AI and the media: a (r)evolution in investigative reporting?

Arti­fi­cial Intel­li­gence (AI) is every­where, grad­u­al­ly find­ing it’s way into the infor­ma­tion sec­tor. Could AI be the next rev­o­lu­tion in the media? The answer is more complex…

For the past ten years or so, the online press has been test­ing gen­er­a­tive AI to auto­mat­i­cal­ly write arti­cles by “suck­ing” infor­ma­tion from reli­able data­bas­es. In March 2015, before Chat­G­PT, Le Monde news­pa­per had used AI devel­oped by the com­pa­ny Syl­labs to write 36,000 arti­cles cov­er­ing the results of the region­al elec­tions in each com­mune, thanks to data from the Min­istry of the Inte­ri­or. A few months lat­er, France Bleu fol­lowed suit, this time for the region­al elec­tions. Since 2021, L’Équipe has also been using AI to auto­mate the pro­duc­tion of con­tent list­ing upcom­ing match­es, fol­lowed by the time and chan­nel of broad­cast. Are jour­nal­ists being replaced by robots?

“Arti­fi­cial intel­li­gence is first and fore­most a tool” says Ioana Manoles­cu, a com­put­er sci­ence researcher at Inria spe­cial­is­ing in large-scale data pro­cess­ing and fact-check­ing. “In edi­to­r­i­al depart­ments, there is a mix­ture of mis­trust and fas­ci­na­tion” she con­tin­ues. “While text gen­er­a­tion works well with Amer­i­can AI – which is more high­ly trained – French AI is still rather clum­sy.” What’s more, the abil­i­ty to analyse, inter­view, cross-check sources and car­ry out in-depth inves­ti­ga­tions remain exclu­sive­ly human skills for the time being. So, there is no risk of jour­nal­ists being replaced by “robot edi­tors” for day-to-day work, as some peo­ple envis­aged a decade ago.

“On the oth­er hand, AI is very good at retriev­ing, organ­is­ing or com­par­ing con­tent with an exist­ing data­base” explains Ioana Manoles­cu. This is the ori­gin of the Statcheck tool, which the researcher and her team have been devel­op­ing in col­lab­o­ra­tion with Radio France since 2021. This AI allows sta­tis­ti­cal infor­ma­tion found in an arti­cle to be cross-ref­er­enced with the INSEE data­base and, more recent­ly, the Euro­stat database.

“To do this, we had to train the AI on a lot of texts to enable it to learn, for exam­ple, that the con­cepts of “employ­ment” and “unem­ploy­ment” are relat­ed…” explains the sci­en­tist. Today, StatCheck can rec­on­cile the way jour­nal­ists write with the ter­mi­nol­o­gy used by sta­tis­ti­cians. New func­tions con­tin­ue to be added by the Inria teams, such as open­ing the sys­tem to exter­nal data­bas­es, then stan­dar­d­is­ing sources in a sin­gle for­mat. But the aim is to go even further…

It all start­ed with an inves­ti­ga­tion car­ried out by a jour­nal­ist from Le Monde a few years ago, who combed through hun­dreds of doc­u­ments to iden­ti­fy the links between sci­en­tists and indus­tri­al lob­bies… all by hand. “It’s a colos­sal job!” com­ments Ioana Manoles­cu. So, the idea was born to cre­ate a tool to auto­mate the work. Called Con­nec­tion­Lens, it is now capa­ble of inter­con­nect­ing data tak­en from a high­ly het­ero­ge­neous cor­pus of doc­u­ments (PDF, Excel, URL, etc.). “The AI extracts the infor­ma­tion con­tained in these doc­u­ments, such as names, organ­i­sa­tions, dates, e‑mails, etc., and links them togeth­er. The name of a com­pa­ny in the acknowl­edge­ments of a the­sis, for exam­ple, will be linked to the name of the author.”

But the algo­rithm is nev­er the last link in the chain! For both StatCheck and Con­nec­tion­Lens, a jour­nal­ist sys­tem­at­i­cal­ly re-checks the work and the sources. “In my opin­ion, the only thing you shouldn’t try to ask the AI to do is to think” con­cludes the researcher.

Le Monde’s Décodeurs teams have adopt­ed this approach in their work. “AI is a very good way of sift­ing through infor­ma­tion and sav­ing time, by going from 3,000 names, for exam­ple, to the 200 that inter­est us” explains Maxime Vau­dano, who coor­di­nates the Les Décodeurs inves­tiga­tive unit. The jour­nal­ist, who spe­cialis­es in open-source and col­lab­o­ra­tive inves­ti­ga­tions, admits that they are not “very organ­ised or very impres­sive at this stage” when it comes to edi­to­r­i­al work.

AI is a very good way of sift­ing through infor­ma­tion and sav­ing time.

Indeed, although sev­er­al large-scale inves­ti­ga­tions involv­ing very large data­bas­es, such as the Pana­ma Papers, have already been car­ried out in the past using more basic algo­rithms, this does not guar­an­tee the reli­a­bil­i­ty of AI today. “We keep in mind that it is an imper­fect tool! There are, of course, false pos­i­tives, which are pieces of infor­ma­tion kept by the AI when they are not rel­e­vant, but the biggest risk remains false neg­a­tives” explains Maxime Vau­dano. “When a piece of infor­ma­tion falls by the way­side even though it’s rel­e­vant to our subject.”

So, the use of AI is far from sys­tem­at­ic, and sev­er­al “clas­sic” inves­tiga­tive tech­niques are reg­u­lar­ly used in par­al­lel. “We are still involved in very long-term research, with a huge num­ber of ver­i­fi­ca­tion steps.” So, what­ev­er the ini­tial method used to obtain the infor­ma­tion, it will be checked sev­er­al times… by humans.

Is this then a case of sim­ple evo­lu­tion rather than rev­o­lu­tion? That’s cer­tain­ly the view of Ioana Manoles­cu, who remains very cau­tious about the real capa­bil­i­ties of AI, but also about those of the hypo­thet­i­cal AGI (Arti­fi­cial Gen­er­al Intel­li­gence) of the future. “Yes, com­put­ers can beat humans at chess, so in that respect it’s all over. But we’re talk­ing here about a very spe­cif­ic sys­tem! As for the rest, “intel­li­gent” robots don’t know that water is wet or that time only flows in one direc­tion, where­as a baby does…”

As for AI as an inves­tiga­tive tool, the researcher is more opti­mistic. “I tell myself that young jour­nal­ists will be able to learn these tech­niques and bring them into their news­rooms.” Because the use of AI remains linked to a cer­tain tech­no­log­i­cal cul­ture with­in the media, but also to the train­ing avail­able to jour­nal­ists to get to grips with these systems.

“We don’t have enough time to put these tools in place, even if the tech­nol­o­gy is already avail­able” adds Maxime Vau­dano. But the sit­u­a­tion is chang­ing. Since 2023, Reporters sans fron­tières has been devel­op­ing an AI pro­to­type for jour­nal­ists, ded­i­cat­ed to envi­ron­men­tal issues and news ver­i­fi­ca­tion. It has been under­go­ing tri­als since April 2024 with 12 part­ner media out­lets. At the same time, in March 2024 the news­pa­per Le Monde entered into a finan­cial part­ner­ship with Ope­nAI, the par­ent com­pa­ny of Chat­G­PT, to give their AI access to their archives. This alliance is being care­ful­ly mon­i­tored by teams of jour­nal­ists, but it her­alds an unprece­dent­ed part­ner­ship between the media and intel­li­gent technologies.

Sophie Podevin