Generative AI: the risk of cognitive atrophy

Less than three years after the launch of Chat­G­PT, 42% of young French peo­ple already use gen­er­a­tive AI on a dai­ly basis¹. How­ev­er, stud­ies are begin­ning to point to the neg­a­tive impact of these tech­nolo­gies on our cog­ni­tive abil­i­ties. Ioan Rox­in, pro­fes­sor emer­i­tus at Marie et Louis Pas­teur Uni­ver­si­ty and spe­cial­ist in infor­ma­tion tech­nol­o­gy, answers our questions.

Ioan Rox­in. The wide­spread use of the inter­net and social media has already weak­ened our rela­tion­ship with knowl­edge. Of course, these tools have tremen­dous appli­ca­tions in terms of access to infor­ma­tion. But con­trary to what they claim, they are less about democ­ra­tis­ing knowl­edge than cre­at­ing a gen­er­alised illu­sion of knowl­edge. I don’t think it’s an exag­ger­a­tion to say that they are dri­ving intel­lec­tu­al, emo­tion­al and moral medi­oc­rity on a glob­al scale. Intel­lec­tu­al because they encour­age over­con­sump­tion of con­tent with­out any real crit­i­cal analy­sis; emo­tion­al because they cre­ate an ever-deep­en­ing depen­dence on stim­u­la­tion and enter­tain­ment; and moral because we have fall­en into pas­sive accep­tance of algo­rith­mic decisions.

Yes. Back in 2011, a study high­light­ed the “Google effect”: when we know that infor­ma­tion is avail­able online, we do not remem­ber it as well. How­ev­er, when we no longer train our mem­o­ry, the asso­ci­at­ed neur­al net­works atro­phy. It has also been proven that the inces­sant noti­fi­ca­tions, alerts and con­tent sug­ges­tions on which dig­i­tal tech­nolo­gies rely heav­i­ly sig­nif­i­cant­ly reduce our abil­i­ty to con­cen­trate and think. Reduced mem­o­ry, con­cen­tra­tion and ana­lyt­i­cal skills lead to dimin­ished cog­ni­tive process­es. I very much fear that the wide­spread use of gen­er­a­tive AI will not improve the situation.

There are neu­ro­log­i­cal, psy­cho­log­i­cal and philo­soph­i­cal risks. From a neu­ro­log­i­cal stand­point, wide­spread use of this AI car­ries the risk of over­all cog­ni­tive atro­phy and loss of brain plas­tic­i­ty. For exam­ple, researchers at the Mass­a­chu­setts Insti­tute of Tech­nol­o­gy (MIT) con­duct­ed a four-month study² involv­ing 54 par­tic­i­pants who were asked to write essays with­out assis­tance, with access to the inter­net via a search engine or with Chat­G­PT. Their neur­al activ­i­ty was mon­i­tored by EEG. The study, the results of which are still in preprint, found that using the inter­net, and even more so Chat­G­PT, sig­nif­i­cant­ly reduced cog­ni­tive engage­ment and “rel­e­vant cog­ni­tive load”, i.e. the intel­lec­tu­al effort required to trans­form infor­ma­tion into knowledge. 

More specif­i­cal­ly, par­tic­i­pants assist­ed by Chat­G­PT wrote 60% faster, but their rel­e­vant cog­ni­tive load fell by 32%. EEG showed that brain con­nec­tiv­i­ty was almost halved (alpha and theta waves) and 83% of AI users were unable to remem­ber a pas­sage they had just written. 

Oth­er stud­ies sug­gest a sim­i­lar trend: research³ con­duct­ed by Qatari, Tunisian and Ital­ian researchers indi­cates that heavy use of LLM car­ries the risk of cog­ni­tive decline. The neur­al net­works involved in struc­tur­ing thought, writ­ing texts, but also in trans­la­tion, cre­ative pro­duc­tion, etc. are com­plex and deep. Del­e­gat­ing men­tal effort to AI leads to a cumu­la­tive “cog­ni­tive debt”: the more automa­tion pro­gress­es, the less the pre­frontal cor­tex is used, sug­gest­ing last­ing effects beyond the imme­di­ate task.

Gen­er­a­tive AI have every­thing it takes to make us depen­dant on it: it express­es itself like humans, adapts to our behav­iour, seems to have all the answers, is fun to inter­act with, always keeps the con­ver­sa­tion going and is extreme­ly accom­mo­dat­ing towards us. How­ev­er, this depen­dence is harm­ful not only because it increas­es oth­er risks but in and of itself. It can lead to social iso­la­tion, reflex­ive dis­en­gage­ment (“if AI can answer all my ques­tions, why do I need to learn or think for myself?”) and even a deep sense of humil­i­a­tion when faced with this tool’s incred­i­ble effi­ca­cy. None of this gives a par­tic­u­lar­ly opti­mistic out­look for our men­tal health.

Gen­er­alised cog­ni­tive atro­phy is already a philo­soph­i­cal risk in itself… but there are oth­ers. If this type of tool is wide­ly used – and this is already the case with younger gen­er­a­tions – we are at risk of a stan­dard­i­s­a­tion of thought. Research⁴ car­ried out by British researchers showed that when authors asked Chat­G­PT to improve their work, the indi­vid­ual ben­e­fits could be great, but the over­all cre­ativ­i­ty of the group reduced. Anoth­er risk relates to our crit­i­cal thinking. 

One study⁵ car­ried out by Microsoft on 319 knowl­edge work­ers showed a sig­nif­i­cant neg­a­tive cor­re­la­tion (r=-0.49) between the fre­quen­cy with which AI tools were used and crit­i­cal think­ing scores (Bloom’s tax­on­o­my). The study con­clud­ed that is an increased ten­den­cy to offload men­tal effort as trust in the sys­tem exceeds trust in our own abil­i­ties. How­ev­er, it is essen­tial to main­tain a crit­i­cal mind­set as AI can not only make mis­takes or per­pet­u­ate bias­es but also con­ceal infor­ma­tion or sim­u­late compliance.

A vast major­i­ty are sole­ly con­nec­tion­ist AI, which rely on arti­fi­cial neur­al net­works trained using phe­nom­e­nal amounts of data. They learn to gen­er­ate plau­si­ble answers to all our ques­tions through sta­tis­ti­cal and prob­a­bilis­tic pro­cess­ing. Their per­for­mance has improved con­sid­er­ably with the intro­duc­tion of Google’s “Trans­former” tech­nol­o­gy in 2017. Thanks to this tech­nol­o­gy, AI can analyse all the words in a text in par­al­lel and weigh their impor­tance for mean­ing, which allows for greater sub­tle­ty in responses. 

But the back­ground remains prob­a­bilis­tic: while their answers always seem con­vinc­ing and log­i­cal, they can be com­plete­ly wrong. In 2023, users had fun ask­ing Chat­G­PT about cow eggs: the AI dis­cussed the ques­tion at length with­out ever answer­ing that they did not exist. This error has since been cor­rect­ed through rein­force­ment learn­ing with human feed­back, but it illus­trates well how these tools work.

Some com­pa­nies are start­ing to com­bine con­nec­tion­ist AI, which learns every­thing from scratch, with old­er tech­nol­o­gy, sym­bol­ic AI, in which rules to fol­low and basic knowl­edge are explic­it­ly pro­grammed. It seems to me that the future lies in neu­ro-sym­bol­ic AI. This hybridi­s­a­tion not only improves the reli­a­bil­i­ty of respons­es but also reduces the ener­gy and finan­cial costs of training.

Yes. There are two types. The first can be delib­er­ate­ly intro­duced by the AI cre­ator. LLMs are trained on all kinds of unfil­tered con­tent avail­able online (an esti­mat­ed 4 tril­lion words for ChatGPT4, com­pared to the 5 bil­lion words con­tained in the Eng­lish ver­sion of Wikipedia!). Pre-train­ing cre­ates a “mon­ster” that can gen­er­ate all kinds of horrors.

A sec­ond step (called super­vised fine-tun­ing) is there­fore nec­es­sary: it con­fronts the pre-trained AI with val­i­dat­ed data, which serves as a ref­er­ence. This oper­a­tion enables, for exam­ple, “teach­ing” it to avoid dis­crim­i­na­tion, but can also be used to guide its respons­es for ide­o­log­i­cal pur­pos­es. A few weeks after its launch, DeepSeek made head­lines for its eva­sive respons­es to user ques­tions about Tianan­men Square and Tai­wanese inde­pen­dence. It is impor­tant to remem­ber that con­tent gen­er­a­tors of this type may not be neu­tral. Blind­ly trust­ing them can lead to the spread of ide­o­log­i­cal­ly biased theories.

These bias­es appear spon­ta­neous­ly, often with­out a clear expla­na­tion. Lan­guage mod­els (LLMs) have “emer­gent” prop­er­ties that were not antic­i­pat­ed by their design­ers. Some are remark­able: these text gen­er­a­tors write flaw­less­ly and have become excel­lent trans­la­tors with­out any gram­mar rules being cod­ed in. But oth­ers are cause for con­cern. The MASK⁶ bench­mark (Mod­el Align­ment between State­ments and Knowl­edge), pub­lished in March 2025, shows that, among the thir­ty mod­els test­ed, none achieved more than 46% hon­esty, and that the propen­si­ty to lie increas­es with the size of the mod­el, even if their fac­tu­al accu­ra­cy improves. 

It seems to me that the future lies in neu­ro-sym­bol­ic AI

MASK proves that LLMs “know how to lie” when con­flict­ing objec­tives (e.g., charm­ing a jour­nal­ist, respond­ing to com­mer­cial or hier­ar­chi­cal pres­sures) pre­dom­i­nate. In some tests, AI delib­er­ate­ly lied⁷, threat­ened users⁸, cir­cum­vent­ed eth­i­cal super­vi­sion rules⁹ and even repro­duced autonomous­ly to ensure its sur­vival¹⁰.

These behav­iours, for which the deci­sion-mak­ing mech­a­nisms remain opaque, are not pos­si­ble to pre­cise­ly con­trol. These capa­bil­i­ties emerge from the train­ing process itself: it is a form of algo­rith­mic self-organ­i­sa­tion, not a design flaw. Gen­er­a­tive AI devel­ops rather than being designed, with its inter­nal log­ic form­ing in a self-organ­ised man­ner, with­out a blue­print. These devel­op­ments are suf­fi­cient­ly wor­ry­ing that lead­ing fig­ures such as Dario Amod­ei¹¹ (CEO of Anthrop­ic), Yoshua Ben­gio¹² (founder of Mila), Sam Alt­man (cre­ator of Chat­G­PT) and Geof­frey Hin­ton (win­ner of the Nobel Prize in Physics in 2024), are call­ing for strict reg­u­la­tion to favour AI that is more trans­par­ent, eth­i­cal and aligned with human val­ues, includ­ing a slow­down in the devel­op­ment of these technologies.

No. The flu­id­i­ty of their con­ver­sa­tion and these emerg­ing prop­er­ties can give the illu­sion of intel­li­gence at work.

But no AI is intel­li­gent in the human sense of the word. They have no con­scious­ness or will, and do not real­ly under­stand the con­tent they are han­dling. Their func­tion­ing is pure­ly sta­tis­ti­cal and prob­a­bilis­tic, and these devi­a­tions only emerge because they seek to respond to ini­tial com­mands. It is not so much their self-aware­ness as the opac­i­ty of their func­tion­ing that wor­ries researchers.

Yes, but this requires both active­ly engag­ing our crit­i­cal think­ing and con­tin­u­ing to exer­cise our neur­al path­ways. AI can be a tremen­dous lever for intel­li­gence and cre­ativ­i­ty, but only if we remain capa­ble of think­ing, writ­ing and cre­at­ing with­out it.

By apply­ing a sys­tem­at­ic rule: always ques­tion the answers giv­en by text gen­er­a­tors and make a con­scious effort to think care­ful­ly about what we read, hear or believe. We must also accept that real­i­ty is com­plex and can­not be under­stood with a few super­fi­cial pieces of knowl­edge… But the best advice is undoubt­ed­ly to get into the habit of com­par­ing your point of view and knowl­edge with those of oth­er peo­ple, prefer­ably those who are knowl­edge­able. This remains the best way to devel­op your thinking.

Interview by Anne Orliac