ISO/IEC TS 6254:2025

Technical
Specification
ISO/IEC TS 6254
First edition
Information technology — Artificial
2025-09
intelligence — Objectives and
approaches for explainability
and interpretability of machine
learning (ML) models and artificial
intelligence (AI) systems
Technologies de l'information — Intelligence artificielle —
Objectifs et approches pour l'explicabilité et l'interprétabilité
des modèles d'apprentissage automatique (AA) et des systèmes
d'intelligence artificielle (IA)
Reference number
© ISO/IEC 2025
All rights reserved. Unless otherwise specified, or required in the context of its implementation, no part of this publication may
be reproduced or utilized otherwise in any form or by any means, electronic or mechanical, including photocopying, or posting on
the internet or an intranet, without prior written permission. Permission can be requested from either ISO at the address below
or ISO’s member body in the country of the requester.
ISO copyright office
CP 401 • Ch. de Blandonnet 8
CH-1214 Vernier, Geneva
Phone: +41 22 749 01 11
Email: copyright@iso.org
Website: www.iso.org
Published in Switzerland
© ISO/IEC 2025 – All rights reserved
ii
Contents Page
Foreword .v
Introduction .vi
1 Scope . 1
2 Normative references . 1
3 Terms and definitions . 1
4 Symbols and abbreviated terms. 5
5 Overview . 6
6 Stakeholders’ objectives . 6
6.1 General .6
6.2 AI user .7
6.3 AI developer .7
6.4 AI product or service provider .7
6.5 AI platform provider .8
6.6 AI system integrator.8
6.7 Data provider .8
6.8 AI evaluator .8
6.9 AI auditor .8
6.10 AI subject .8
6.11 Relevant authorities . .8
6.11.1 Policy makers .8
6.11.2 Regulators .8
6.11.3 Other authorities .9
7 Explainability considerations throughout the AI system life cycle . 9
7.1 General .9
7.2 Inception .10
7.3 Design and development .10
7.3.1 General .10
7.3.2 Development of the explainability component .10
7.3.3 Explainability’s contribution to development .11
7.4 Verification and validation .11
7.4.1 General .11
7.4.2 Evaluation of the explainability component .11
7.4.3 Explainability’s contribution to evaluation . 13
7.5 Deployment .14
7.5.1 General .14
7.5.2 Deployment of the explainability component .14
7.5.3 Explainability’s contribution to deployment .14
7.6 Operation and monitoring . .14
7.7 Continuous validation .14
7.8 Re-evaluation .14
7.9 Retirement . . 15
8 Property taxonomy of explainability methods and approaches .15
8.1 General . 15
8.2 Properties of explanation needs .16
8.2.1 General .16
8.2.2 Expertise profile of the targeted audience .16
8.2.3 Frame activity of interpretation or explanation.17
8.2.4 Scope of information .17
8.2.5 Completeness .17
8.2.6 Depth . .18
8.2.7 Reasoning path .18
8.2.8 Implicit and explicit explanations .19

© ISO/IEC 2025 – All rights reserved
iii
8.3 Forms of explanation .19
8.3.1 General .19
8.3.2 Numeric .19
8.3.3 Visual .19
8.3.4 Textual . 20
8.3.5 Structured . 20
8.3.6 Example-based . 20
8.3.7 Interactive exploration tools . 20
8.4 Technical approaches towards explainability . 20
8.4.1 General . 20
8.4.2 Empirical analysis .21
8.4.3 Post hoc interpretation .21
8.4.4 Inherently interpretable components .21
8.4.5 Architecture- and task-driven explainability . 22
8.5 Technical constraints of the explainability method. 22
8.5.1 General . 22
8.5.2 Genericity of the method . . 22
8.5.3 Transparency requirements . 23
8.5.4 Display requirements . 23
9 Approaches and methods to explainability .23
9.1 General . 23
9.2 Empirical analysis methods . .24
9.2.1 General .24
9.2.2 Fine-grained evaluation . 25
9.2.3 Error analysis . 25
9.2.4 Analysis-oriented datasets . 25
9.2.5 Ablation . 26
9.2.6 Known trends . . 26
9.3 Post hoc methods .27
9.3.1 Local .27
9.3.2 Global .32
9.4 Inherently interpretable components . 36
9.4.1 General . 36
9.4.2 Legible models .37
9.4.3 Meaningful models . 39
9.4.4 Models with explicit knowledge .41
9.5 Architecture- and task-driven methods .43
9.5.1 General .43
9.5.2 Informative features .43
9.5.3 Rich and auxiliary inputs . 44
9.5.4 Multi-step processing. 44
9.5.5 Rich outputs.45
9.5.6 Rationale-based processing . 46
9.5.7 Rationale generation as auxiliary output . 46
9.6 Data explanation .47
Annex A (informative) Extent of explainability and interaction with related concepts .48
Annex B (informative) Illustration of methods’ properties .51
Annex C (informative) Concerns and limitations. 61
Bibliography .65

© ISO/IEC 2025 – All rights reserved
iv
Foreword
ISO (the International Organization for Standardization) and IEC (the International Electrotechnical
Commission) form the specialized system for worldwide standardization. National bodies that are
members of ISO or IEC participate in the development of International Standards through technical
committees established by the respective organization to deal with particular fields of technical activity.
ISO and IEC technical committees collaborate in fields of mutual interest. Other international organizations,
governmental and non-governmental, in liaison with ISO and IEC, also take part in the work.
The procedures used to develop this document and those intended for its further maintenance are described
in the ISO/IEC Directives, Part 1. In particular, the different approval criteria needed for the different types
of document should be noted. This document was drafted in accordance with the editorial rules of the ISO/
IEC Directives, Part 2 (see www.iso.org/directives or www.iec.ch/members_experts/refdocs).
ISO and IEC draw attention to the possibility that the implementation of this document may involve the
use of (a) patent(s). ISO and IEC take no position concerning the evidence, validity or applicability of any
claimed patent rights in respect thereof. As of the date of publication of this document, ISO and IEC had not
received notice of (a) patent(s) which may be required to implement this document. However, implementers
are cautioned that this may not represent the latest information, which may be obtained from the patent
database available at www.iso.org/patents and https://patents.iec.ch. ISO and IEC shall not be held
responsible for identifying any or all such patent rights.
Any trade name used in this document is information given for the convenience of users and does not
constitute an endorsement.
For an explanation of the voluntary nature of standards, the meaning of ISO specific terms and expressions
related to conformity assessment, as well as information about ISO's adherence to the World Trade
Organization (WTO) principles in the Technical Barriers to Trade (TBT) see www.iso.org/iso/foreword.html.
In the IEC, see www.iec.ch/understanding-standards.
This document was prepared by Joint Technical Committee ISO/IEC JTC 1, Information technology,
Subcommittee SC 42, Artificial intelligence.
Any feedback or questions on this document should be directed to the user’s national standards
body. A complete listing of these bodies can be found at www.iso.org/members.html and
www.iec.ch/national-committees.

© ISO/IEC 2025 – All rights reserved
v
Introduction
When AI systems are used to help make decisions that affect people’s lives, it is important that people
understand how those decisions are made. Achieving useful explanations of the behaviour of AI systems
and their components is a complex task. Industry and academia are actively exploring emerging methods for
enabling explainability, as well as scenarios and reasons why explainability can be required.
Due to the multitude of stakeholders and communities contributing to this effort, the field is suffering
from a certain terminological inconsistency. Most notably, the methods to provide such explanations of the
behaviour of an AI system are discussed under the banner of “explainability”, “interpretability”, (sometimes
even other terms like “transparency”), raising the question of how these terms relate to each other. This
document aims to provide practical guidance for stakeholders regarding compliance with regulatory
requirements labelled one way or another. With this goal in mind, it uses the umbrella term “explainability”
and provides a non-exhaustive taxonomy and list of approaches that stakeholders can use to comply with
regulatory requirements.
While the overarching goal of explainability is to evaluate the trustworthiness of AI systems, at different
stages of the AI system life cycle, diverse stakeholders can have more specific objectives in support of the
goal. To illustrate this point, several examples are provided. For developers, the goal can be improving the
safety, reliability and robustness of an AI system by making it easier to identify and fix bugs. For users,
explainability can help to decide how much to rely on an AI system by uncovering potential sources
or existence of unwanted bias or unfairness. For service providers, explainability can be essential for
demonstrating compliance with legal requirements. For policy makers, understanding the capabilities and
limitations of different explainability methods can help to develop effective policy frameworks that best
address societal needs while promoting innovation. Explanations can also help to design interventions to
improve business outcomes.
This document describes the applicability and the properties of existing approaches and methods
for improving explainability of machine learning (ML) models and AI systems. This document guides
stakeholders through the important considerations involved with selection and application of such
approaches and methods.
While methods for explainability of ML models can play a central role in achieving the explainability
of AI systems, other methods such as data analytics tools and fairness frameworks can contribute to
the understanding of AI systems’ behaviour and outputs. The description and classification of such
complementary methods are out of scope for this document.

© ISO/IEC 2025 – All rights reserved
vi
Technical Specification ISO/IEC TS 6254:2025(en)
Information technology — Artificial intelligence —
Objectives and approaches for explainability and
interpretability of machine learning (ML) models and
artificial intelligence (AI) systems
1 Scope
This document describes approaches and methods that can be used to achieve explainability objectives
of stakeholders with regard to machine learning (ML) models and artificial intelligence (AI) systems’
behaviours, outputs and results. Stakeholders include but are not limited to, academia, industry, policy
makers and end users. It provides guidance concerning the applicability of the described approaches and
methods to the identified objectives throughout the AI system’s life cycle, as defined in ISO/IEC 22989.
2 Normative references
The following documents are referred to in the text in such a way that some or all of their content constitutes
requirements of this document. For dated references, only the edition cited applies. For undated references,
the latest edition of the referenced document (including any amendments) applies.
ISO/IEC 22989:2022, Information technology — Artificial intelligence — Artificial intelligence concepts and
terminology
3 Terms and definitions
For the purposes of this document, the following terms and definitions apply.
ISO and IEC maintain terminology databases for use in standardization at the following addresses:
— ISO Online browsing platform: available at https:// www .iso .org/ obp
— IEC Electropedia: available at https:// www .electropedia .org/
3.1
stakeholder
any individual, group, or organization that can affect, be affected by or perceive itself to be affected by a
decision or activity
[SOURCE: ISO/IEC 22989:2022, 3.5.13]
3.2
explainability
property of an AI system (3.4) that enables a given human audience to comprehend the reasons for the
system's behaviour (3.22)
Note 1 to entry: Explainability methods are not limited to the production of explanations, but also include the enabling
of interpretations.
© ISO/IEC 2025 – All rights reserved
3.3
transparency
property of a system that appropriate information about the system is communicated to relevant
stakeholders (3.1)
Note 1 to entry: Appropriate information for system transparency can include aspects such as features, performance,
limitations, components, procedures, measures, design goals, design choices and assumptions, data sources and
labelling protocols.
Note 2 to entry: Inappropriate disclosure of some aspects of a system can violate security, privacy, or confidentiality
requirements.
[SOURCE: ISO/IEC 22989:2022, 3.5.15, modified — "made available" replaced by "communicated".]
3.4
artificial intelligence system
AI system
engineered system that generates outputs such as content, forecasts, recommendations or decisions for a
given set of human-defined objectives
Note 1 to entry: The engineered system can use various techniques and approaches related to artificial intelligence to
develop a model to represent data, knowledge, processes, etc. which can be used to conduct tasks.
[SOURCE: ISO/IEC 22989:2022, 3.1.4]
3.5
machine learning
ML
process of optimizing model parameters through computational techniques, such that the model's behaviour
(3.22) reflects the data or experience
[SOURCE: ISO/IEC 22989:2022, 3.3.5]
3.6
trustworthiness
ability to meet stakeholder (3.1) expectations in a verifiable way
Note 1 to entry: Depending on the context or sector and also on the specific product or service, data and technology
used, different characteristics apply and need verification to ensure stakeholders’ (3.1) expectations are met.
Note 2 to entry: Characteristics of trustworthiness include, for instance, reliability, availability, resilience, security,
privacy, safety, accountability, transparency, integrity, authenticity, quality and usability.
Note 3 to entry: Trustworthiness is an attribute that can be applied to services, products, technology, data and
information as well as, in the context of governance, to organizations.
[SOURCE: ISO/IEC TR 24028:2020, 3.42, modified — "stakeholders’ expectations" replaced by "stakeholder
expectations"; comma between quality and usability replaced by “and”.]
3.7
feature
measurable property of an object or event with respect to a set of characteristics
Note 1 to entry: Features play a role in training and prediction.
Note 2 to entry: Features provide a machine-readable way to describe the relevant objects. As the algorithm will not
go back to the objects or events themselves, feature representations are designed to contain all useful information.
[SOURCE: ISO/IEC 23053:2022, 3.3.3]

© ISO/IEC 2025 – All rights reserved
3.8
global
property of an explanation (3.27) or an interpretation (3.28) that describes how model predictions are
determined
Note 1 to entry: A global explanation provides an overall understanding of the model’s typical operation. For instance,
a list of rules or features (3.7) that determine the model outputs is an example of global explanation.
3.9
local
property of an explanation (3.27) or an interpretation (3.28) that describes how a single model prediction
was determined
Note 1 to entry: Compared to a global explanation, a local explanation does not try to explain the whole model.
3.10
post hoc explanation
explanation (3.27) built by applying analysis on the model after it has been trained or developed
Note 1 to entry: Post hoc explanations are often used with opaque box (3.16) models, but they are not limited to opaque
box models.
3.11
feature-based explanation
explanation (3.27) of model behaviour (3.22) based on input features (3.7)
Note 1 to entry: For instance, a measure of how much each input feature contributes to a model’s output for a given
data point is an example of feature-based explanation.
Note 2 to entry: An input feature for a model does not necessarily correspond to the inputs a user gives as entry as
several layers of processing can be applied.
3.12
application programming interface
API
boundary across which a software application uses facilities of programming languages to invoke software
services
[SOURCE: ISO/IEC 13522-6:1998, 3.3]
3.13
backpropagation
neural network training method that uses the error at the output layer to adjust and optimise the weights
for the connections from the successive previous layers
[SOURCE: ISO/IEC 23053:2022, 3.2.1]
3.14
classification model
machine learning model whose expected output for a given input is one or more classes
[SOURCE: ISO/IEC 23053:2022, 3.1.1]
3.15
closed box
black box
property of an AI system (3.4) or a model within an AI system, whereby only its outputs can be
obtained programmatically
© ISO/IEC 2025 – All rights reserved
3.16
opaque box
black box
property of an AI system (3.4) or a model within an AI system, whereby it does not offer
intrinsic interpretability (3.17)
3.17
intrinsic interpretability
inherent interpretability
property of an AI model that holds its criteria and decision process (3.21) in an intelligible way in its structure
or content
Note 1 to entry: Intrinsic interpretability is not limited to access only, but also implies an ability to understand the
provided information. For instance, a structure of millions of parameters does not usually constitute an intelligible
way of holding it.
Note 2 to entry: Intrinsic interpretability is opposed to opaque box (3.16).
3.18
decision
content or item produced by the AI system (3.4) as a fulfilment of its task, based on a given input
Note 1 to entry: The decision can be a class, but also any other form of structured or unstructured data (e.g. a sentence,
an image).
3.19
outcome
one of the various options that the AI system (3.4) considers when choosing a given decision (3.18)
Note 1 to entry: Outcomes are candidate decisions.
3.20
output
any data or information returned by the AI system (3.4) when processing a given input
Note 1 to entry: Outputs encompass decisions but also any additional data or information that is returned together
with a decision, e.g. to contextualize or explain it.
3.21
decision process
set of steps and criteria used by the AI system (3.4) to analyse an input and choose the decision (3.18) among
the possible outcomes (3.20)
Note 1 to entry: Depending on the design of the AI system, that decision process can be embedded in part or in whole,
implicitly or explicitly, into the AI system’s models.
3.22
behaviour
any observable effect of a given decision process (3.21), such as a particular decision (3.18), the
preferences made among different outcomes (3.19), a relationship among multiple decisions or a statistical
property of the complete set of decisions made by the AI system (3.4) (including future decisions)
Note 1 to entry: Depending on the design of the AI system, the behaviour of the AI system can be attributed to the
behaviour of the AI system’s models or to their interplay.
3.23
factor
element, property or other characteristic that is considered during the decision process (3.21) and can have
an effect on the chosen decision (3.18)

© ISO/IEC 2025 – All rights reserved
3.24
cause
any type of circumstance that can lead to a given decision (3.18), including for instance the presence, absence
or value of a factor (3.23), but also the analysis made of that factor, its similarity or interaction with other
factors, or the presence or absence of a given step or criterion in the decision process (3.21)
3.25
rationale
piece of information or the analysis made of that information, based on which decisions (3.18) are made
Note 1 to entry: A rationale provided for a single decision identifies one or more causes as having affected the decision
process (3.21) of the AI system (3.4) when choosing that particular decision. A rationale provided without the context
of a specific decision identifies a set of causes (3.24) that can affect the behaviour (3.22) of the AI system during past or
future decisions.
3.26
justification
piece of information or the analysis made of that information, that is sufficient to choose a given decision
(3.18) among the possible outcomes (3.19)
Note 1 to entry: A justification identifies causes relevant to a given decision, without assumption on the set of causes
that have affected the decision process of the AI system (3.4).
3.27
explanation
result of expressing a given rationale (3.25) or justification (3.26) in a way that humans can understand
Note 1 to entry: Explanations can pertain to a decision (3.18) or to an AI system (3.4).
3.28
interpretation
result of understanding (by a human) a given rationale (3.25) or justification (3.26)
Note 1 to entry: Interpretations can pertain to a decision (3.18) or to an AI system (3.4).
Note 2 to entry: Interpretations can be produced either based on a received explanation or directly from observation
without an explicit act of expression.
3.29
behavioural accuracy
adequacy between the outcomes (3.19) to which the explanation (3.27) leads and the actual decisions (3.18)
made by the AI system (3.4)
3.30
simulatability
ability of humans to process the provided information and apply the corresponding criteria mentally to
obtain the output
4 Symbols and abbreviated terms
CEM contrastive explanations method
CEM-MAF contrastive explanations method with monotonic attribute functions
CNN convolutional neural networks
ML machine learning
XAI explainable artificial intelligence

© ISO/IEC 2025 – All rights reserved
5 Overview
Explainability is the property of an AI system that enables a given human audience to comprehend the
reasons for the system’s behaviour. Reasons are rationales or justifications, as defined in this document with
respect to the system’s behaviour. The appropriate way of achieving explainability depends on the context
and stakeholder characteristics. Stakeholder-appropriate explainability helps to achieve concrete objectives
such as:
— identifying the causes of an incorrect decision;
— ensuring that a decision was taken for the right reasons;
— strengthening the confidence in the system.
Users of this document are advised that this concept of explainability (and thus the corresponding set of
methods) is more encompassing than some existing uses of the term “explainability”, while more restrictive
than others. For instance, it includes some analysis and visualization methods, and not only explanations
given by the system itself, but it does not include transparency or AI literacy. Some circles call this concept
"interpretability" and use “explainability” in a different way, but this document does not make this kind of
distinction. See Annex A for further explanations on the exact technical scope targeted in this document.
The relevance of specific objectives depends on the stakeholders and what they are trying to achieve. The
stakeholders can be interested in achieving one or several of the objectives. Stakeholders’ objectives are
discussed in more details in Clause 6, subject to the limitations and concerns discussed in Annex C.
Achieving explainability in a system warrants specific methodological considerations throughout the
whole life cycle. Guidance on that process is offered in Clauses 7, 8 and 9 provide further technical material
(taxonomy of properties for needs assessment and corresponding landscape of methods) to support that
methodology.
6 Stakeholders’ objectives
6.1 General
XAI is a broad field, and stakeholders can have very different reasons to seek explainability. As a result,
there is a large variety of corresponding expectations. It is important to consider them and ensure clarity
on the stakeholders’ objectives, because the utility of an explanation can depend a lot on the stakeholder
receiving it, or on the action taken based on it.
Stakeholders can be characterized into various different types as defined in ISO/IEC 22989. This
characterization is neither unique, comprehensive, nor non-overlapping, but serves to make a point about
different types of explanations. The stakeholder can be someone who participates in developing the AI
system (e.g. a developer) or someone who uses the AI system (e.g. an end user applying for a loan).
Each stakeholder can perform different actions. The developer, for instance, is trying to improve the system
or decide whether to deploy it and can modify the system based on the explanation. It is therefore more
useful if the explanation refers to system characteristics. Such explanations are concerned with the inner
workings of a system or are associated with the system's development process. Such explanations can be
incomprehensible to end users who are unfamiliar with the system's inner workings.
Alternatively, the system can make decisions that the end user wants to address. Therefore, it is more useful
for explanations to emphasize aspects that the end user can comprehend and control. Often there are several
influential paths from input to output and the explanation can pick aspects of the input under the end user's
control. In the process, the explanation can optimize for comprehension (and control) over faithfulness to
the inner-workings of the system.
Figure 1 illustrates the various AI stakeholder roles as they are defined in ISO/IEC 22989. This Clause 6
discusses (and illustrates on practical scenarios) the various objectives that stakeholders can have
depending on their role. See ISO/IEC 22989 for further information on those stakeholder roles.

© ISO/IEC 2025 – All rights reserved
Figure 1 — AI stakeholder roles and their sub-roles
6.2 AI user
A typical objective for AI users seeking explainability is to be reassured on the appropriateness of decisions
that the AI system will induce, and that they will not cause harm to others or themselves by using and being
influenced by that AI system. This can be linked with a reluctance to interact with automated systems that
are perceived as having lower cognitive capabilities than humans, and a desire to maintain human control
and enable interventions.
Another common objective is the desire to be able to justify their own actions (driven by the AI system’s
decisions) when interacting with other stakeholders.
AI users also seek explainability for their own purposes in order to gain a better understanding of a given
domain, topic or item.
Example scenarios where AI users seek explainability include: a judge who needs to decide whether to use
an automated pretrial bail risk assessment system for upcoming cases; or the same judge who for a given
case wants to know the extent to which the generated risk score is reliable; a medical doctor considering to
overrule the recommendation of an automated diagnosis tool because the proposed diagnostic is atypical;
the medical doctor who announces a serious illness to a patient; or a biologist who can discover new causes
for known illnesses thanks to “explainable” analysis of health data.
Another scenario is to learn new information (e.g. motivations for customer churn) that can help
...