Behind every algorithm: the unseen workers powering AI

Ant­o­nio Casilli. First, an import­ant cla­ri­fic­a­tion: this work is not “invis­ible”. If one stud­ies this phe­nomen­on as I do, by meet­ing work­ers in coun­tries such as Mad­a­gas­car, Kenya or Brazil, one dis­cov­ers that their work is hid­den but not invis­ible. So, the defin­i­tion I can give is that of work involving the pre­par­a­tion, veri­fic­a­tion and some­times imit­a­tion of arti­fi­cial intel­li­gence. That is to say, everything to do with train­ing, qual­ity con­trol and mod­el alignment.

This work is essen­tial for the exist­ence of cur­rent arti­fi­cial intel­li­gence mod­els, as they are largely based on machine learn­ing sys­tems. And “learn­ing” means that people have to teach the machine. And des­pite what one might think, this train­ing is not only provided by experts paid hun­dreds of thou­sands of euros a year. It is also, and above all, people on very low wages and in rather pre­cari­ous situ­ations who keep the machines run­ning. These hid­den work­ers have been there for about twenty years, ever since we star­ted devel­op­ing data-driv­en and machine learn­ing-based arti­fi­cial intelligence.

Since we began our research in 2016 as part of the DiPLab, we have indeed observed changes. And I can tell you that these changes are not where one might have expec­ted them. We would have expec­ted them to be more evid­ent in terms of task com­plex­ity, giv­en that we are deal­ing with mod­els that are becom­ing increas­ingly power­ful and large in scale. When you have to man­age up to a tril­lion para­met­ers, you’d ima­gine the mod­el is so com­plex that the people tasked with train­ing and veri­fy­ing it must also be top spe­cial­ists – not just “data work­ers”, but genu­ine “AI tutors”. The type of are pro­files that are fre­quently advert­ised on plat­forms like Linked­In. But this is not actu­ally the case.

Cur­rently, in the post-Chat­G­PT era (i.e. post-2022), a few spe­cial­ists have cer­tainly been recruited, but they are very much in the minor­ity. Their vis­ib­il­ity is merely a com­mu­nic­a­tion strategy employed by the tech giants and arti­fi­cial intel­li­gence pro­du­cers, aimed at their investors. Most of this work­force still con­sists of the people we encounter in our invest­ig­a­tions, who are poorly paid and carry out frag­men­ted work. What is some­times referred to as “piece­work”.

Many of them suf­fer from severe post-trau­mat­ic stress dis­order, and we see this quite con­sist­ently. Some people can even be left broken.

Since 2019–2020, we have observed that these work­ers are increas­ingly young­er. Inter­na­tion­ally, even in Europe, these work­ers are now between 18 and 35 years old, although some may be as old as 44, or even approach­ing retire­ment age. Many of these work­ers are young people, with a dis­pro­por­tion­ately high level of edu­ca­tion rel­at­ive to the tasks they per­form, their pay, and the nature of the work.

So, these people are over­qual­i­fied and over-edu­cated for the work they do. The AI mar­ket, in this respect, is a labour mar­ket that does not func­tion prop­erly, because it is not effect­ively alloc­at­ing the best work­ers, the young­est or the most highly qual­i­fied, to the jobs that pay the most. Moreover, these data work­ers are not even offered per­man­ent pos­i­tions. This work­force, recruited via plat­forms, is paid on a piece­work basis, without a prop­er employ­ment con­tract. Very often, these are simply stand­ard terms and con­di­tions of use, which must be accep­ted online, and which do not guar­an­tee work­ers’ rights. If they do have a con­tract, it is a fixed-term con­tract last­ing between one and three months. Moreover, they often reside in coun­tries where work­er pro­tec­tion is so weak that such a con­tract remains a pipe dream.

Data work­ers are a group that is highly exposed to psychoso­cial risks. A sig­ni­fic­ant pro­por­tion of them exper­i­ence isol­a­tion, for example when work­ing from home. Oth­ers work in highly restrict­ive envir­on­ments, which leave them vul­ner­able. To give you an example, con­tent and chat­bot mod­er­at­ors, who are also data work­ers, work in secure offices where they are sub­ject to very close super­vi­sion yet are unable to com­mu­nic­ate with one anoth­er. Para­dox­ic­ally, this cre­ates a situ­ation of both intense pres­sure and severe isol­a­tion at the same time.

Many of them suf­fer from severe post-trau­mat­ic stress dis­order, and we see this quite con­sist­ently. Some people can even be left broken. I remem­ber tak­ing the metro in a European city with someone, a former mod­er­at­or, who had spent months watch­ing videos of people throw­ing them­selves under trains. This per­son was so deeply trau­mat­ised that they had to stand far away from the edge of the plat­form, avoid the trains, unable to con­trol them­selves or bear being near the metro.

Yes, there are def­in­itely coun­ter­trends, which are not a recent phe­nomen­on, nor are they spon­tan­eous, in the sense that it is not the sys­tem that is self-cor­rect­ing or sta­bil­ising itself. These counter-trends are linked to social con­flicts, dis­putes, and trade uni­on struggles in cer­tain sec­tors, and to the fact that some man­age to attract the atten­tion of the pub­lic and decision-makers. This is the case with con­tent mod­er­at­ors, whose cause is now widely known.

I can give you a con­crete example from Kenya, where our col­leagues and the trade uni­on­ists we work with on the ground are extremely well organ­ised, par­tic­u­larly because Kenya is a coun­try with a strong trade uni­on tra­di­tion and a rel­at­ively pro­gress­ive con­sti­tu­tion. There is both a move­ment and a highly developed civil soci­ety. And although these indi­vidu­als are some­times migrants them­selves from oth­er coun­tries, such as Somalia, Niger­ia, Ethiopia, or South Africa, they are embed­ded in wider net­works, and these net­works are giv­ing rise to a pro­lif­er­a­tion of trade uni­ons, asso­ci­ations and oth­er alli­ances. The name may change, the label may change, but there are a huge num­ber of ini­ti­at­ives of this kind at present.

In Europe, oth­er move­ments are also tak­ing shape. I am think­ing, for example, of Ger­many, where a fem­in­ist trade uni­on called Superrr is work­ing to mobil­ise pub­lic opin­ion and decision-makers. In the United States, too, “cross-col­lar” move­ments (i.e. cross-sec­tor­al uni­ons bring­ing togeth­er white-col­lar and blue-col­lar work­ers, such as the Alpha­bet Work­ers Uni­on) are organ­ising, even though indus­tri­al­ists are very power­ful there.

Firstly, the scope of AI is enorm­ous, as this is a dual-use tech­no­logy: it has both civil­ian and mil­it­ary applic­a­tions, which can obvi­ously be used sim­ul­tan­eously. This is, in fact, par for the course, since the struc­ture of fund­ing and forms of research sup­port have incor­por­ated the mil­it­ary sphere from the out­set. This extends even into the sov­er­eign sphere, as we are deal­ing with issues of sov­er­eignty in the strictly nation­al­ist sense of the term. It is a ques­tion of a country’s inde­pend­ence and resi­li­ence, of hav­ing its own infra­struc­ture, but also of pro­ject­ing an aggress­ive and, alas, war­mon­ger­ing image and ideology.

Today, the situ­ation has become far more com­plex, par­tic­u­larly with the over­lap and inter­sec­tion between arti­fi­cial intel­li­gence train­ing and mil­it­ary object­ives. Some data work­ers thus find them­selves work­ing in their own coun­tries on tools that could sub­sequently be deployed by for­eign gov­ern­ments against them. Very spe­cif­ic cases have been doc­u­mented in Syr­ia, Palestine, and sev­er­al Afric­an countries.

We have also observed a sort of con­tinuum between human­it­ari­an ini­ti­at­ives and mil­it­ary implic­a­tions in act­ive theatres of war. The activ­it­ies of major AI play­ers have been doc­u­mented across numer­ous theatres of oper­a­tion, via facial recog­ni­tion applic­a­tions, for example. In Ukraine, many arti­fi­cial intel­li­gence giants have been iden­ti­fied as sup­pli­ers of tools and data. I am refer­ring here to Google, Palantir Tech­no­lo­gies, Clear­view AI, Microsoft, SpaceX, Anthrop­ic and OpenAI. The activ­it­ies of these com­pan­ies are prob­lem­at­ic, in that they com­bine labour-intens­ive AI pro­duc­tion with applic­a­tions that are increas­ingly being used for mil­it­ary purposes.

Per­son­ally, as a con­scien­tious object­or, I am out­raged by this situ­ation. As an aca­dem­ic too, espe­cially as it makes our work even more dif­fi­cult. Because in a geo­pol­it­ic­al con­text char­ac­ter­ised by ten­sions, not only with adversar­ies but also with his­tor­ic­al allies, the situ­ation becomes par­tic­u­larly com­plic­ated for us researchers.

In this con­text, car­ry­ing out the basic work of our research becomes com­plex. The simple act of cor­rectly doc­u­ment­ing the nature of mil­it­ary inter­ven­tion and the mis­use of tech­no­logy by polit­ic­al and eco­nom­ic act­ors whose interests align with armed con­flicts becomes a sens­it­ive, even dan­ger­ous, undertaking.

Interview by Marie Varasson