The addictive design of social media under fire from regulators 

On 6^th Feb­ru­ary 2026, the European Com­mis­sion issued a pre­lim­in­ary opin­ion on the viol­a­tion of the Digit­al Ser­vices Act by the Chinese social net­work Tik­Tok¹. The Com­mis­sion con­siders that Byte­Dance, the com­pany behind Tik­Tok, has failed to “take reas­on­able, pro­por­tion­ate and effect­ive meas­ures to mit­ig­ate the risks arising from the addict­ive design” of its product. It refers to fea­tures such as infin­ite scrolling, auto­play and noti­fic­a­tions, but also its recom­mend­a­tion AI, i.e. the algorithm that responds to every click, swipe or scroll by a user look­ing for new con­tent. This is the first time that a demo­crat­ic insti­tu­tion has taken on an algorithm at the heart of con­trolling the flow of inform­a­tion in the mod­ern world.
 
But this leg­al attack is not an isol­ated one. At the end of Janu­ary 2026, in the pro­vi­sion­al ver­sion of the draft bill aimed at pro­tect­ing minors from the risks to which they are exposed through the use of social net­works, Art­icle 1 bis was intro­duced, which also spe­cific­ally tar­gets recom­mend­a­tion AI. “For the inform­a­tion thus high­lighted, the pro­vider may be held liable as a pub­lish­er,” states para­graph 6–8‑1-II. At the same time, in Cali­for­nia on 9^th Feb­ru­ary 2026, a law­suit accused Meta and You­Tube of “man­u­fac­tur­ing addic­tion”, par­tic­u­larly through recom­mend­a­tion AI, thus going far bey­ond the simple host­ing of con­tent, which is gov­erned by Sec­tion 230, which is very leni­ent towards host­ing pro­viders². In two weeks, recom­mend­a­tion AI has been the sub­ject of more leg­al scru­tiny than in its pre­vi­ous two dec­ades of exist­ence.
 
Yet for at least a dec­ade, sci­ent­ists, journ­al­ists and human rights advoc­ates have been warn­ing about the civil­isa­tion­al risk posed by these recom­mend­a­tion AI sys­tems, some­times called cur­a­tion sys­tems, or some­times just “the algorithm”, par­tic­u­larly by con­tent cre­at­ors and con­sumers on social media. The accus­a­tions range from exploit­ing users’ psy­cho­lo­gic­al vul­ner­ab­il­it­ies to incit­ing sui­cide, cog­nit­ive impair­ment, com­pli­city in cyber­bul­ly­ing, amp­li­fic­a­tion of mass dis­in­form­a­tion, destabil­isa­tion of demo­crat­ic elec­tions and know­ingly con­trib­ut­ing to gen­o­cide. In this art­icle, we will com­pile some of the most damning sci­entif­ic argu­ments against these AIs.

On 27^th Novem­ber 2025, Sci­ence pub­lished an art­icle on the monu­ment­al impact of con­tent recom­mend­a­tions on users³. The study invited a group of X/Twitter users to install a browser exten­sion. This exten­sion then assigned users to dif­fer­ent groups. For some, no inter­ven­tion took place. For oth­ers, the exten­sion would reorder the mes­sages pro­posed by X/Twitter’s recom­mend­a­tion AI, put­ting those that were most angry or hate­ful towards the oppos­ing group first. For oth­ers, the exten­sion would pri­or­it­ise the most calm or polit­ic­ally bene­vol­ent con­tent. After only one week of inter­ven­tion, strik­ingly, this min­im­al inter­ven­tion is already affect­ing users’ emo­tion­al polar­isa­tion, i.e. their degree of emo­tion­al anti­pathy towards the oppos­ing clan, to an extent that the authors estim­ate to be “com­par­able to three years of atti­tude change in the United States”.

In a very insi­di­ous way, even though users knew they were par­ti­cip­at­ing in an exper­i­ment (they had to install an exten­sion!), they were unable to guess the nature of the inter­ven­tion they were the tar­get of. The reorder­ing of pub­lished mes­sages was, in a sense, sub­lim­in­al, or at least below their threshold of con­scious­ness. This echoes para­graph 5.1.a of the AI Act, which came into force in the European Uni­on and pro­hib­its AI sys­tems that “use sub­lim­in­al tech­niques […] with the aim or effect of sub­stan­tially alter­ing a person’s beha­viour”. Of course, oth­er con­di­tions are men­tioned in this para­graph, and the study in Sci­ence is cer­tainly not enough to con­clude that recom­mend­a­tion AI should be banned. Nev­er­the­less, it raises dizzy­ing ques­tions about the invest­ig­a­tions that should be pri­or­it­ised to pro­tect their 3 bil­lion users, and about the applic­a­tion of exist­ing reg­u­la­tions.
 
Moreover, the Sci­ence study is not an isol­ated case. Elev­en years earli­er, on 2^nd June 2014, a study was pub­lished in PNAS that already showed this monu­ment­al impact of AI⁴. The inter­ven­tion was also min­im­al. It involved remov­ing pos­it­ive or neg­at­ive emo­tion­al con­tent from the recom­mend­a­tion feed, with a prob­ab­il­ity of between 10% and 90%. Amaz­ingly, after just one week, simply remov­ing a frac­tion of one type of emo­tion­al con­tent changed users’ beha­viour in line with that remov­al. In oth­er words, the less neg­at­ive con­tent users saw, the less neg­at­ive the con­tent they pos­ted. Equally strik­ing was the fact that redu­cing emo­tion­al con­tent, wheth­er pos­it­ive or neg­at­ive, reduced user activ­ity by approx­im­ately 5 stand­ard devi­ations.
 
There is no doubt that this last res­ult is the sub­ject of recur­ring obser­va­tions by all social net­works, giv­en that user engage­ment is dir­ectly linked to the abil­ity of plat­forms to show them highly tar­geted advert­ise­ments. We are talk­ing here about a mar­ket worth hun­dreds of bil­lions of dol­lars a year, ten times lar­ger than that of chat­bots today. Des­pite ques­tion­able cir­cu­lar fin­an­cing, OpenAI is fore­cast­ing “only” £12 bil­lion in rev­en­ue in 2025. This is a drop in the ocean com­pared to Meta’s £201 bil­lion. Giv­en the amounts at stake, it is not sur­pris­ing to see recom­mend­a­tion AI being over-optim­ised to cap­ture users. In doing so, we can expect them to massively recom­mend extremely emo­tion­ally charged con­tent, both pos­it­ive and neg­at­ive, in order to make social media as addict­ive as possible.

While the addict­ive design of recom­mend­a­tion AI is now the sub­ject of leg­al pro­ceed­ings by the European Com­mis­sion and in Cali­for­nia, incrim­in­at­ing evid­ence is mount­ing regard­ing its cata­stroph­ic role in user well-being, the safety of eth­nic pop­u­la­tions and the integ­rity of demo­cra­cies. Examples include a Brit­ish sur­vey reveal­ing that nearly half of young people say they would have pre­ferred to grow up without the Inter­net⁵, the accus­a­tion by the UN and Amnesty Inter­na­tion­al that Face­book know­ingly con­trib­uted to the gen­o­cide of the Rohingya by massively amp­li­fy­ing calls for murder⁶, and the can­cel­la­tion of the 2024 pres­id­en­tial elec­tions in Romania fol­low­ing inform­a­tion inter­fer­ence on TikTok.

Incrim­in­at­ing evid­ence is mount­ing regard­ing its cata­stroph­ic role in user well-being, the safety of eth­nic pop­u­la­tions and the integ­rity of democracies. 

Bey­ond the strik­ing cor­rel­a­tion between the wide­spread adop­tion of social media around 2012 and the wor­ry­ing trend in indic­at­ors such anxi­ety and depres­sion rate among young people⁷, the num­ber of deaths in wars⁸, and the lib­er­al demo­cracy index⁹, there are a num­ber of wor­ry­ing argu­ments regard­ing the vari­ous neg­at­ive impacts of social media, such as the testi­mon­ies of vic­tims and wit­nesses, epi­demi­olo­gic­al stud­ies¹⁰, ran­dom­ised clin­ic­al tri­als¹¹, nat­ur­al exper­i­ments¹²and leaks of intern­al doc­u­ments.
 
In par­tic­u­lar, as in pre­vi­ous cases of mer­chants of doubt (tobacco, oil, sug­ar, pesti­cides, PFAS, etc.), these intern­al doc­u­ments reveal that man­u­fac­tur­ers con­cluded that their products were harm­ful, some­times dec­ades before research­ers came to the same con­clu­sion. The Face­book Files in par­tic­u­lar, exposed by whis­tleblower Frances Hau­gen in 2021, are extremely damning in terms of the cata­stroph­ic implic­a­tions of recom­mend­a­tion AI for the men­tal health of young girls, the qual­ity of journ­al­ism and polit­ic­al dis­course, push­ing the lat­ter in par­tic­u­lar to sys­tem­at­ic­ally adopt divis­ive tones.
 
How­ever, unlike oth­er indus­tries, digit­al giants have the advant­age of enorm­ous influ­ence over aca­dem­ic research¹³and a mono­poly on com­prom­ising data, which they act­ively seek to pro­tect even with­in a leg­al frame­work. In this con­text, any sci­ent­ist who claims “it has not been proven” without spe­cify­ing that the lack of evid­ence is fuelled by the delib­er­ate opa­city of the accused seems to me to be offer­ing a prob­lem­at­ic, and per­haps even irre­spons­ible, inter­pret­a­tion of our cur­rent under­stand­ing of the com­plex rela­tion­ship between sci­ence and industry.
 
On the con­trary, giv­en the moun­tain of incrim­in­at­ing evid­ence against the digit­al giants and their recom­mend­a­tion AI, and know­ing, moreover, their dis­turb­ing prox­im­ity to Chinese author­it­ari­an­ism or to a US power that Pres­id­ent Mac­ron has described as neo-colo­ni­al, it seems wise to invest heav­ily in the design of sov­er­eign altern­at­ives that com­ply with demo­crat­ic stand­ards, as pro­posed by pro­jects such as EU OS, FramaSoft, Mas­to­don, EuroSky and Tournesol, a non-profit pro­ject spe­cial­ising in the design of recom­mend­a­tion AI for the pub­lic good. It also seems wise to adopt a rad­ic­ally more accus­at­ory tone, as the European Com­mis­sion has done with TikTok.