Seite - 129 - in The Future of Software Quality Assurance

Testing Artificial Intelligence 129 3.3 Under-confidence inAI The other side of this is under-confidence. A rational debate on whether to use AI can be blurredby uncertainty, irrational fear or bias in the media (or sci-fi movies). Accidentswithself-drivingcarsgetmoreheadlines thanordinaryaccidents.People areafraid to becomeobsoleteor that a maliciousghost in themachinemightarise. 3.4 Traceability Withnon-AI-systemsthealgorithmis thecode.This isnot thecasewithAI-systems so we don’t know the exact criteria by which the AI-system takes decisions. Next to that it’s hard to oversee the total population of training data and therefore get a good understanding of how the AI system will behave. So when the outcome is evidently incorrect, it is hard to pinpoint the cause and correct it. Is it the training data, the parameters, the neural network or the labelling? Lack in traceability fuels over-confidenceandunder-confidence(aswasshownabove)andcausesuncertainty in liability (was it thesoftware, thedata, the labellingor thecontext thatdid it?)and lackofmaintainability (what tocorrect?). 4 TestingAI Thekeytomitigationof theAIrisks is transparency.Inbiasweneedinsight into the representativenessof trainingdataandlabelling,butmostofallweneedinsight into how important expectations and consequences for all parties involved are reflected in the results. Building the rightamountofconfidenceand traceabilityneeds transparencytoo. Transparency will not be achieved by illuminating the code. Even if this were possible, by showing a heat-map of the code indicating which part of the neural network is active when a particular part of an object is analysed or a calculation in a layer isproduced,meansclose tonothing.Lookinginsideabrainwillnevershow a thoughtordecision. It couldshowwhichpart isactivatedbutallmentalprocesses always involve multiple brain parts to be involved and most of all experience from thepast. AI systems are black boxes, so we should test them like we do in black box testing: fromthe outside, developing test cases that are modelled on real-life input. From there expectationson the output are determined. Sounds traditional and well known,doesn’t it? The basic logic of testing AI might be familiar, the specific tasks and elements areverydifferent.