal Society A, 376(2133).
- Raji, I. D., & Yang, J. (2019). “ABOUT ML: Annotation and Benchmarking on Understanding and Transparency of Machine Learning Lifecycles.” Proceedings of the Workshop on Challenges and Opportunities for Privacy and Fairness in Social Science Research.
- Selbst, A. D., & Barocas, S. (2018). “The Intuitive Appeal of Explainable Machines.” Fordham Law Review, 87(3), 1085–1139.
- AlgorithmWatch (2024). ADM Watch: Tracking Automated Decision-Making in Europe. AlgorithmWatch. https://algorithmwatch.org/en/adm-watch/