How can AI be tested?
A product or service must meet user expectations, fulfill contractual requirements, and comply with legal regulations and standards. All of these aspects must also be appropriately verified and tested in an AI project.
However, the requirements for AI-based systems are different. The behavior of artificial intelligence is difficult to predict deterministically, which means that traditional test design methods, such as boundary value analysis or the white-box method, can no longer be used—or at least not as easily.
Consequently, the established approach to testing conventional software (i.e., non-AI software)—namely, using a test oracle to determine how a system should respond and then verifying whether the actual response matches the predicted one—can no longer be applied so readily. New testing methods, such as metamorphic testing, are therefore needed, and the question arises as to where the existing testing approach can continue to be used and where the testing process and the methods employed must be adapted or redesigned.
However, the existing testing approach still has its place in this context. This is because, in some systems, artificial intelligence performs only specific sub-tasks. Other systems, while AI-based, contain conventional components or work in conjunction with them. In these cases, the conventional parts of these systems, as well as the interfaces between AI and non-AI components, are tested using conventional methods. Existing testing procedures and quality metrics therefore remain applicable in this scenario as well, and a mix of different methods and approaches will become established.
Regarding the question, “Which testing methods are applicable and cost-effective for my project?” imbus can advise you and offer support in the planning, specification, and execution of tests.
Ihr Ansprechpartner bei imbus
Mr. Tilo Linz