Artificial general intelligence: the true, the false, and the uncertain

It is only a mat­ter of months. Accord­ing to sev­er­al Sil­ic­on Val­ley heavy­weights, the arrival of arti­fi­cial gen­er­al intel­li­gence (AGI), arti­fi­cial intel­li­gence that can match human intel­li­gence, is immin­ent. Unlike today’s spe­cial­ised AI sys­tems, AGI could learn, reas­on, adapt and trans­fer its know­ledge from one domain to anoth­er without being spe­cific­ally trained for each new task. Bey­ond the hype and mar­ket­ing rhet­or­ic that is dif­fi­cult to assess sci­en­tific­ally, AGI and new cog­nit­ive sys­tems are rede­fin­ing our under­stand­ing of intel­li­gence. To bet­ter under­stand the implic­a­tions, Laurence Dev­illers, a pro­fess­or at Sor­bonne Uni­ver­sity and research­er at the CNRS in the Inter­dis­cip­lin­ary Labor­at­ory for Digit­al Sci­ences (LISN) in Saclay, shares her expert­ise on AI.

TRUE

OpenAI, Google Deep­Mind and Anthrop­ic make no secret of it: their aim is to devel­op arti­fi­cial gen­er­al intel­li­gence (AGI) cap­able of per­form­ing any cog­nit­ive task as well as we can, or even a super­in­tel­li­gence that would sur­pass that of humans.

In early 2025, five renowned research­ers in the field of AI (includ­ing Daniel Kokota­jlo, formerly of OpenAI’s gov­ernance divi­sion) under­took a for­ward-look­ing exer­cise that sought to describe pos­sible scen­ari­os for the evol­u­tion of AI. The report, entitled AI 2027¹, con­cluded that it was plaus­ible that AI sys­tems would sur­pass human cog­nit­ive per­form­ance as early as 2027 in a large num­ber of tasks. In Feb­ru­ary 2026, the authors reas­sessed their pre­dic­tions² and con­cluded that the advent of AGI would take slightly longer than anti­cip­ated: the new estim­ate points to a time­frame between 2029 and 2032. Accord­ing to them, “the impact of super­in­tel­li­gence over the next dec­ade will be enorm­ous, sur­pass­ing that of the Indus­tri­al Revolution.”

UNCERTAIN

The term AGI is over­used and lacks a con­sist­ent defin­i­tion or shared meas­ure­ment threshold. This ambi­gu­ity was illus­trated in 2026 by the dis­crep­ancy between the CEO of NVIDIA, who claimed that AGI had already been achieved, and the ARC-AGI‑3 bench­mark³, a test designed to eval­u­ate agent-based AI sys­tems. Accord­ing to the bench­mark, the most advanced mod­els (OpenAI, Google, Anthrop­ic) were capped at 1 % on reas­on­ing tasks that humans solve 100 % of the time, due to a lack of train­ing suited to this type of envir­on­ment. As things stand, there is there­fore no sci­entif­ic data to con­firm this time­frame, nor even wheth­er such a sys­tem is achiev­able. For the research­er, there is “a real buzz sur­round­ing the immin­ent arrival of AGI, fuelled by Big Tech’s mar­ket­ing rather than by sci­ence.” That said, the lack of trans­par­ency on the part of com­pan­ies leaves the aca­dem­ic world some­what in the dark: there is a genu­ine lack of under­stand­ing regard­ing the exact nature of these sys­tems. This hype sur­round­ing AGI there­fore hinges largely on the com­mer­cial appet­ite sur­round­ing it: AI can be owned and sold, and entre­pren­eurs see it as an oppor­tun­ity to demon­strate their power to reshape soci­ety and the world of work, a com­pet­it­ive mind­set that is omni­present in cur­rent dis­cus­sions sur­round­ing AGI.

UNCERTAIN

The latest LLMs (large lan­guage mod­els) are incred­ibly power­ful. They gen­er­ate incred­ible text and images, and they undeni­ably appear intel­li­gent. They are designed to mim­ic human inter­ac­tions, sus­tain cred­ible con­ver­sa­tions, but primar­ily to tell stor­ies, regard­less of wheth­er they are true or false. So, if you scratch beneath the sur­face, the story is rather more pro­sa­ic. LLMs are, first and fore­most, highly soph­ist­ic­ated stat­ist­ic­al archi­tec­tures, cap­able of identi­fy­ing lin­guist­ic abstrac­tions and pat­terns without, how­ever, achiev­ing a genu­ine under­stand­ing of the world, sub­ject­ive exper­i­ence or true intentionality.

“We’ve had to incor­por­ate super­vised learn­ing (or fine-tun­ing, which involves retrain­ing an LLM on examples spe­cif­ic to a par­tic­u­lar domain) and, by exten­sion, rein­force­ment learn­ing based on human feed­back (RLHF),” notes Laurence Dev­illers. These steps are neces­sary to lim­it hal­lu­cin­a­tions, responses that appear coher­ent but are in fact false, fab­ric­ated and, above all, not based on real data. 

Fur­ther­more, align­ment refers to the action taken by humans to make sys­tems more robust, reli­able, con­trol­lable, assess­able and veri­fi­able; and to ensure that an AI sys­tem using LLMs behaves in ways that align with the inten­tions of developers and end-users, includ­ing when faced with quer­ies spe­cific­ally designed to manip­u­late and deceive (“adversari­al prompts”), dis­tri­bu­tions of data unknown dur­ing train­ing, or “unfore­seen oper­a­tion­al con­texts”, as out­lined by Van­ina Paoli-Gagin, Sen­at­or for Aube and mem­ber of the Par­lia­ment­ary Mis­sion on the Align­ment of Arti­fi­cial Intel­li­gence Sys­tems (SIA). 

For a long time, using more data has been syn­onym­ous with bet­ter per­form­ance, but this cor­rel­a­tion is now reach­ing its lim­its, which is fuel­ling the debate between those who believe that AGI will ulti­mately achieve this simply through scale and those who believe that archi­tec­tur­al break­throughs or new types of data (mul­timod­al, inter­act­ive, embod­ied) are needed to break through this ceiling.

If we ever achieve AGI, it will prob­ably be thanks to mod­els that dif­fer from gen­er­at­ive mod­els, and not­ably, that incor­por­ate a phys­ic­al rep­res­ent­a­tion of the world (a “world model”).

TRUE & FALSE

Human intel­li­gence dif­fers from that of cur­rent AI in its ver­sat­il­ity. Humans can learn new tasks without hav­ing been spe­cific­ally trained to do so, reas­on in the face of nov­el situ­ations, and act on intu­ition that includes a cer­tain innate under­stand­ing of the laws of phys­ics. Young chil­dren, for example, quickly learn numer­ous con­cepts (the effects of grav­ity, the phys­ic­al con­sequences of an action, people’s inten­tions…) without need­ing to be exposed to the same situ­ation count­less times. That said, this ver­sat­il­ity is not abso­lute either: humans can­not do everything.

As such, for cer­tain tasks, machines are already far more effi­cient than we are. Even a cal­cu­lat­or eas­ily out­per­forms us when it comes to cal­cu­lat­ing large num­bers, for example, as Laurence Dev­illers points out. As for LLMs, their mod­el­ling power far exceeds our own when it comes to trans­la­tion, doc­u­ment sum­mar­isa­tion, reac­tion speed… Pre­dict­ive AI, for its part, out­per­forms us in sort­ing, ana­lys­ing vast data­sets, and detect­ing pat­terns.

World mod­els, as pro­posed by Yann LeCun, would, how­ever, rep­res­ent a fur­ther step towards gen­er­al intel­li­gence: they would poten­tially be cap­able of under­stand­ing a situ­ation in a short space of time, tak­ing into account con­text and his­tory (without los­ing pre­vi­ously acquired know­ledge). Fur­ther­more, some research­ers speak of the emer­gence of meta­cog­ni­tion in LLMs, refer­ring to the fact that some of their “beha­viours” resemble reflec­tion on their own know­ledge or reas­on­ing. These beha­viours appear without hav­ing been expli­citly programmed.

FALSE

There are two major bar­ri­ers that will lim­it arti­fi­cial intelligence’s abil­ity to match human cog­nit­ive per­form­ance: emo­tion and spir­itu­al­ity⁴. Cur­rent mod­els are incap­able of emo­tion and lack both inten­tion and con­scious­ness. These sys­tems do not exper­i­ence feel­ings; they merely imit­ate human express­ive­ness, she emphas­ises. Wheth­er it be empathy or oth­er emo­tions, they “replay some­thing con­tex­tu­ally sim­il­ar, but without any real inter­ac­tion, if only because there is no biofeedback”.

The lack of a “body” in machines is, moreover, anoth­er major obstacle. A sig­ni­fic­ant pro­por­tion of our intel­li­gence stems from our phys­ic­al inter­ac­tions with the world. Some research­ers there­fore believe that a true AGI will need to be able to inter­act in some way with an envir­on­ment, wheth­er phys­ic­al or simulated.

UNCERTAIN

There are two approaches under dis­cus­sion, the spe­cial­ist points out: either we are talk­ing about func­tion­al con­scious­ness or phe­nom­en­al con­scious­ness. The first term refers more to meta­cog­ni­tion (reas­on­ing about what the AI knows or does not know) than to sub­ject­ive exper­i­ence; in oth­er words, a sort of sim­u­lated con­scious­ness. The second term con­cerns sub­ject­ive exper­i­ence, the act of liv­ing and feel­ing some­thing. We also speak of qualia, that is, the per­ceived qual­it­ies of cer­tain men­tal states, con­scious exper­i­ences: exper­i­en­cing a sen­sa­tion, smelling a scent, see­ing a col­our… There is no evid­ence to sug­gest that AI sys­tems can acquire this type of bio­lo­gic­al con­scious­ness without a bio­lo­gic­al basis or a phys­ic­al body.

AGI remains a the­ory, a research object­ive, which faces numer­ous chal­lenges: a deep under­stand­ing of the phys­ic­al world, gen­er­al reas­on­ing, con­tinu­ous learn­ing, inter­ac­tion with the envir­on­ment, access to emotions…

Interview by Célia Chaboud