Profiling: how algorithms predict and influence our needs

When dis­cussing upcom­ing vaca­tions or pur­chas­es with friends, have you ever won­dered why high­ly tar­get­ed ads sud­den­ly appear on your Face­book wall or Insta­gram feed? We some­times feel like our elec­tron­ic devices are watch­ing us. This con­cern is not unfound­ed: the traces we leave online reveal valu­able infor­ma­tion about our lives, often with­out us being ful­ly aware of it.

Fur­ther­more, on 14^th Feb­ru­ary 2025, the Human Rights League filed a com­plaint in France against Apple for vio­la­tion of pri­va­cy through the unau­tho­rised col­lec­tion of user data via the Siri voice assis­tant. This case rais­es the issue of the pro­tec­tion of our per­son­al data, which has become a valu­able resource cov­et­ed by companies.

Take cook­ies, for exam­ple, those ele­ments that we are asked to accept before access­ing web­sites. Behind their appetis­ing name lie oppor­tu­ni­ties for com­pa­nies to access our per­son­al data. As Philippe Hune­man (CNRS research direc­tor at the Insti­tute for the His­to­ry and Phi­los­o­phy of Sci­ence and Tech­nol­o­gy) shows in his book Les sociétés du pro­fi­lage (Pro­fil­ing Soci­eties), it is impor­tant to dis­tin­guish between “nec­es­sary” cook­ies, which ensure the prop­er func­tion­ing of a web­site, and “option­al” cook­ies, which are intend­ed to improve the user’s brows­ing expe­ri­ence or per­son­alise “more rel­e­vant” adver­tise­ments for the user¹. By accept­ing these cook­ies on a par­tic­u­lar web­site, we con­sent to some of our online behav­iour being observed. This behav­iour is often linked to data bro­kers, com­pa­nies that buy, col­lect, and aggre­gate data from mul­ti­ple web­sites and ulti­mate­ly resell it. Among the best known are Acx­iom in the Unit­ed States.

But why share and sell our per­son­al data? One of the main objec­tives is to influ­ence our online pur­chas­ing behav­iour through tar­get­ed adver­tis­ing. As Oana Goga (INRIA research direc­tor at Ecole Poly­tech­nique, IP Paris) points out: “In the field of online adver­tis­ing, track­ing [Editor’s note: mon­i­tor­ing users’ online behav­iour on the web] is the basis of two tar­get­ing meth­ods: the first is retar­get­ing, a method that involves tar­get­ing Inter­net users who have already vis­it­ed a web­site by dis­play­ing adver­tise­ments on oth­er sites they vis­it. The oth­er tech­nique is pro­fil­ing-based tar­get­ing, which involves cre­at­ing a user profile.”

The traces we leave on the web in the dig­i­tal age can there­fore be col­lect­ed to build a pro­file. This prac­tice, known as “pro­fil­ing”, is defined by the GDPR as “the auto­mat­ed pro­cess­ing of per­son­al data to eval­u­ate cer­tain per­son­al aspects relat­ing to a nat­ur­al per­son²”. It is used to analyse, pre­dict or influ­ence indi­vid­u­als’ behav­iour, includ­ing through the use of algo­rithms. To illus­trate this con­cept, let’s take the exam­ple giv­en by the researcher on Facebook’s pro­fil­ing tar­get­ing and how it has evolved: “In 2018, users were clas­si­fied on Face­book into 250,000 cat­e­gories by algo­rithms, based on their pref­er­ences on the plat­form. Today, this clas­si­fi­ca­tion is no longer explic­it. Algo­rithms no longer place users in cat­e­gories so that adver­tis­ers can choose who they want to tar­get but instead decide for adver­tis­ers who to send adver­tise­ments to.”

The algo­rithms used today to pre­dict and influ­ence our actions are extreme­ly effec­tive. They are said to be bet­ter than humans at under­stand­ing our behav­iour and could even influ­ence it. For exam­ple, research shows that com­put­er mod­els are much more effec­tive and accu­rate than humans at per­form­ing an essen­tial socio-cog­ni­tive task: per­son­al­i­ty assess­ment³.

But this effec­tive­ness rais­es sev­er­al ques­tions: to what extent can these algo­rithms know and pre­dict our behav­iour? And how do they work? To date, the answers remain unclear. Oana Goga says: “One of the big prob­lems with rec­om­men­da­tion algo­rithms is that they are dif­fi­cult to audit because the data is pri­vate and belongs to com­pa­nies.” Philippe Hune­man adds: “Right now, we don’t know how algo­rithms use our data to pre­dict our behav­iour, but their mod­els are get­ting bet­ter and bet­ter. Just as with gen­er­a­tive AI, we don’t know how the data is put togeth­er. We have to choose: do we want a world where this soft­ware is effec­tive, or ethical?”

The eth­i­cal impli­ca­tions of these algo­rithms are fun­da­men­tal. In 2016, the Cam­bridge Ana­lyt­i­ca scan­dal⁴ high­light­ed the pos­si­bil­i­ty of exploit­ing user data with­out their con­sent for polit­i­cal pur­pos­es, in par­tic­u­lar by devel­op­ing soft­ware capa­ble of tar­get­ing spe­cif­ic user pro­files and influ­enc­ing their votes, as in the case of Brex­it and the elec­tion of Don­ald Trump. How­ev­er, prov­ing that these manoeu­vres actu­al­ly influ­enced the out­comes of these events remains dif­fi­cult. More recent cas­es include the can­cel­la­tion of the pres­i­den­tial elec­tion in Roma­nia by the Con­sti­tu­tion­al Court in Decem­ber 2024, fol­low­ing sus­pi­cions of an ille­gal sup­port cam­paign on Tik­Tok⁵. The algo­rithms used by plat­forms such as Face­book and X could also be more like­ly than oth­ers to rein­force echo cham­bers, i.e. lim­it­ing expo­sure to diverse per­spec­tives and encour­ag­ing the for­ma­tion of groups of like-mind­ed users, there­by rein­forc­ing cer­tain com­mon nar­ra­tives⁶.

In this con­text, Oana Goga and her team have been run­ning the MOMENTOUS pro­gramme since 2022, fund­ed by a Euro­pean Research Coun­cil (ERC) grant. Its aim is to under­stand how algo­rithms can exploit psy­cho­log­i­cal and cog­ni­tive traits to influ­ence people’s pref­er­ences and behav­iour. The pro­gramme offers a new mea­sure­ment method­ol­o­gy based on ran­domised con­trolled tri­als in social media. As Oana Goga points out: “It is impor­tant to dis­tin­guish between algo­rith­mic bias­es, such as algo­rithms that dis­crim­i­nate against cer­tain pop­u­la­tions, and cog­ni­tive bias­es, which are bias­es held by humans. With MOMENTOUS, we are look­ing at whether algo­rithms can exploit cog­ni­tive bias­es.” On a sim­i­lar note, Philippe Hune­man also men­tions the con­cept of nudge in pro­fil­ing: “Pro­fil­ing con­veys the idea of soft pater­nal­ism, or lib­er­tar­i­an­ism aimed at influ­enc­ing an individual’s behav­iour by act­ing on the bias­es that gov­ern them. Adver­tise­ments and web­site inter­faces exploit these bias­es to influ­ence users’ deci­sions; this is nudging.”

In addi­tion, among the eth­i­cal issues raised by tar­get­ing, the first to men­tion is the issue of chil­dren: “From a legal stand­point, we don’t have the right to tar­get chil­dren based on pro­fil­ing. How­ev­er, our stud­ies on YouTube have revealed that it is pos­si­ble to tar­get them con­tex­tu­al­ly, for exam­ple by dis­play­ing adver­tise­ments on Pep­pa Pig videos or on influ­encer chan­nels,” explains Oana Goga. “Although plat­forms pro­hib­it tar­get­ing chil­dren under the age of 18, they can tar­get the con­tent they watch. The prob­lem is that dig­i­tal reg­u­la­tors are focus­ing on ban­ning pro­fil­ing-based tar­get­ing of chil­dren, but not con­tex­tu­al tar­get­ing, which takes into account con­tent specif­i­cal­ly aimed at them, even though these strate­gies are well known to adver­tis­ers. There is a lack of data on adver­tis­ing that tar­gets children’s chan­nels, which rais­es the issue of the risks of rad­i­cal­i­sa­tion among young peo­ple,” adds the researcher.

How can we make things hap­pen? From a reg­u­la­to­ry per­spec­tive, Oana Goga believes that one of the most press­ing issues is ensur­ing more trans­par­ent access to online plat­form data: “Con­crete mea­sures must be tak­en to enable bet­ter access to data so that effec­tive action can be tak­en. This could be done in two ways: i) through leg­is­la­tion; and ii) through cit­i­zen par­tic­i­pa­tion. It is essen­tial to be able to col­lect data in an eth­i­cal man­ner that com­plies with the GDPR.”

With this in mind, Oana Goga has been devel­op­ing tools such as AdAn­a­lyst and Check­MyNews for Meta and YouTube for sev­er­al years. Their goal is to col­lect user data to con­duct research on the con­tent and sources of infor­ma­tion they receive on these net­works while respect­ing their pri­va­cy as much as pos­si­ble, in par­tic­u­lar by not col­lect­ing people’s emails, by going through ethics com­mit­tees, and in com­pli­ance with the GDPR. “It would also be inter­est­ing to have a pan­el of users at the Euro­pean lev­el. A plat­form obser­va­to­ry, with 1,000 to 2,000 users in France, Ger­many, etc., could pro­vide access to data inde­pen­dent­ly of the plat­forms,” she adds.

These are ques­tions at the heart of our soci­ety, which should be at the cen­tre of dis­cus­sions on democ­ra­cy in the com­ing years.

Lucille Caliman