Demystifying generative AI: true, false, uncertain

Generative AI, a different type of AI – TRUE

In the realm of arti­fi­cial intel­li­gence, there are a num­ber of dif­fer­ent types. Among these, gen­er­a­tive AI, as its name sug­gests, stands out for its abil­i­ty to gen­er­ate con­tent: text, images, video, etc. Some of the best-known cur­rent sys­tems are Chat­G­PT, Bard, Mid­jour­ney and DALL‑E.

Their prin­ci­ple is based on prob­a­bil­i­ties: they pre­dict the next word or the neigh­bour­ing pix­el, accord­ing to what seems most like­ly. To do this, gen­er­a­tive AI relies on a large lan­guage mod­el, i.e. a deep net­work of arti­fi­cial neu­rons that has been trained on a vast amount of data. In this way, the soft­ware iden­ti­fies the most like­ly match­es accord­ing to the context.

Generative AI is intelligent – FALSE

This is how gen­er­a­tive AI per­forms remark­ably well. They are able to estab­lish links between mul­ti­ple ele­ments, based on an enor­mous vol­ume of data. This is a com­plex process, involv­ing a large num­ber of math­e­mat­i­cal oper­a­tions, car­ried out very quickly.

But can we real­ly call this “intel­li­gence”? While the results can be aston­ish­ing, the way in which they are achieved has noth­ing to do with human intel­li­gence. Nor is it “gen­er­al AI”, capa­ble of learn­ing any task per­formed by a human being. Today, gen­er­a­tive AI is more like a mul­ti­pli­ca­tion of nar­row AIs, brought togeth­er with­in the same model.

Generative AI can do anything – UNCERTAIN

Gen­er­a­tive AI is cur­rent­ly used in many fields: some use it to cre­ate music, oth­ers video game land­scapes… As for the lan­guage mod­els ini­tial­ly used to cap­ture the seman­tics of words, they can now gen­er­ate text, answer ques­tions, trans­late con­tent, or even gen­er­ate code. But these tools still have their lim­i­ta­tions, linked in par­tic­u­lar to the datasets used dur­ing their train­ing. Cor­re­la­tions iden­ti­fied at this stage can lead to errors at the gen­er­a­tion stage. In addi­tion, any bias­es encoun­tered dur­ing the train­ing phase are reflect­ed in the results. For exam­ple, a trans­la­tion sys­tem will tend to trans­late “the nurse” (in Eng­lish) as “l’infirmière” (in French, a fem­i­nine ver­sion of the word), because of the stereo­types asso­ci­at­ed with the profession.

Gen­er­a­tive AIs are not always very sta­ble. Try it out with Chat­G­PT: ask the same ques­tion but vary the word­ing, and you’ll some­times get dif­fer­ent answers! These sys­tems are based on math­e­mat­i­cal oper­a­tions that trans­form infor­ma­tion into high-dimen­sion­al vec­tors, which makes them dif­fi­cult to explain. Research is cur­rent­ly under­way on this subject.

Generative AI can be wrong – TRUE

It is impor­tant to bear in mind that gen­er­a­tive AI does not aim to deliv­er the truth, but to max­imise plau­si­bil­i­ty, based on its train­ing data. It some­times pro­duces false cor­re­la­tions between words. What’s more, if the train­ing data con­tains errors or bias­es, the sys­tem will inevitably repro­duce them. In any case, it does not seek to know whether the infor­ma­tion pro­vid­ed is accu­rate or sourced! This leads to the fre­quent and unpre­dictable appear­ance of “hal­lu­ci­na­tions”, i.e. incor­rect respons­es or inco­her­ent images.

For exam­ple, accord­ing to a study by the Uni­ver­si­ty of Hong Kong¹, Chat­G­PT (ver­sion GPT‑3.5) has an accu­ra­cy rate of 64%. Would you take the word of some­one who has more than a one in three chance of being wrong?

Generative AI will rebel and take over – FALSE

As soon as arti­fi­cial intel­li­gence seems to reach a new stage, fan­tasies of machine upris­ings, influ­enced by sci­ence fic­tion, resur­face. We shouldn’t indulge in exces­sive anthro­po­mor­phism: gen­er­a­tive AI sim­ply pre­dicts prob­a­bil­i­ties – in admit­ted­ly com­plex ways. They do not feel emo­tions, nor do they have con­scious­ness. So, they can­not have a “will” to rebel.

In 2015, the Amer­i­can AI researcher Andrew Ng² said that fear­ing a pos­si­ble AI revolt was like “wor­ry­ing about over­pop­u­la­tion on Mars”, when “we’ve nev­er set foot on the plan­et before.” Even if the tech­nol­o­gy has evolved con­sid­er­ably in recent years, the com­par­i­son still rings true!

Generative AI raises security and confidentiality issues – UNCERTAIN

Today, we need to be aware that most gen­er­a­tive AI mod­els are host­ed on Amer­i­can servers. Under the Patri­ot Act and the Cloud Act, the data sent can be retrieved by the Amer­i­can author­i­ties. In addi­tion, the data sup­plied to these gen­er­a­tive AIs is undoubt­ed­ly reused to improve the mod­els, mak­ing it pos­si­ble to retrieve this data in future queries. This can there­fore rep­re­sent a risk, par­tic­u­lar­ly for busi­ness­es, whose data secu­ri­ty and con­fi­den­tial­i­ty are under threat. There are, how­ev­er, host­ing solu­tions with ded­i­cat­ed, closed spaces, or open-source gen­er­a­tive AI alter­na­tives that can be installed on local servers.

How­ev­er, as is often the case, reg­u­la­tions even­tu­al­ly adapt to the new tech­no­log­i­cal con­text. For exam­ple, at the end of 2023, the Coun­cil of the Euro­pean Union and the Euro­pean Par­lia­ment reached agree­ment on leg­is­la­tion on arti­fi­cial intel­li­gence (AI Act)³. The text will undoubt­ed­ly be refined, but it will pro­vide a bet­ter frame­work for the use of AI, in com­pli­ance with Euro­pean law (includ­ing the RGPD).

Generative AI can replace humans for certain tasks – TRUE

The capa­bil­i­ties of gen­er­a­tive AI make it very use­ful in the pro­fes­sion­al sphere. It can draft con­tent, write lines of code, draw up a train­ing plan, etc. But what it pro­duces gen­er­al­ly requires the human eye to check its accu­ra­cy, per­son­alise the mes­sage, add a more sen­si­tive touch, etc. It is there­fore a tool for increas­ing pro­duc­tiv­i­ty, free­ing up time to work differently.

Some jobs could dis­ap­pear, how­ev­er, for lack of suf­fi­cient added val­ue. But isn’t that always the case with tech­ni­cal progress? Didn’t lamp­lighters, for exam­ple, dis­ap­pear with the advent of elec­tric lighting?

Generative AI will put millions out of work – FALSE

But let’s remain mea­sured about the pro­fes­sion­al con­se­quences of gen­er­a­tive AI. At the end of the day, it’s just a new tool – a very use­ful one – at the ser­vice of human beings. And changes in the labour mar­ket depend on many para­me­ters… Have auto­mat­ic check­outs com­plete­ly replaced the need for cashiers? Has e‑learning replaced schools and teachers?

What’s more, the rise of gen­er­a­tive AI is like­ly to be accom­pa­nied by new pro­fes­sions, such as prompt engi­neer­ing, a dis­ci­pline that aims to opti­mise the queries for­mu­lat­ed to the AI, so as to obtain the best pos­si­ble results. Accord­ing to the Inter­na­tion­al Labour Organ­i­sa­tion (ILO)⁴, “gen­er­a­tive AI is more like­ly to increase than destroy jobs by automat­ing cer­tain tasks rather than replac­ing a role entirely”.

How far will generative AI go? – UNCERTAIN

What will hap­pen in the long term? How will gen­er­a­tive AI evolve? Pre­dict­ing its future is tricky: who could have pre­dict­ed the cur­rent sit­u­a­tion a few years ago? Nev­er­the­less, cer­tain trends are emerg­ing, such as the hybridi­s­a­tion of sys­tems. For exam­ple, RAG (retrieval-aug­ment­ed gen­er­a­tion) involves com­bin­ing gen­er­a­tive AI with a search engine to improve the rel­e­vance of results and lim­it hallucinations.

Final­ly, gen­er­a­tive AI can­not devel­op with­out address­ing its eco­log­i­cal foot­print. Its mod­els require an enor­mous amount of data and com­put­ing pow­er. A new approach is already being explored to opti­mise the nec­es­sary resources: fru­gal AI.

Bastien Contreras