Artificial general intelligence: how will it be regulated?

Arti­fi­cial Intel­li­gence (AI) is cur­rent­ly boom­ing and trans­form­ing many aspects of our dai­ly lives. It opti­mis­es the oper­a­tion of search engines and makes it pos­si­ble to analyse queries more effec­tive­ly in order to pro­pose the most rel­e­vant results¹. It improves sur­veil­lance sys­tems, which now use it to detect sus­pi­cious behav­iour². It offers invalu­able assis­tance in the health­care sec­tor for analysing med­ical images, devel­op­ing new drugs and per­son­al­is­ing treat­ments³. How­ev­er, there is a fun­da­men­tal dis­tinc­tion between the AI we know today, often referred to as “clas­si­cal AI”, and a more ambi­tious con­cept: Arti­fi­cial Gen­er­al Intel­li­gence (AGI).

Clas­si­cal AI is designed to excel at spe­cif­ic tasks and can out­per­form the best experts or spe­cialised algo­rithms. AGI, on the oth­er hand, aspires to an intel­li­gence com­pa­ra­ble to that of a human being. It aims to under­stand the world in all its com­plex­i­ty, to learn autonomous­ly and to adapt to new sit­u­a­tions. In oth­er words, AGI would be capa­ble of solv­ing a wide vari­ety of prob­lems, rea­son­ing, cre­at­ing and being self-aware⁴.

Warn­ings about the rise of gen­er­al-pur­pose AI are mul­ti­ply­ing, point­ing to a bleak future for our civil­i­sa­tion. Sev­er­al lead­ing fig­ures in the world of tech­nol­o­gy have warned of the harm­ful effects of this tech­nol­o­gy. Stephen Hawk­ing has expressed fears that AI could sup­plant humans, lead­ing to a new era in which machines could dom­i­nate⁵. Emi­nent Amer­i­can pro­fes­sors, such as Stu­art Rus­sell, Pro­fes­sor at the Uni­ver­si­ty of Cal­i­for­nia, Berke­ley, have also high­light­ed the shift towards a uni­verse where AI will play a role that is unknown at this stage, with new risks to be tak­en into account and antic­i­pat­ed⁶. Fur­ther­more, Jerome Glenn of the Mil­len­ni­um Project has stat­ed⁷ that “gov­ern­ing AGI could be the most com­plex man­age­ment prob­lem human­i­ty has ever faced” and that “the slight­est mis­take could wipe us off the face of the Earth.” These asser­tions sug­gest an extreme­ly pes­simistic, even cat­a­stroph­ic, out­look on the devel­op­ment of the AGI.

A fun­da­men­tal crit­i­cism of the immi­nence of AGI is based on the “prob­lem of the com­plex­i­ty of val­ues”, a key con­cept addressed by Nick Bostrom in Super­in­tel­li­gence: Paths, Dan­gers, Strate­gies⁸. The evo­lu­tion­ary process of human life and civil­i­sa­tion spans bil­lions of years, with the devel­op­ment of numer­ous com­plex sys­tems of feel­ings, but also of con­trols and val­ues, thanks to the many and var­ied inter­ac­tions with an envi­ron­ment that is phys­i­cal, bio­log­i­cal, and social. From this per­spec­tive, it is hypoth­e­sised that an autonomous and high­ly sophis­ti­cat­ed AGI can­not be achieved in just a few decades.

The Aus­tralian Rod­ney Brooks, one of the icons and pio­neers of robot­ics and the­o­ries of “embod­ied cog­ni­tion”, main­tains that what will deter­mine whether an intel­li­gence is tru­ly autonomous and sophis­ti­cat­ed is its inte­gra­tion with­in a body and con­tin­u­ous inter­ac­tion with a com­plex envi­ron­ment over a suf­fi­cient­ly long peri­od⁹. These ele­ments rein­force the the­sis that AGI, as described in the alarmist sce­nar­ios, is still a long way from becom­ing a reality.

Recent years have seen the rise of large lan­guage mod­els (LLMs) such as Chat­G­PT, Gem­i­ni, Copi­lot and so on. These have demon­strat­ed an impres­sive abil­i­ty to assim­i­late many implic­it human val­ues, based on mas­sive analy­sis of writ­ten doc­u­ments. Because of its archi­tec­ture and the way it works, Chat­G­PT has a num­ber of lim­i­ta­tions¹⁰. It does not sup­port log­i­cal rea­son­ing, its respons­es are some­times unre­li­able, its knowl­edge base is not adapt­ed in real-time, and it is sus­cep­ti­ble to “prompt injec­tion” attacks. Although these mod­els have sophis­ti­cat­ed val­ue sys­tems, they do not appear to be autonomous. In fact, they do not seem to aim for auton­o­my or self-preser­va­tion with­in an envi­ron­ment that is both com­plex and vari­able. In this respect, it is impor­tant to remem­ber that a very impor­tant part of com­mu­ni­ca­tion is linked to into­na­tion and body lan­guage¹¹, ele­ments that are not at all con­sid­ered in inter­ac­tions with gen­er­a­tive AIs.

A sim­ple reminder of this (pro­found) dis­tinc­tion seems cru­cial to bet­ter under­stand the extent to which con­cerns over mali­cious super­in­tel­li­gence are unfound­ed and exces­sive. Today, LLMs can only be con­sid­ered as par­rots pro­vid­ing prob­a­bilis­tic answers (“sto­chas­tic par­rots” accord­ing to Emi­ly Ben­der¹²). Of course, they rep­re­sent a break with the past, and it appears nec­es­sary to reg­u­late their use now.

It seems to us that future intel­li­gence can­not be “arti­fi­cial” in the strict sense of the word, i.e. designed from scratch. But it would be high­ly col­lab­o­ra­tive, emerg­ing from the knowl­edge (and even wis­dom) accu­mu­lat­ed by humankind. It is real­is­tic to con­sid­er that cur­rent AIs, as such, are large­ly tools and embod­i­ments of col­lec­tive thought pat­terns, tend­ing towards benev­o­lence rather than con­trol or dom­i­na­tion. This col­lec­tive intel­li­gence is noth­ing less than a deep mem­o­ry that is nour­ished by civilised val­ues such as help­ing those in need, respect for the envi­ron­ment and respect for oth­ers. We there­fore need to pro­tect this intan­gi­ble her­itage and ensure that it is aimed at pro­vid­ing sup­port and help to human beings rather than trans­mit­ting mis­in­for­ma­tion or incit­ing them to com­mit rep­re­hen­si­ble acts. At the risk of being Manichean, LLMs can be used for good¹³, but they can also be used for evil¹⁴.

From a log­i­cal point of view, alarmist sce­nar­ios in which mali­cious actors would be led, in the short term, to pro­gramme man­i­fest­ly harm­ful objec­tives into the heart of AI appear a pri­ori to be exag­ger­at­ed. The argu­ment of the com­plex­i­ty of val­ues sug­gests that these neg­a­tive val­ues would be poor­ly inte­grat­ed into the mass of pos­i­tive val­ues learned. Fur­ther­more, it seems like­ly that well-inten­tioned pro­gram­mers (white hats) will cre­ate AIs that can counter the destruc­tive strate­gies of mali­cious AIs (black hats). This could lead, quite nat­u­ral­ly, to a clas­sic “arms race”. Anoth­er counter-argu­ment to a mali­cious takeover of AIs is their eco­nom­ic poten­tial. At present, AI for the gen­er­al pub­lic is being dri­ven by major play­ers in the eco­nom­ic sec­tor (Ope­nAI, Google, Microsoft, etc.), at least some of whom have a prof­it ratio­nale. This requires user con­fi­dence in the use of the AI made avail­able, but also the preser­va­tion of the data and algo­rithms that make up AI as an intan­gi­ble asset at the heart of eco­nom­ic activ­i­ty. The resources required for pro­tec­tion and cyber-defence will there­fore be considerable.

Ini­tia­tives have already been tak­en to reg­u­late spe­cialised AI. How­ev­er, the reg­u­la­tion of arti­fi­cial gen­er­al intel­li­gence will require spe­cif­ic mea­sures. One such ini­tia­tive is the AI Act cur­rent­ly being draft­ed by the Euro­pean Union¹⁵. The authors make the fol­low­ing addi­tion­al proposals:

• The intro­duc­tion of a sys­tem of nation­al licences to ensure that any new AGI com­plies with the nec­es­sary safe­ty standards,
• Sys­tems for ver­i­fy­ing the safe­ty of AI in con­trolled envi­ron­ments before they are autho­rised and deployed,
• The devel­op­ment of more advanced inter­na­tion­al coop­er­a­tion, which could lead to UN Gen­er­al Assem­bly res­o­lu­tions and the estab­lish­ment of con­ven­tions on AI.

Ratio­nal reg­u­la­tion of AI requires an informed analy­sis of the issues at stake and a bal­ance between pre­vent­ing risks and pro­mot­ing ben­e­fits. Inter­na­tion­al insti­tu­tions and tech­ni­cal experts will play an impor­tant role in coor­di­nat­ing the efforts required for the safe and eth­i­cal devel­op­ment of AI. Good gov­er­nance and effec­tive reg­u­la­tion of the AGI will require a dis­pas­sion­ate approach.