The Ger­man Fede­ral Office for Infor­ma­ti­on Secu­ri­ty (BSI) is the cen­tral body in Ger­ma­ny for infor­ma­ti­on secu­ri­ty at the natio­nal level and the aut­hor of the BSI basic pro­tec­tion. One of its focal points is arti­fi­ci­al intel­li­gence; to this end, the BSI main­ta­ins a Topic pagewhe­re you can find stu­dies worth reading.

On August 5, 2024, the BSI published a rela­tively short noti­ce dated July 1, 2024. White paper on trans­pa­ren­cy in AI systems has been published. It is unli­kely to remain visi­ble for long in the ple­tho­ra of publi­ca­ti­ons, but attempts to shed light on the topic from a fun­da­men­tal perspective.

Accor­ding to the BSI, trans­pa­ren­cy in AI systems means pro­vi­ding infor­ma­ti­on about the system, inclu­ding its limi­ta­ti­ons, over the enti­re life cycle, from plan­ning to design, deve­lo­p­ment, vali­da­ti­on, com­mis­sio­ning, use and ongo­ing eva­lua­ti­on. Trans­pa­ren­cy should not only enable stake­hol­ders to make infor­med decis­i­ons, but also streng­then trust­wort­hi­ness and the pro­tec­tion of fun­da­men­tal rights, as requi­red by the AI Act. The paper refers in par­ti­cu­lar to Art. 13 AIA (fun­da­men­tal Trans­pa­ren­cy requi­re­ment for high-risk AI systems; basis for the Ope­ra­ting ins­truc­tionswith which the Pro­vi­der must inform the Deployer in accordance with Art. 3(15) AEOI (this obli­ga­ti­on is not assi­gned to the Pro­vi­der in Art. 16 AEOI, but rather in the defi­ni­ti­on of the ope­ra­tio­nal deri­va­ti­on) and in accordance with Art. 53 AEOI (obli­ga­ti­on of pro­vi­ders of GPAIto crea­te tech­ni­cal docu­men­ta­ti­on for the system and make it available to pro­vi­ders who­se systems are based on the GPAI). Howe­ver, the­re are various other pro­vi­si­ons in the AIA that are in the ser­vice of trans­pa­ren­cy, in par­ti­cu­lar Art. 50 for Chat­bots and other systems desi­gned to inter­act direct­ly with natu­ral persons).

It is inte­re­st­ing to note that the BSI right­ly points out that trans­pa­ren­cy can also be harmful – it can reve­al new attack vec­tors and limi­ta­ti­ons can be exploi­ted. An appro­pria­te balan­ce must the­r­e­fo­re be found.