AI Certification: An Accreditation Perspective

Artificial intelligence (AI) applications are increasingly relevant in every business sector. Some AI systems pose a risk to important commodities worthy of protection, such as health, environment or fundamental rights. This rationale drove the European Union’s approach in developing Regulation (EU) 2024/1689 (AI Act) [9]. Although the EU AI Act may be the most comprehensive, other governments like the US or China have provisions with different approaches and regulatory depth [11, 10].

Where regulations exist, compliance must be ensured without establishing new trade barriers. Therefore, the World Trade Organization (WTO) demands for processes of conformity assessment and accreditation²²2Accreditation is a third-party attestation with a conformity assessment body (CAB) as object of conformity. The purpose is to ensure a CAB’s formal demonstration of its competence, impartiality and consistent operation in performing specific conformity assessment activities. The authority of an accreditation body can be derived from governments. In Europe, accreditation bodies are government-authorized bodies according to Regulation (EC) No 765/2008. [12]. With the accreditation of a conformity assessment body (CAB), it is ensured that its results can be recognized by other WTO member states as equivalent to their own national CAB results. In this way, accreditation supports international trade since it helps companies avoid duplicating costly conformity assessment procedures in other WTO member countries for companies. In order to enable this system, accreditation bodies are tasked with assessing the competence, impartiality, and independence of conformity assessment bodies in the sense that these CABs perform their work reliably and in a comparable as well as reproducible manner. Thus, all the procedures and methods applied and used by CABs to perform conformity assessment need to be objective, reproducible and comparable. Therefore, one purpose of this paper is to outline how the established system of conformity assessment works in order to ensure comparability and reproducibility and how this can be applied in the context of AI systems – in an international context as well as with reference to the AI Act. The other is to clarify the distinction of conformity assessment and AI system “testing” and “validation”.

This paper introduces the methodology of the international conformity assessment system in section 2 which is based on the internationally accepted standards series of ISO/IEC 17000, the so-called ISO CASCO toolbox. It will be used to subsequently focus in section 3 on the conformity assessment activity of certification. Here, it will be outlined how certifications with focus on AI systems (mainly seen as products) should be developed in general based on the internationally existent system. It will also briefly explain the distinction between conformity assessment and AI system “testing” and “validation” with respect to the development cycle. This will be followed by the specifications of conformity assessment for the European context in accordance with the EU AI Act in section 4. The paper concludes by outlining some paradigmatic gaps at the technical level of AI systems and by directing to areas where further scientific and standardization work is needed from an accreditation point of view.

According to ISO/IEC 17000, conformity assessment means the demonstration that specified requirements are fulfilled.³³3Those specified requirements are very often determined in standards and less commonly in technical specifications (in terms of regulation (EU) 1025/2012). Seldom are they directly determined within national regulations. Conformity assessment must account for all different kinds of objects of conformity assessment such as products, processes, organizations, persons, services, data, systems, materials, designs. Hence, a variety of different types of conformity assessment activities exist. These are mainly testing (including calibration and proficiency testing), inspection, validation/verification and certification (of management systems; persons; products, processes and services). While those activities are all different, they follow the same functional approach determined in ISO/IEC 17000, Annex A. In order to ensure the comparability and reproducibility of the results, each activity and the needed procedure to fulfil the functional approach are defined in the respective conformity assessment standard (ISO/IEC 17000 series), with every one of it internationally agreed upon to support the WTO mutual recognition approach of statements of conformity.

Conformity assessment can be performed by three different entities: the organization providing a product or service (first-party), organizations with user interest (second-party) or independent accredited conformity assessment bodies (third-party). The accredited third-party conformity assessment body is the only one that is allowed to conduct certification as special type of conformity assessment activity according to ISO/IEC 17000 [2].

Third-party certification is therefore an important independent evaluation method to demonstrate certain levels of quality, transparency and trust of a product or service, management system or a competent person to other market players and authorities.

For AI systems, the following certifications⁴⁴4Certifications of persons according to ISO/IEC 17024 can be important but are not considered here. The focus is on the needed conformity assessment activities for market access that determines the AI system as a “product”. are particularly relevant: management system certifications (ISO/IEC 17021-1) [3] and certifications for products, processes and services (ISO/IEC 17065) [4].

With the AI Act in force, specific requirements need to be fulfilled by European AI operators.

Operators of no- or low-risk AI systems (Art. 3, No. 8, AI Act), only need to fulfil transparency requirements. In contrast, operators of high-risk AI systems must meet all requirements in Chapter III, and demonstrate it by using pre-defined conformity assessment procedures. According to Article 43 AI Act, there are two options: self-declaration (first-party) or conformity assessment by a notified body (third-party) (as product certification). Yet, where the AI system is part of another product (Annex I, especially Section A), the relevant conformity assessment under those legal acts is required.

All of these options align with the New Legislative Framework (NLF)⁸⁸8The NLF was adopted in 2008 in order to establish a common legal framework for placing goods on the internal market and ensure a high quality of those products placed on the market. The NLF consists of regulation (EC) 765/2008, decision 768/2008/EC and regulation (EU) 2019/1020. and reflect procedures outlined in different modules of decision 768/2008/EC. Module A of this decision applies to the self-declaration and module H1⁹⁹9The European Cooperation for Accreditation (EA) declares in its EA Accreditation for Notification (AfN) Project report from July 2024 [7] that the preferred standard for regulation (EU) 2024/1689 (AI Act) is ISO/IEC 17065. to a notified body. For AI systems listed in Annex I, the modules of these legal acts apply with the preferred conformity assessment standard (of the 17000 series) to be used according to EA-2/17 [8]. Often, this is also ISO/IEC 17065.

Looking at the second option outlined in Article 43 of the AI Act, notified bodies need to perform conformity assessment (product certification according to ISO/IEC 17065). Thus, also in the European context, the general approach of product certification as outlined in section 3.2 remains applicable. However, the legal consequences differ. In the EU, only a positive assessment by a notified body for a high-risk AI system (here as result of a product certification), where required, permits affixing the CE-marking. Furthermore, the certification process is outlined in Annex VII AI Act. It includes the assessment of the quality management system (QMS) in accordance with Art. 17 of the AI Act via the conformity assessment activity audit following ISO/IEC 17021-1, and the technical documentation in accordance of Art. 11 of the AI Act regarding the high-risk AI system via an inspection in order to evaluate if all requirements of Chapter III are comprised by the technical documentation.

Accordingly, it applies for the European context as well that specific certification schemes according to ISO/IEC 17067 are developed to ensure compliance with the AI Act based on the already existent CAA (ISO/IEC 17000 series). Since many other harmonized regulations may interfere with the AI Act, it may be feasible to develop this within the different economic sectors in order to take the specific sectoral regulations as well as business practices and applications into account. Thus, for a notified body to attest conformity for an operator, these scheme developments are now crucial. For potential EU-specific objective measurands (e.g. for high quality data), it is necessary that first the standardization experts define the technical requirements for the respective objective of conformity assessment (e.g. data), and subsequently determine which kind of conformity assessment activities may need a specification (e.g. specific testing measurand according to ISO/IEC 17025 for “high quality data” or the development of a procedure to provide “reference data” according to ISO 17034 [1]).

In conclusion, we can summarize that the general framework for conducting conformity assessment and especially certification is provided by the ISO/IEC 17000 series and it is generally sufficient for complex systems. This paper demonstrated that this also applies for complex (black-box) and non-complex (white-box) AI systems, since both can be assessed based on these conformity assessment activities. Hence, no additional standards specifically targeting the level of the conformity assessment activities (ISO/IEC 17000 series) are required. However, this paper also made clear what is missing. In particular, there is a need to develop sector and technology-specific certification schemes according to ISO/IEC 17067 – both internationally and at the European level. Furthermore, there is still a gap in standardization regarding objective scientifically proven measurands for certain components (e.g. bias in datasets; security features; robustness) and their calibration as well as the applicable reproducible testing methods. Here, there might be even the need to develop European specific standard testing methods. However, this can only be evaluated once the applicable AI system standards to meet the requirements of the AI Act are developed. From an accreditation perspective, there might also be the need to develop objective synthetic test and reference datasets to ensure a high quality of the testing procedures in laboratories and to reliably assess AI model validation procedures. With regard to complex (black-box) AI systems, there is also a lack of scientific basis that determines whether conformity is presumed or not. Thus, more technical work needs to be done here.

On the other hand, it became evident that in all AI system contexts, AIMS certifications play a vital role for maintaining trust in AI systems and can be considered as even more important than other management systems being used more conventionally as quality assurance measure. The reason is that an AI management system allows to control the context of deployment and to some degree the processes of data input and generated output of an AI system. With regard to complex (black-box) AI systems and the limited predictability of their behaviour, AIMS certifications are thus the only trust anchor available. Consequently, the importance and trustworthiness of the system of accreditation and conformity assessment in general, and in particular of those accredited CABs for these AIMS certifications, should not be underestimated.