Luc Julia, Sci­en­tif­ic Direc­tor with the Group Renault, is also the author of the book “L’intelligence arti­fi­cielle n’existe pas” [“AI doesn’t exist”], pub­lished by First Edi­tions. In favour of reg­u­lat­ing AI, he nonethe­less calls for intel­li­gent regulation.

Although the expres­sion ‘AI’ was first coined in 1956*, intel­li­gent machines designed to per­form spe­cif­ic tasks have always exist­ed. Such is the case, accord­ing to Luc Julia, of Pascal’s machine in the 17th Cen­tu­ry, or the aba­cus that appeared 1 000 years ear­li­er. “The arti­fi­cial intel­li­gence that doesn’t exist is the one that works like ours. On the con­trary, all arti­fi­cial intel­li­gences are tools designed to han­dle a giv­en task and, in this respect, they are bet­ter than we are”, he believes.

From expert sys­tems to machine learn­ing, deep learn­ing and, more recent­ly, gen­er­a­tive AI, these tools have evolved over time. “AIs are just tools in the tool­box that is AI in gen­er­al. Just like the ham­mer, nails, screw­driv­er and screws found in a con­ven­tion­al tool­box. Each AI is spe­cif­ic and designed for a par­tic­u­lar prob­lem. But you have to be care­ful, since, as with a ham­mer, which I can use both to dri­ve a nail or to smash someone’s head, AI can be used for good or bad pur­pos­es. Hence the need for reg­u­la­tion, even if, at the end of the day, it’s still a per­son hold­ing the ham­mer han­dle”, he asserts.

While Luc Julia stress­es the impor­tance of reg­u­lat­ing AI both in terms of the tools devel­oped — what kind of tools are accept­able? — and in the use to which they are put, he is nonethe­less scep­ti­cal about the “Euro­pean AI Act” — the first com­pre­hen­sive reg­u­la­tion of AI on a glob­al scale by a reg­u­la­to­ry insti­tu­tion of this impor­tance. In par­tic­u­lar, he fears the dis­in­cen­tives to inno­va­tion it could entail and advo­cates smart reg­u­la­tion. “EU reg­u­la­tion clas­si­fies AI appli­ca­tions into three lev­els of risk: appli­ca­tions and sys­tems that cre­ate an unac­cept­able risk such as the social rat­ing sys­tems deployed in some coun­tries, high-risk appli­ca­tions such as CV scan­ning tools that could clas­si­fy appli­cants accord­ing to gen­der or eth­nic­i­ty, for exam­ple, and final­ly appli­ca­tions that are not list­ed as high-risk. A clas­si­fi­ca­tion of this nature is too gen­er­al and ignores cer­tain gran­u­lar­i­ties. Take facial recog­ni­tion, for exam­ple. One imme­di­ate­ly thinks of gen­er­al­ized sur­veil­lance of pop­u­la­tions, which is of course unac­cept­able, yet it can be use­ful for facial repair in surgery. The EU may well come to accept its use in this field, but the prob­lem is its lack of agili­ty. If the EU could tell me with­in a month that facial recog­ni­tion applied to surgery is no longer con­cerned by this lev­el of risk, I’d be delight­ed. The prob­lem is that it will take three years while oth­er coun­tries move for­ward,” con­cludes Luc Julia.

* AI as a sci­en­tif­ic dis­ci­pline can trace its roots back to the Dart­mouth Sum­mer Research Project on Arti­fi­cial Intel­li­gence, held at Dart­mouth Col­lege in 1956.

MSD