TFL Vsebine / Revija SIR*IUS

The risks and control options in artificial intelligence solutions

Razvoj informacijskih tehnologij je pripeljal do obdobja, v katerem ljudje ustvarjamo, delimo in si izmenjujemo čedalje večje količine podatkov. Vpeljava metod umetne inteligence nam vse pogosteje pomaga te obsežne vire podatkov učinkovito obdelati ter nad njimi razviti inovativne informacijske rešitve in storitve, ki prinašajo uporabnikom vedno nove funkcionalnosti. Sočasno z razvojem tehnologij se spreminjajo tudi koncepti njihove uporabe, uporabniki so ozavestili inteligentno obnašanje naprav ter storitev, in to od ponudnikov v vedno večji meri tudi zahtevajo. Podjetja so tako primorana nadgrajevati ter razvijati produkte in rešitve s pomočjo umetne inteligence oziroma z uporabo metod in tehnologij strojnega učenja. Tovrsten razvoj predstavlja za razvijalce vse kompleksnejše zahteve in izzive, za revizorje pa nova in pogosto še neznana tveganja. Kljub velikanskemu napredku na področju strojnega učenja je to še vedno zahtevna naloga, ki zahteva obilico domenskega znanja. Izhajajoč iz dveh desetletij praktičnih izkušenj pri razvoju rešitev umetne inteligence, bomo predstavili pogoste pasti in tveganja, katerim so izpostavljeni premalo izkušeni razvijalci, ki pod vplivom pogosto prevelikih pričakovanj in nerealnih zahtev vodstva vidijo metode umetne inteligence kot novo "silver bullet" tehnologijo. Nakazali bomo tudi nekaj možnosti, s katerimi lahko revizorji tovrstne rešitve kontrolirajo, in na kaj morajo biti pri tem pozorni.

The development of information technology has brought us to a period of increasingly rapid creation, sharing and exchange of information. The introduction of artificial intelligence methods is increasingly helping us to process these extensive data sources efficiently and to develop innovative information solutions and services above them, which bring users new functionalities. Simultaneously With the development of technology, the concepts of their use are also changing. The users have realized the intelligent behaviour of devices and services, and they are increasingly demanding this from providers. Companies are thus compelled to upgrade and develop products and solutions towards integration with machine learning methods and artificial intelligence. Such development poses evermore complex requirements and challenges for developers, and new and often as yet unknown risks for auditors. Despite the enormous progress in machine learning, the latter is still a challenging task requiring an abundance of domain knowledge and experience. Drawing from two decades of practical experience in developing artificial intelligence solutions, we will present some common pitfalls and risks to which inexperienced developers are exposed. Under the influence of often excessive expectations and unrealistic requirements from the management, they see artificial intelligence methods as a new "silver bullet" technology. We will point out some of the options that auditors can use to control this type of solutions and what they need to look out for.