A responsible AI framework: pipeline contextualisation

Abstract

Incorporating ethics and values within the life cycle of an AI asset means securing its development, deployment, use, and decommission under these perspectives. These approaches depend on the market domain where AI is operational – considering the interaction and the impact on humans if any process does not perform as expected – and the legal compliance, both required to ensure adequate fulfilment of ethics and values. Specifically, in the manufacturing sector, standards were developed since the 1990’s to guarantee, among others, the correct use of mechanical machinery, systems robustness, low product variability, workers safety, system security, and adequate implementation of system constraints. However, it is challenging to blend the existing practices with the needs associated with deployments of AI in a trustworthy manner. This document provides an extended framework for AI Management within the Manufacturing sector. The framework is based on different perspectives related to responsible AI that handle trustworthy issues as risk. The approach is based on the idea that ethical considerations can and should be handled as hazards. If these requirements or constraints are not adequately fulfilled and managed, it is expected severe negative impact on different sustainable pillars. We are proposing a well-structured approach based on risk management that would allow implementing ethical concerns in any life cycle stages of AI components in the manufacturing sector. The framework follows a pipeline structure, with the possibility of being extended and connected with other industrial Risk Management Processes, facilitating its implementation in the manufacturing domain. Furthermore, given the dynamic condition of the regulatory state of AI, the framework allows extension and considerations that could be developed in the future.

TAI-PRM: trustworthy AI—project risk management framework towards Industry 5.0

“…This section covers necessary considerations before proceeding with the TAI-PRM. As a reminder, readers are encouraged to check previous publications in check approaches of extensions to other domains, regulatory conditions, or management approaches [51,52].…”

“…and guidelines for managing risks. The first two components are discussed in [51,52], while the focus of this work is on developing and testing the protocols for the risk management process. The following considerations were also defined: (i) the communication and consultation activities are part of the RASP architecture, not the risk management process itself; and (ii) the ethical risk management process includes the ISO-defined Monitoring and Review process as the primary component after Risk Evaluation and risk treatment.…”

“…AI assets were viewed as individual components that interact with other elements and can be embedded within systems or standalone. Dependencies of AI assets are classified in [52].…”

“…The legislation heterogeneity applied in different countries on the use of AI can be varied; therefore, flexibility is key. In this stage, and as explained in reference [52] a regulatory check process is included. Furthermore, since the approach proposed is fundamentally constructed for TAI requirements (with possible extensions, as covered in this manuscript), the possibility of extending the applicability of the framework within a similar environment of application should be possible.…”

“…TAI-PRM builds on two earlier works: the first one provides a scope contextualization that aids in comprehending the risk management process and evaluation [52]; the second establishes the groundwork for AI development and management, establishing connections between requirements and risk components [51]. These two works, together with the current one, can be used to define operationalization approaches for the framework proposed and facilitate the translation of the framework on domains different from the main one focused here (i.e.…”