Comparison and Analysis of 3 Key AI Documents: EU's Proposed AI Act, Assessment List for Trustworthy AI (ALTAI), and ISO/IEC 42001 AI Management System
Irish Conference on Artificial Intelligence and Cognitive Science (AICS)
✍ Delaram Golpayegani* , Harshvardhan J. Pandit , Dave Lewis
publication 🔓copies: DORAS , harshp.com , TARA , zenodo
An analysis of the three key documents related to regulating AI and identifying how to represent the identified analysis in the form of linked data
Conforming to multiple and sometimes conflicting guidelines, standards, and legislations regarding development, deployment, and governance of AI is a serious challenge for organisations. While the AI standards and regulations are both in early stages of development, it is prudent to avoid a highly-fragmented landscape and market confusion by finding out the gaps and resolving the potential conflicts. This paper provides an initial comparison of ISO/IEC 42001 AI management system standard with the EU trustworthy AI assessment list (ALTAI) and the proposed AI Act using an upper-level ontology for semantic interoperability between trustworthy AI documents with a focus on activities. The comparison is provided as an RDF resource graph to enable further enhancement and reuse in an extensible and interoperable manner.
Keywords:Trustworthy AI, AI management system, ALTAI, AI Act, ISO/IEC 42001, Ontology, Activity, Comparison.
The wide application of AI systems urges governments, legislators, standardisation bodies, and think tanks to encourage and sometimes obligate organisations to develop and use AI in a trustworthy manner. AI regulations, standards, and guidelines developed separately and in isolation risk a highly fragmented landscape that can lead to regulatory and market confusion. Consequently, organisations are compelled to navigate a large number of competing and changing requirements from multiple sources regarding AI development and use. The lack of alignment between different sources of requirements, such as laws and standards, creates difficulties in identifying and fulfilling obligations.
In this paper, we identify the commonality, inconsistencies, and gaps across the following three dominant AI documents within the scope of EU’s regulatory regime: the proposed AI Act , Assessment List for Trustworthy AI (ALTAI) , and the draft ISO/IEC 42001 standard for AI management systems1.
Amongst these three, we utilise ISO/IEC 42001 as the primary source of requirements given its distinct role as a certifiable standard, and compare the others with it to indicate adherence towards guidelines (ALTAI) and regulations (AI Act). More specifically, we investigate the following questions:
To what extent can ALTAI’s trustworthy AI requirements be integrated into ISO/IEC 42001’s AI management system activities?
To what extent can AI Act’s high-risk AI obligations be integrated into ISO/IEC 42001’s AI management system activities?
We address the aforementioned questions by proposing a methodology to compare AI documents using an upper-level trustworthy AI ontology , which enables modelling and linking concepts within AI documents (see Section 2). We then demonstrate the comparison of ISO/IEC 42001 with ALTAI’s trustworthy AI (Section 3) and the AI Act (Section 4). The comparison is made available online as an RDF resource to enable further enhancement and reuse2. We discuss semantic modelling of activities extracted from the documents in Section 5. In Section 6, related work on ontology-based comparison of policies, regulations, and standards is mentioned and we conclude the paper and identify avenues for future work in Section 7.
Methodology for Comparison and Analysis
AI documents can be compared on the basis of different semantic building blocks: key terms defined within them, activities mentioned, and normative requirements or obligations required to be met for compliance. Considering the central focus of management system standards on organisational activities and processes, we limit the scope of our comparison to activities.
Given that different standards, regulations, and policies are being created for evaluating trustworthiness of AI, there is bound to be some overlap between them. To assist in the task of comparing them, a conceptual model and framework is essential to identify and link together the relevant concepts within different documents. An ontological representation permits formalisation of the conceptual model and its application in use-cases. With this view, Fig. [fig:TAI ontology] presents the core ontology for supporting mapping of concepts between different emerging AI standards. It is based on activities carried out within ISO/IEC (more specifically sub-committee 42) regarding AI standardisation and incorporates existing ISO/IEC standards and outputs for ‘characteristics’ expressed by trustworthy AI systems.
The premise of the ontology rests on the fact that several of trustworthy characteristics are yet to be clarified and defined in relation to AI and AI development activities. Therefore, it focuses on specifying the relationships between activities, entities, assets, and characteristics (exhibited for trustworthiness), agents, stakeholders, and organisations. The ontology is based on Basic Formal Ontology (BFO) - a generic upper-level ontology used in formalisations across domains, and the PROV-O ontology which is a W3C standard for expressing provenance.
The ontology provides a way to express activities of organisations that relate to AI where the trustworthiness is manifested through characteristics of Entities that make up a product or service employing AI. It also provides a way to depict the influence of entities, activities, and agents in these processes, and captures the role of stakeholders in disclosing and exhibiting trustworthiness of AI through its characteristics. The ontology thus enables representing use of AI from both within and outside the perspective of an organisation or service, and is useful for comparing different AI guidelines by using its conceptual model as a framework for identifying and aligning concepts.
We utilise the trustworthy AI ontology to compare AI documents in order to assess the degree of alignment between them by modelling and linking trustworthy AI activities mentioned within them. The following describes the steps taken for analysis and comparison of documents:
The documents are analysed to extract relevant activities to trustworthy AI, which then modelled as
partOfrelationship is used to bridge the isolated sets of
Activitiesidentified from the documents.
An analysis is carried out to identify the overlaps and potential conflicts through investigation of activities that are mapped or could not be mapped using the
Comparison of ALTAI with ISO/IEC 42001
ALTAI suggests a set of questions, grouped by the ethical principle under assessment, for assessing whether an AI system adheres to trustworthy AI requirements specified in  (see the structure of ALTAI in Fig. [fig:ALTAI structure]). Designed for trustworthy AI self-assessment, ALTAI provides useful hints regarding development and use of AI systems. One of the aspects of trustworthiness assessment is execution of particular activities; for example, ‘Are end-users or other subjects adequately made aware that a decision, content, advice or outcome is the result of an algorithmic decision?’, which is a question listed under Human Agency and Oversight requirements, implies execution of an activity to inform end-users or other subjects that a decision, content, advice or outcome is the result of an algorithmic decision. For the purpose of comparison, we made the management activities implied by ALTAI questions explicit.
AI Management System Activities
The ISO/IEC 42001 standard for AI management systems, being developed by JTC 1/SC 42, is currently (Nov’22) in DIS or draft stage, implying relative maturity awaiting final comments before publication. It follows the ‘harmonised structure’ of all management system standards developed by ISO, which is defined in the openly available ISO/IEC Directives on procedures for ISO technical work3. Based on the harmonised structure, Lewis et al.  identified AI management system activities, where each is given an identifier, a label, and a ‘see also’ attribute which is a link to the relevant harmonised structure clause. The entities generated and used by each activity are represented in a similar manner. The updated list of AI management system activities, which reflects the latest version of the Directive published in 2022, is presented in Table 1.
|o No.||ID||AIMS activity (label)||HS clause (see also)|
|1||UOC||Understanding organisation and its context||4.1|
|2||USE||Understanding stakeholder needs and expectation||4.2|
|3||DS||Determine AIMS scope||4.3|
|4||EIMI||Establish, implement, maintain and continually improve management system and its processes||4.4|
|5||DLC||Demonstrate leadership and commitment to the management system||5.1|
|6||EP||Establish AIMS policy||5.2|
|7||ARRA||Assign roles, responsibilities and authorities||5.3|
|8||ARO||Address risks and opportunities||6.1|
|9||EPAO||Establish and plan to achieve AI objectives||6.2|
|10||ARRA||Assign roles, responsibilities and authorities||6.3|
|11||DAR||Determine and allocate resources for AIMS||7.1|
|12||DEC||Determine and ensure competence of people affecting AI performance||7.2|
|14||DC||Determine AIMS communication||7.4|
|15||CUCD||Create, update, and control documented information||7.5|
|16||PCP||Plan and control AI processes||8.1|
|17||MMAE||Monitor, measure, analyse and evaluate AI||9.1|
|18||IA||Internal (AIMS) audit||9.2|
|19||UMR||Undertake management review||9.3|
|20||DNCA||Detect non-conformance and take corrective action||10.1|
|21||CI||AIMS Continual improvement||10.2|
ALTAI - ISO/IEC 42001 Activity Comparison
By comparing ALTAI with ISO/IEC 42001, we aim to investigate the following:
Is there any organisational activity required for trustworthy AI that cannot be integrated into an AI management system?
Which AI management systems activities do not play a role in achieving trustworthiness?
What management systems activities are involved in achieving a particular trustworthy AI requirement, e.g. privacy and data governance?
In the comparison process, a number of commonly occurring structures
are identified. For instance, multiple ALTAI activities that refer to
achieving AI objectives such as Accuracy,
Explainability, Privacy, and Fairness are
partOf ‘establish and plan to achieve AI objectives’
activity. We categorise these structures into the 17 alignment groups
listed in Table 2.
|ID||ALTAI activity structure||partOf (AIMS activity)|
|AG1||Assess the impact of the AI system||ARO|
|AG2||Assess the system vulnerabilities or threats||ARO|
|AG3||Assess whether the AI system respects a specific right||ARO|
|AG4||Establish processes to test or monitor AI impacts or risks||PCP & ARO & MMAE|
|AG5||Establish processes to measure and assess AI risks||PCP & ARO|
|AG6||Establish processes to mitigate, rectify, or avoid AI risks||PCP & ARO|
|AG7||Establish processes to achieve an AI objective||PCP & EPAO|
|AG8||Assess whether an AI objective is achieved||EPAO & MMAE|
|AG9||Establish processes to test and monitor AI objectives||PCP & EPAO & MMAE|
|AG10||Establish processes to measure and assess AI objectives||PCP & EPAO & MMAE|
|AG11||Provide information about a design decision||UOC|
|AG12||Determine compliance / Align the systems with a specific standard or guideline||PCP & UOC|
|AG13||Designate a role||ARRA|
|AG14||Establish a broad (e.g. ethics review board)||ARRA|
|AG15||Provide employee training / Ensure workers competence||DEC|
|AG16||Communicate with or inform users or third parties||DC|
|AG17||Inform staff and employees about the AI policy||PA|
The comparison revealed that ALTAI is centred around trustworthy AI issues and principles rather than how to manage trustworthy AI processes and policies within an organisation. In comparison, the draft AI management system standard does not specifically refer to any trustworthy principle, however, it provides a foundation for implementing these principles in an organisation. The two are therefore complementary regarding effective implementation and assessment of trustworthy AI, with the comparison providing a way to achieve trustworthiness through management system activities.
Table 3 presents the number of ALTAI activities that are mapped into each AI management system activity. It should be noted that the total number indicates the number of times an AI management system activity is individually mapped to ALTAI activities as the mapping between the two is many-to-many. Activities within AI management system that do not have a corresponding ALTAI activity are omitted from the table (8 in total).
As shown in the table, approximately 50 percent (73 of 144) of ALTAI activities refer to risk management which makes the fact that ALTAI adopts a risk-oriented approach towards trustworthy AI clear. The missing management system activities in the table, which are nearly half of total, demonstrates that processes and tasks at a high level of organisational governance and management are not covered in ALTAI.
|AIMS activity||AIMS activity (label)||Nos. ALTAI activities|
|ARO||Address risks and opportunities||73|
|PCP||Plan and control AI processes||54|
|EPAO||Establish and plan to achieve AI objectives||44|
|DC||Determine AIMS communication||22|
|MMAE||Monitor, measure, analyse and evaluate AI||20|
|UOC||Understanding organisation and its context||12|
|DEC||Determine and ensure competence of people affecting AI performance||7|
|ARRA||Assign roles, responsibilities and authorities||2|
Comparison of AI Act with ISO/IEC 42001
The AI Act Activities
In April 2021, the European Commission published the proposal for EU AI regulation, called AI Act, to create a legal framework for trustworthy AI by laying down obligations which are proportionate to the level of risk imposed by AI systems. Under the AI Act, providers of high-risk AI systems, i.e. systems that are likely to cause harm to health, safety, and rights of individuals, are required to implement a quality management system (Art. 17), among other requirements. The AI Act relies on creation of harmonised AI standards to facilitate conformity to its requirements by providing technical solutions (Art. 40).
Conformity with the AI Act’s high-risk AI obligations requires
performing organisational as well as technical activities. By analysis
of the requirements for high-risk AI systems and the obligations of
providers of those systems, described in title III, Chapters 2 and 3, we
identified 52 high-level organisational activities that are
associatedWith high-risk AI providers, which are modelled
Agents. It is important to note that our list of
activities is not exhaustive, and therefore performing the identified
activities is essential for conformity to the AI Act but not necessarily
AI Act - ISO/IEC 42001 Activity Comparison
Using the methodology described earlier, we mapped the activities identified from the AI Act to the ones extracted from ISO/IEC 42001. Table 4 shows mapping of AI Act’s risk management activities into AI management system.
|AI Act risk management activity||partOf (AIMS)|
|Establish risk management system||DC & EIMI & ARO|
|Implement risk management system||EIMI & ARO|
|Document risk management system||EIMI & ARO & CUCD|
|Maintain risk management system||EIMI & ARO|
|Identify/ Analyse/ Evaluate/ Mitigate Risks||ARO|
|Communicate Residual Risk to Users||PA & AIRO|
|Identify Impact On Stakeholders (e.g. children)||USNE & ARO|
Our analysis indicates activities to establish management systems, address risks, create documentation, and communicate with external entities are among the most mapped management system activities. This shows that in conformity to the AI Act’s legal requirements, documentation and sharing information with external stakeholders are as important as conducting risk management.
Identification of the degree to which compliance to ISO/IEC 42001 assists in conformity to AI Act’s high-risk AI obligations needs further investigation as our focus was primarily on the organisational activities explicitly referenced therein.
Semantic Modelling of Activities
Documents that specify guidelines generally refer to activities and processes across three distinct phases: ex-ante where a plan of activity must exist; ongoing or during where an activity is currently in the process of being executed; and ex-post where an activity has finished execution or has produced artefacts. For AI guidelines, it is important to model the corresponding semantic representation of activities in a similar manner so as to distinguish when an organisation or system must have a plan in place representing some future activity versus having carried out that activity i.e. in the past. This notion is also applicable and demonstrated in the area of legal and regulatory compliance where an obligation can entail provenance of both a plan as well as executed activities, and therefore requires documentation at both ex-ante and ex-post phases .
Intended for self-assessment purposes, ALTAI predominately refers to the ex-post phase. This means that to provide answers to ALTAI questions we have to look into the results and artefacts of executed activities. Furthermore, separation between ex-ante and ex-post phases of ALTAI activities enables ex-ante planning for trustworthiness and ex-post trustworthy AI (self-) assessment as outlined by AI management system activities. However, for semantic representation of the activities extracted from ALTAI both planning and execution phases should be taken into account. For example, from ‘establish processes to assess AI risks’ two activities are inferred: plan for AI risk assessment (ex-ante) and AI risk assessment (ex-post). A semantic model of the former should be able to represent plans for risk assessment, intended steps and actions, responsible parties, and entities generated and used during the planning. This can be done by extending the Ontology for Provenance and Plans (P-Plan)4. Naja et al.  have adopted the same approach for recording accountability plans. Representing ex-post activities is possible by extending the PROV-O ontology.
To model previously introduced alignment groups we consider the ex-post phase. Each alignment group can be represented as an ontology design pattern (ODP) . An example of one such pattern for AG17 (providing training for employees to ensure competence) that uses the PROV-O ontology to represent agents5 and activities is shown in fig. [fig:training activity]. By modelling training activities using this pattern all processes and activities which are part of DEC (Determine and ensure competence of people affecting AI performance) can be uniformly represented, and retrieved e.g. using SPARQL queries.
Using the pattern as a generic template for different activities and roles regarding training enables a uniform mechanism to answer questions such as:
Did the organisation provide training to staff on risk management?
Who provided the training? When? To whom? On what topic?
What activities are relevant to training?
What are the subjects that the organisation provides training on?
Who is trained on a specific topic, e.g. risk management?
Boer et al.  used an ontology-based approach to facilitate comparison of similar regulations, i.e. in a specific area such as tax, within different jurisdictions. Despres and Szulman  proposed an approach for integrating ontologies created from the European community directives. Fiorentini et al.  proposed an approach for harmonisation which compares documents using informal analysis, typology of standards, use-cases, and ontologies. Pardo et al.  created h3mO - an ontology for harmonisation of reference models and standards utilised in software process improvement. Koelle et al.  proposed a tool for ATM security which harmonises relevant standards and regulations. Lewis et al.  presented an analysis of the normative content of trustworthy AI guidelines presented by IEEE, EU HLEG, and OECD and mapped these guidelines into ISO 26000 social responsibility issues.
This paper presented a comparison and analysis between the EU AI Act, ALTAI, and ISO/IEC AI management system standard to identify the potential alignment between these 3 key documents. The assessment compared management-level activities mentioned in the documents and is represented formally using the trustworthy AI upper-level ontology proposed by .
Implications of Comparison and Analysis of AI documents
Identification of the gaps existed in the AI documents being developed assists standardisation bodies in determining the areas that need creation or modification of standards. Legislators can use the comparison to determine the degree to which compliance with existing AI standards contributes to conformity to legal obligations and identify the aspects of trustworthy AI that are not subject to regulation. Furthermore, comparison of activities provides a baseline for the communications between authorities and standardisation bodies for development of harmonised regulations and standards.
The comparison assists AI providers and developers in adoption of standards and guidelines required for satisfying legal requirements by helping them identify inconsistencies and areas of overlaps. It can also be used to ensure organisational AI policies are effective in satisfying normative and legal requirements.
Given the potential of AI research to cause harm, recently some AI conferences, such as NeurIPS6, provide ethical guidelines and ask researchers to assess the impact of their work on key areas of concern, e.g. safety, fairness, and privacy. The comparison methodology can be applied in assessing the alignment of ethical guidelines provided by different conferences, universities’ policies on ethics and data protection as well as ethical assessment approaches.
The comparison presented in this paper will be expanded to provide a more comprehensive analysis and alignment of key terms, technical activities, and requirements detailed within AI documents. Starting with the analysis provided in this paper, we aim to identify a common set of AI risk and impact assessment activities from the AI Act, ALTAI, and ISO risk management and management system standards and extend AIRO - an ontology for describing AI risks , to represent provenance of activities. Future work also includes updating this work based on changes made in the subsequent drafts and finalisations of the AI Act and ISO/IEC 42001 standard.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 813497, as part of the ADAPT SFI Centre for Digital Media Technology is funded by Science Foundation Ireland through the SFI Research Centres Programme and is co-funded under the European Regional Development Fund (ERDF) through Grant#13/RC/2106\_P2. Harshvardhan J. Pandit has received funding under the Irish Research Council Government of Ireland Postdoctoral Fellowship Grant#GOIPD/2020/790.
The PROV concepts of agents and entities are different from ALTAI and AIMS. In PROV, an entity is an artefact such as an input to an activity, and an agent is what is referred to as an entity within ALTAI, AIMS, and the general use of the words.↩︎
NeurIPS 2022 ethics guidelines https://neurips.cc/public/EthicsGuidelines↩︎