Seite - 131 - in The Future of Software Quality Assurance

Testing Artificial Intelligence 131 the search engine of a museum. The tourist wants to know if a particular picture is fordisplay, the art loveralsowantsbackgroundinformationandsketches. Next to that:as theworldchanges, theusersand their expectationscouldchange overnight.Thinkof what the fire in the NotreDame did to what usersmightexpect whensearchingfor“NotreDame”or“fire inParis”.AI recognisingviruses inDNA sequencesshould take intoconsiderationpossiblemutations thatoccurconstantly. So testing AI starts with identifying the users or the perspectives from which output from the system will be used. This means studying data analytics on the usageof thesystem, interviewingprocessownersor interviewingrealusers. 4.3 ProfilingUsers Identifyingusersorgroupsofdataisone,determiningwhat theywant,expect,need, are afraid of or will behave like, is another. What the tester needs is profiles of the users and perspectives: what is their typical background, what do they want, what turns themofforupsets themandwhatdo theyexpect? A technique to create profiles is “Persona”. Key to this technique is to not think of an entire group of users but to pick one from this group and make her or him as concreteas possible. The benefit of Persona is that it makes the user come alive. It’s a technique to take the perspective of a user from the inside out. For example: thePersonaforAmericantouristscouldbeJoe,aplumber, livinginChicago,white, aged 45, married, two children. He is not well read but loves colourful and well- craftedpaintings.Hishobbiesarefishingandrefurbishingoldaudioequipment.He is turnedoffbyprofoundtheoriesbut likesthehumansideandbackgroundof things (Fig.1). 4.4 CreatingTest Cases Thispart isprobablywheremostof thework is for the tester.Peruserprofile, input andexpectedoutput isdetermined.Goodprofileswill provideagoodbasisbutwill probablyneedextra informationcomingfromresearchand interviews. Identifying test cases will never be complete nor definitive: you can’t test everything, also not in AI. The world and the users change so this needs to be reflected in the requirements. It starts with the most important cases; it will grow constantlyandneedspermanentmaintenance.