Sergey V. Garbuk

The problem of assessment and trust validation of applied artificial intelligence systems is examined. Despite the diversity of studies in this area, it is shown that in fact there is no unified approach to trust validation. At the same time, different researchers emphasize one or another aspect of trust without attempting to cover the problem as a whole. The trust feature is reduced to an emotional characteristic, the constructive meaning of which in the context of artificial intelligence systems lifecycle management is not defined. In some cases there is an unjustified ‘humanization’ of the concept of trust, as a result of which the characteristics of a technical system are completely replaced by the subjective treatment of its creators, suppliers or owners.

The paper attempts to substantiate a holistic trust model for artificial intelligence systems based on formalization of the relationship between trust and system quality. In this case the paper shows that the characteristic of trust can be used to make a decision by the operator, regulators and other interested parties about the possibility (or impossibility) to use the system as intended in certain operating conditions. The main aspects of trust are identified, taking into account both the general issues of trust in automated systems (technical reliability, information security and import independence) and the specifics of AI systems based on machine learning algorithms (functional correctness), and an approach to evaluating the characteristics of the artificial intelligence systems according to these aspects is proposed.

The obtained results can be used for the development of conformity assessment system in the field of AI technologies, as well as for the selection of systems that provide the solution of intelligent data processing tasks in various industries and social sphere.