ISO/IEC TR 24028: AI Trustworthiness Overview, Bias, Transparency & Risk
What Is ISO/IEC TR 24028?
As a Technical Report rather than a full standard, ISO/IEC TR 24028 does not establish mandatory requirements - it provides an authoritative reference document that maps the AI trustworthiness landscape, defines key concepts, identifies known risks and failure modes and describes approaches for addressing those risks. It is one of the foundational documents in the ISO/IEC JTC 1/SC 42 AI standards family, intended to be read alongside ISO/IEC 42001 for AI management systems, ISO/IEC 23894 for AI risk management and ISO/IEC 24029 for AI robustness assessment.
The report addresses the full scope of AI trustworthiness concerns - from data quality and model bias through system robustness and security to transparency, explainability, human oversight and societal impact. It is relevant to any organization that develops AI systems, deploys AI solutions, or governs the use of AI in products, services, or operational processes - and provides the conceptual foundation that informed the development of the AI Act in the European Union and similar AI governance frameworks globally.
ISO/IEC TR 24028 helps organizations understand AI trustworthiness by addressing bias, transparency, robustness, privacy, security and risk management - Pacific Certifications
AI Trustworthiness Concepts
Reliability
A reliable AI system consistently produces correct and expected outputs across the range of inputs and conditions it is designed to handle. Reliability encompasses both functional correctness - producing the right output for a given input - and operational consistency - maintaining that correctness across time, environments and edge cases.
Reliability failures in AI systems frequently arise from training data that does not adequately represent the deployment environment, causing the model to perform well in testing but unpredictably in production.
Safety
AI safety refers to the property of an AI system that ensures it does not cause harm - to individuals, organizations, or society - whether through malfunction, misuse, or unintended consequences.
Safety considerations are particularly critical for AI systems deployed in high-stakes contexts such as healthcare, autonomous vehicles, industrial control and financial decision-making, where a system failure can have severe real-world consequences.
Security
AI systems are subject to a distinct set of security threats that do not apply to conventional software - including adversarial attacks that manipulate model inputs to produce incorrect outputs, model inversion attacks that extract training data from deployed models and data poisoning attacks that corrupt model behavior by introducing malicious training samples.
AI security encompasses the measures taken to protect AI systems against these threats throughout their lifecycle.
Resilience
AI system resilience refers to the ability to maintain acceptable performance in the face of unexpected inputs, operational disruptions, distribution shifts and adversarial conditions.
A resilient AI system degrades gracefully rather than failing catastrophically when encountering conditions outside its training distribution.
Accountability
Accountability requires that there are identifiable parties - individuals or organizations - who are responsible for the behavior and outcomes of an AI system and that those parties can be held answerable when the system causes harm or operates outside its intended boundaries.
Accountability mechanisms include audit trails, decision logging, governance documentation and human oversight procedures.
Practical Tip: Treat AI trustworthiness as a full governance responsibility, not only as a technical model performance issue.
Transparency and Explainability
Transparency refers to the degree to which information about an AI system - its purpose, capabilities, limitations, training data, performance characteristics and decision-making processes - is available to relevant stakeholders.
Transparency operates at multiple levels: system-level transparency covers what the AI system does and under what conditions; process-level transparency covers how the system was developed, validated and deployed; and decision-level transparency covers how the system arrived at a specific output for a specific input.Explainability refers specifically to the ability to provide understandable reasons for an AI system's outputs - making it possible for users, affected individuals and oversight bodies to understand why the system produced a particular result.
Explainability is particularly important for AI systems that make or inform consequential decisions affecting individuals - such as credit scoring, medical diagnosis support, criminal risk assessment and employment screening.
ISO/IEC TR 24028 recognizes that explainability requirements vary by context - a radiologist reviewing an AI-assisted diagnostic suggestion needs a different level of explanation than a regulator auditing a financial institution's AI governance program.
The report identifies several dimensions of explainability:
Local explainability: Explaining a specific individual prediction or decision - why did the model produce this output for this particular input?
Global explainability: Explaining the overall behavior of the model - what features drive its predictions across the entire input space?
Contrastive explainability: Explaining why the model produced one output rather than another - what would need to change for the outcome to be different?
Writer’s view: Transparent AI systems are easier to trust when users can understand what the system does, why it matters and where its limits are.
Bias and Robustness
AI Bias
Bias in AI systems refers to systematic errors or unfair outcomes that arise from flaws in training data, model design, or deployment context. ISO/IEC TR 24028 identifies several categories of bias:
Data bias: Arises when training data does not accurately represent the population or phenomenon the model is intended to address - including historical biases encoded in past human decisions, under-representation of certain demographic groups and label noise introduced by biased human annotators.
Algorithmic bias: Arises from model architecture choices, optimization objectives, or feature selection decisions that cause the model to systematically disadvantage certain groups or produce systematically skewed outputs in certain contexts.
Interaction bias: Arises from the way users interact with AI systems - for example, when user feedback used to train or update a model reflects the biases of the user population rather than objective ground truth.
Deployment bias: Arises when a model trained for one context is deployed in a different context where its assumptions no longer hold - producing systematically incorrect outputs for the new deployment population.
Addressing AI bias requires action at multiple stages of the AI development lifecycle - from data collection and labeling through model training and evaluation to deployment monitoring and post-deployment audit. ISO/IEC TR 24028 identifies fairness metrics, bias testing methodologies and data diversity requirements as key technical approaches for bias mitigation.
Strength
AI strength refers to the ability of an AI system to maintain correct and consistent behavior when faced with variations in input data, distribution shifts between training and deployment environments and deliberate adversarial perturbations. A robust AI system produces reliable outputs not only on clean, well-formed inputs representative of its training data, but also on noisy, corrupted, or out-of-distribution inputs that arise in real-world deployment. ISO/IEC TR 24028 distinguishes between natural robustness - resistance to naturally occurring input variations - and adversarial robustness - resistance to deliberately crafted inputs designed to cause model failure.
Practical Tip: AI systems should be tested against changing inputs, edge cases, failure modes and misuse scenarios before deployment.
Risks and Failure Modes
Data-Related Risks
Training data that is incomplete, unrepresentative, mislabeled, or contaminated
Privacy risks arising from personal data used in training or accessible through model outputs
Data drift - changes in the statistical distribution of production data over time that degrade model performance
Model-Related Risks
Overfitting - model performs well on training data but fails to generalize to new data
Underfitting - model is too simple to capture the patterns required for the intended application
Model degradation over time as the deployment environment diverges from the training environment
Operational Risks
Deployment in contexts or for purposes beyond the validated scope of the model
Failure to maintain human oversight for high-stakes AI-assisted decisions
Inadequate monitoring of production AI system performance
Security Risks
Adversarial attacks - crafted inputs designed to cause incorrect model outputs
Data poisoning - introduction of malicious training samples to corrupt model behavior
Model theft - extraction of model parameters or behavior through repeated querying
Societal Risks
Amplification of existing societal biases at scale through automated decision-making
Erosion of privacy through large-scale inference of personal attributes from aggregated data
Opacity of consequential decisions made or informed by AI systems
Identify AI risks early by reviewing data quality, model limitations, operational misuse, security threats and societal impact together.
Relation to ISO 42001 and ISO 23894
Final Remark: ISO/IEC TR 24028 explains AI trustworthiness concepts, while ISO/IEC 42001 and ISO/IEC 23894 help organizations turn them into governance and risk controls.
Implementation Examples
Financial Services - Credit Scoring AI
A bank deploying an AI-assisted credit scoring model uses ISO/IEC TR 24028 as its reference framework for addressing trustworthiness obligations under applicable financial services regulations. The bias assessment covers the training data for demographic representation gaps and applies fairness metrics to evaluate differential approval rates across protected demographic groups.
Explainability is implemented at the decision level - providing applicants with understandable reasons for credit decisions as required by consumer credit regulations. Model performance monitoring is implemented in production to detect data drift and performance degradation over time.
Healthcare - AI Diagnostic Support
A medical technology company developing an AI-assisted radiology tool applies ISO/IEC TR 24028 to govern its trustworthiness program alongside IEC 62304 for software lifecycle and ISO 13485 for medical device quality management. Transparency documentation covers the training dataset composition, validation performance across demographic subgroups and known limitations.
Robustness testing covers performance on images from different scanner manufacturers and image quality levels. Human oversight is designed into the clinical workflow - the AI output is presented as a decision support recommendation with confidence indicators, with the radiologist retaining final diagnostic authority.
Public Sector - Benefits Eligibility Assessment
A government agency using AI to support benefits eligibility assessment applies ISO/IEC TR 24028 to structure its algorithmic accountability program. A bias audit is conducted covering historical decision data used for training, with demographic parity testing applied across age, gender and geographic groups.
Decision logging is implemented to ensure that every AI-informed eligibility determination can be reviewed and explained in response to an appeal. Human review is mandated for all adverse decisions, with the AI output serving as a screening and prioritization tool rather than an autonomous decision-maker.
Practical Tip: Use implementation examples to show how AI trustworthiness controls apply differently in finance, healthcare, public sector and other high-impact environments.
ISO/IEC TR 24028 Certification Cost
For organizations pursuing ISO/IEC 42001 certification as the primary management system certification alongside ISO/IEC TR 24028 implementation, audit cost is determined by employee count, number of AI systems in scope and the complexity of the AI lifecycle activities covered. A focused AI product company with a single primary AI application will have a modest audit investment. A large enterprise deploying multiple AI systems across diverse business functions - or an AI platform provider serving enterprise clients - will require proportionally more audit days. Where ISO/IEC 27001 and ISO/IEC 27701 are pursued alongside ISO/IEC 42001 to address the security and privacy dimensions of AI trustworthiness, integrated audits reduce total audit days and provide better value than separate certifications. Pacific Certifications provides transparent, fixed-fee proposals so your organization has full cost visibility before the process begins.
Cost planning should consider AI system complexity, number of use cases, governance maturity, employee count and whether ISO/IEC 42001, ISO/IEC 27001 or ISO/IEC 27701 is included.
ISO/IEC TR 24028 Certification Timeline
This includes 2 to 4 weeks for AI system inventory and trustworthiness gap analysis against ISO/IEC TR 24028 concepts, 4 to 8 weeks for bias assessment, explainability implementation, robustness testing and accountability documentation and 2 to 4 weeks for governance policy development and human oversight procedure design. Organizations with large, complex AI portfolios spanning multiple use cases, data modalities and deployment environments should plan for the longer end of this range.
For organizations pursuing ISO/IEC 42001 certification in parallel with ISO/IEC TR 24028 implementation, the full certification process runs approximately 4 to 6 months from initial gap analysis through certificate issuance. Where ISO/IEC 27001 is added to address the security dimensions of AI trustworthiness, the timeline extends to 6 to 9 months. Where ISO/IEC 27701 is also included to address privacy obligations, an integrated program covering all three standards typically requires 6 to 12 months depending on organizational complexity. The ISO/IEC TR 24028 implementation work feeds directly into the ISO/IEC 42001 AIMS evidence base - particularly the risk assessment, transparency and human oversight requirements - reducing duplication of effort across the combined program. Assigning dedicated AI governance ownership, maintaining contemporaneous documentation from the outset of AI system development and conducting a structured internal audit before the Stage 2 certification assessment are the most effective ways to keep the program on track.
A Practical Tip from Pacific Certifications: Organizations can avoid delays by preparing AI inventories, bias assessments, explainability records, robustness testing and governance evidence early.
How Pacific Certifications Can Help?
Accredited by ABIS, Pacific Certifications conducts impartial, evidence-based audits against applicable ISO standards in full conformance with ISO/IEC 17021.
Our services for organizations implementing AI trustworthiness programs include:
Independent certification audits for ISO/IEC 42001, ISO/IEC 27001, ISO/IEC 27701, ISO 9001 and ISO 22301
Stage 1 and Stage 2 audit execution across single and multi-site AI organizations
Clear, transparent audit reports with conformity findings and certification decisions
Issuance of internationally recognized ISO certificates upon successful audit completion
Annual surveillance and triennial recertification audits to maintain certificate validity
Pacific Certifications does not provide consultancy - our role is strictly that of an independent auditor, ensuring your certificate carries full credibility with enterprise clients, regulators and AI governance authorities in every market you operate in.
Contact Us
To get started with your AI trustworthiness certification program or initiate your audit, contact us at support@pacificcert.com or +91-8595603096.
For training programs, contact us at trainings@pacificcert.com.
Also Read: ISO/IEC 23894 and ISO 42001 responsible AI governance
