Are we moving towards global regulation of AI?

There are a num­ber of issues to con­sid­er. First, there is the fear of an exis­ten­tial risk fuelled by sto­ries of an AI that could become autonomous and destroy human­i­ty. How­ev­er, I don’t see this as a real pos­si­bil­i­ty – at least not with the mod­els that are being devel­oped right now. Then there’s the fear of an oli­gop­oly, with heavy depen­dence on a hand­ful of com­pa­nies. There is also the risk of human rights vio­la­tions. And in this case, a new fac­tor comes into play: for a long time, the most ter­ri­ble actions were reserved for the coun­tries and armies of the most pow­er­ful coun­tries. Now it’s com­mon­place, mass-mar­ket tech­nol­o­gy. The same AI that can detect can­cer could also be used to pre­vent part of the pop­u­la­tion from enter­ing air­ports. We have rarely seen such mul­ti-pur­pose tech­nolo­gies. AI also makes mali­cious acts pos­si­ble, such as deep fakes, attacks on our sys­tems, or the manip­u­la­tion of opin­ion. Not to men­tion the imbal­ances it will cre­ate in terms of intel­lec­tu­al prop­er­ty, pro­tec­tion of the pub­lic domain, pro­tec­tion of pri­va­cy, and the upheavals in the work­place. All these chal­lenges, cou­pled with the desire to real­ly ben­e­fit from AI’s bound­less promise, have prompt­ed intense glob­al reflec­tion on a frame­work for shared governance.

The great­est risk, in my opin­ion, is that of a monop­oly. It’s a ques­tion of democ­ra­cy. One of the great dan­gers is that the glob­al econ­o­my, soci­ety and the media will become ultra-depen­dent on a small oli­gop­oly that we are unable to reg­u­late. In line with this analy­sis, I’m try­ing to push for the pro­tec­tion of the dig­i­tal com­mons and open source. We need to make sure that there are mod­els that are free to use so that every­one can ben­e­fit from them. There are already high-qual­i­ty open source mod­els. So, the ques­tion is: are we going to put enough pub­lic mon­ey into train­ing them for the ben­e­fit of every­one? In my opin­ion, that’s where the real bat­tle lies. Ensur­ing that there are resources for the pub­lic good and that it will be pos­si­ble to inno­vate with­out ask­ing for permission.

Among the aspects that could lead to inter­na­tion­al coor­di­na­tion are the indi­rect effects of AI. These tech­nolo­gies could dis­rupt the way in which we have con­struct­ed the pro­tec­tion of pri­va­cy. Today, the gen­er­al prin­ci­ple is to pro­tect per­son­al data in order to pro­tect indi­vid­u­als. With AI, by using pre­dic­tive mod­els, we can learn a great deal, if not every­thing, about a per­son. Pre­dic­tive mod­els can take into account a person’s age, where they live and where they work, and give a very good prob­a­bil­i­ty of their risk of can­cer or their like­li­hood of lik­ing a par­tic­u­lar film. Anoth­er issue being debat­ed at an inter­na­tion­al lev­el is that of intel­lec­tu­al prop­er­ty. It is now pos­si­ble to ask an AI to pro­duce paint­ings in the style of Kei­th Har­ing and to sell them, which pos­es a prob­lem for the artist’s beneficiaries. 

Ten years ago, there was a tac­it agree­ment not to reg­u­late social net­works. Now these com­pa­nies are worth $1,000bn and it’s very dif­fi­cult to change their tra­jec­to­ry. Most devel­oped coun­tries are telling them­selves that they won’t make the same mis­take again and that they mustn’t miss the boat this time. But that means they know what to do. There is a new aware­ness that reg­u­la­tion must be anchored in an inter­na­tion­al frame­work. There are so few bor­ders in the dig­i­tal world that you can set up in any coun­try and still have a pres­ence in anoth­er. Clear­ly, there is intense inter­na­tion­al activ­i­ty around the major issues men­tioned above: exis­ten­tial risk, the ques­tion of eco­nom­ic sov­er­eign­ty and mali­cious actors.

No. The “region­al” lev­el (i.e. a coher­ent group of coun­tries) is also very impor­tant. Europe has learned the hard way that only reg­u­lat­ing all dig­i­tal uses at nation­al lev­el is not enough to bring the giants to heel. When we estab­lish a Euro­pean frame­work, they nego­ti­ate. But that cre­ates oth­er ten­sions, and we don’t want to encour­age an inter­na­tion­al order based on extra-ter­ri­to­r­i­al appli­ca­tion deci­sions. So, the idea that the inter­na­tion­al lev­el is the right size for think­ing about the dig­i­tal world has tak­en hold, and it’s not real­ly ques­tioned any more.

Under the law, we have the right to pro­hib­it the devel­op­ment of cer­tain AI. But we are afraid that the oth­er pow­ers will con­tin­ue to do so, and that we will become weak and obso­lete. It’s a ques­tion of being both pro-inno­va­tion and pro-secu­ri­ty for cit­i­zens, and that’s why every­one would like deci­sions to be col­lec­tive. These tech­nolo­gies are chang­ing very fast, and cre­at­ing a lot of pow­er, so we don’t want uni­lat­er­al disarmament.

The eth­i­cal frame­work is being dis­cussed. Dis­cus­sions are tak­ing place in dozens of forums with­in busi­ness, civ­il soci­ety, among researchers, at the UN, the G7, the OECD and as part of the French ini­tia­tive for a glob­al part­ner­ship for arti­fi­cial intel­li­gence. It’s also a diplo­mat­ic effort that’s com­ing out of the embassies, with debates at the Inter­net Gov­er­nance Forum and the annu­al sum­mit on human rights, RightsCon. Lit­tle by lit­tle, ideas are tak­ing root or becom­ing estab­lished. We are still in the process of iden­ti­fy­ing the con­cepts on which an agree­ment will be based. An ini­tial con­sen­sus is emerg­ing around cer­tain prin­ci­ples: no use should be made that is con­trary to human rights; the tech­nol­o­gy must be in the inter­est of the users; it must be proven that pre­cau­tions have been tak­en to ensure that there is no bias in the edu­ca­tion of the mod­els; the need for trans­paren­cy so that experts can audit the mod­els. Then it will be time to look for treaties.

There are also debates about a demo­c­ra­t­ic frame­work. How can we ensure that these com­pa­nies are not manip­u­lat­ing us? Do we have the right to know what data the AI has been trained with? The notions of secu­ri­ty in the face of exis­ten­tial risk were much dis­cussed in the UK at last year’s world sum­mit. Now the con­ver­sa­tion is turn­ing to the future of work and intel­lec­tu­al prop­er­ty, for exam­ple. In 2025, France will be host­ing a major inter­na­tion­al sum­mit on AI, which will help move these issues forward.

Many peo­ple think not. Per­son­al­ly, I think that a good text is fun­da­men­tal. The Dec­la­ra­tion of Human Rights is still valid, but tech­nolo­gies have changed. The 1978 Data Pro­tec­tion Act has become the GDPR, but the prin­ci­ple of user con­sent for data to be cir­cu­lat­ed has not aged a day. If we can find robust prin­ci­ples, we can pro­duce texts that will stand the test of time. I think we could reg­u­late AI with the GDPR, the respon­si­bil­i­ty of the media and con­tent pub­lish­ers, and two or three oth­er texts that already exist. It’s not a giv­en that we need an entire­ly new framework.

Sirine Azououai