Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims

Ethical Assurance: A practical approach to the responsible design, development, and deployment of data-driven technologies

“…Cobbling together "a robust 'toolbox' of mechanisms to support the verification of claims about AI systems and development processes" (Brundage et al 2020), in this latter sense, leads, in AI ethics and governance, to a kind of functional tardiness of the governance strategies that result. Namely, it leads to an emphasis on narrowly-targeted methods such as "effective assessment" (Brundage et al 2020), "auditability" (Mökander and Floridi 2021;Raji et al 2020), "traceability" (Kroll 2021), and "reviewability" (Cobbe, Lee, and Singh 2021), that show up on the scene a moment too late. Such methods remain ex post facto and external to the innerworkings of sufficiently reflective and responsible modes of technology production and use.…”

“…By this we mean research that seeks to support or develop mechanisms by which the processes and outcomes that characterise ML/AI research and innovation can be made, for example, more transparent, trustworthy, or responsible. Within this broader remit, there is research that includes general overviews or frameworks that support transparent reporting and communication (Mitchell et al 2019;Brundage et al 2020), specific (narrowly-focused) tools that support bias mitigation or algorithmic interpretability (Research 2018;PAIR 2020;Lundberg 2020;ICO 2020), as well as more focused extensions of assurance cases to address the specific challenges of ML (Ashmore, Calinescu, and Paterson 2019;Ward and Habli 2020;Habli et al 2020). Each of these approaches can play a valuable role individually, but collectively add up to a (currently) disorganised toolbox of practical mechanisms with little unifying purpose or direction.…”

“…Each of these approaches can play a valuable role individually, but collectively add up to a (currently) disorganised toolbox of practical mechanisms with little unifying purpose or direction. Brundage et al (2020) provide some means for bringing order to this miscellany, by identifying practical mechanisms to support the trustworthy development of AI systems, categorised according to whether they are 'institutional,' 'software,' or 'hardware' mechanisms. For example, they note that institutional mechanisms, such as third party auditing can be used to create a "robust alternative to selfassessment claims," while bias and safety bounties can "strengthen incentives to to discover and report flaws in AI systems" (Brundage et al 2020, 1).…”

“…Turning to arguably one of the most comprehensive proposals, Ashmore et al (2019) provide a systematic survey of ML assurance, focusing on the generation of evidential artefacts that can be used in the process of developing and evaluating an ML system. In a similar vein to (Brundage et al 2020), their approach covers the methods and mechanisms that are capable of providing evidence for claims about the properties of ML systems. However, their approach categorises these methods according to where in the ML lifecycle they are most relevant.…”

“…It is important to acknowledge thatBrundage et al (2020) recognise that the mechanisms alone are merely tools to support wider processes of governance, and also suggest the need for pursuing argument-based forms of assurance in Appendix III.7 Ashmore et al (2019) also define key desiderata for each of the four stages of their "ML lifecyle": data management, model learning, model verification, and model deployment.…”