Les biais ne sont pas des accidents isolés

Les biais ne sont pas des accidents isolés

0:00

/279

1×

Contrairement aux idees recues, les biais au sein des grands modeles de langage ne sont pas des erreurs isolees ou des accidents de parcours. Ils sont profondement ancres dans la structure meme de ces outils, decoulant directement des donnees massives utilisees pour leur entrainement. Lorsqu'un outil d'intelligence artificielle est utilise dans les ressources humaines, son apparence de neutralite et la fluidite de ses reponses peuvent facilement masquer ces inequites sous-jacentes.

Ces biais se manifestent de maniere subtile, souvent par une qualite de reponse inegale selon le profil, la langue ou le contexte culturel. Les modeles reproduisent les stereotypes et les desequilibres de representation presents sur le web. De plus, les consignes de conception et l'utilisation finale de ces outils dans des processus decisionnels transforment ces biais statistiques en impacts bien reels sur des carrieres humaines.

Key takeaways

• Les biais des LLM ne sont pas des erreurs accidentelles mais des caracteristiques structurelles issues de leurs donnees d'entrainement
• L'apparence fluide et propre d'un texte genere par IA peut masquer des points de vue biaises et une absence de neutralite
• Les inequilites de representation dans les donnees sources entrainent des variations de qualite et de precision selon les profils traites
• La domination de certaines langues sur le web cree d'importantes disparites de performance et de fiabilite culturelle
• L'integration des LLM dans les processus de decision RH necessite une vigilance constante pour eviter des discriminations concretes

Highlights

• 0:12 — Les biais sont structurels et discrets
• 0:58 — Comment un grand modele de langage apprend
• 2:23 — La qualite inegale des reponses comme forme de biais
• 2:55 — L'impact de la domination linguistique sur le web
• 4:15 — La transformation des biais en decisions reelles

00:00:00 L'illusion de neutralite des outils RH
00:00:58 Les sources d'apprentissage des LLM
00:01:51 L'impact des biais dans le recrutement
00:02:55 L'inegalite des performances selon les langues
00:03:28 Les choix de conception et consignes
00:04:03 L'importance d'une vigilance active

WEBVTT

00:00:06.410 --> 00:00:12.370
Il<00:00:06.860> y<00:00:08.300> a<00:00:08.360> un<00:00:08.500> sujet<00:00:08.680> dont<00:00:08.820> on<00:00:09.020> parle<00:00:09.180> peu:<00:00:10.100> les<00:00:10.200> biais<00:00:10.380> dans<00:00:10.600> l'IA<00:00:10.800> ne<00:00:10.980> sont<00:00:11.260> pas<00:00:11.420> des<00:00:11.580> ratés<00:00:11.820> visibles.

00:00:12.560 --> 00:00:16.140
Ils<00:00:12.620> sont<00:00:12.880> structurels,<00:00:14.260> discrets,<00:00:14.940> déjà<00:00:15.220> là.

00:00:16.140 --> 00:00:23.580
On<00:00:16.280> va<00:00:16.480> remonter<00:00:16.740> jusqu'aux<00:00:17.100> fondations<00:00:17.580> des<00:00:17.800> LLM<00:00:18.200> pour<00:00:18.600> comprendre<00:00:18.960> comment<00:00:19.200> un<00:00:19.380> outil<00:00:19.580> RH<00:00:19.980> en<00:00:20.240> apparence<00:00:20.680> neutre<00:00:21.100> peut<00:00:21.500> produire<00:00:21.820> des<00:00:22.060> écarts<00:00:22.320> bien<00:00:22.680> réels.

00:00:24.140 --> 00:00:26.060
Imaginons<00:00:24.240> ensemble<00:00:24.600> un<00:00:24.740> outil<00:00:24.980> RH<00:00:25.280> basé<00:00:25.540> sur<00:00:25.740> l'IA.

00:00:26.396 --> 00:00:32.049
Vous<00:00:26.476> lui<00:00:26.573> envoyez<00:00:26.878> des<00:00:26.990> candidatures,<00:00:27.938> il<00:00:28.082> vous<00:00:28.259> aide<00:00:28.419> à<00:00:28.484> les<00:00:28.596> lire,<00:00:29.383> à<00:00:29.447> reformuler<00:00:30.009> des<00:00:30.186> appréciations,<00:00:31.069> à<00:00:31.133> proposer<00:00:31.535> des<00:00:31.615> profils.

00:00:32.514 --> 00:00:34.875
Il<00:00:32.659> répond<00:00:32.900> vite<00:00:33.317> avec<00:00:33.558> un<00:00:33.622> texte<00:00:34.024> clair,<00:00:34.425> bien<00:00:34.618> tourné.

00:00:35.405 --> 00:00:37.107
À<00:00:35.469> première<00:00:35.822> vue,<00:00:36.433> tout<00:00:36.529> a<00:00:36.609> l'air<00:00:36.834> neutre.

00:00:37.492 --> 00:00:42.984
Il<00:00:37.557> ne<00:00:37.637> crie<00:00:37.878> pas,<00:00:38.295> il<00:00:38.440> ne<00:00:38.472> met<00:00:38.681> pas<00:00:38.761> de<00:00:38.922> gros<00:00:39.162> scores<00:00:39.403> visibles,<00:00:40.222> il<00:00:40.287> ne<00:00:40.447> dit<00:00:40.608> jamais<00:00:40.849> frontalement,<00:00:41.812> prends<00:00:42.005> cette<00:00:42.214> personne,<00:00:42.615> pas<00:00:42.727> celle-là.

00:00:43.514 --> 00:00:45.473
Il<00:00:43.659> produit<00:00:43.980> juste<00:00:44.863> du<00:00:45.104> texte.

00:00:46.244 --> 00:00:47.722
Et<00:00:46.308> c'est<00:00:46.469> là<00:00:46.710> que<00:00:46.790> ça<00:00:46.951> peut<00:00:47.031> devenir<00:00:47.352> trompeur.

00:00:48.316 --> 00:00:53.487
Parce<00:00:48.476> qu'un<00:00:48.717> texte<00:00:49.038> qui<00:00:49.199> a<00:00:49.215> l'air<00:00:49.440> propre,<00:00:50.082> fluide,<00:00:50.724> raisonnable<00:00:51.688> peut<00:00:51.849> donner<00:00:52.025> l'impression<00:00:52.491> que<00:00:52.732> tout<00:00:52.908> est<00:00:52.973> sous<00:00:53.133> contrôle.

00:00:53.760 --> 00:00:55.810
Que<00:00:53.940> l'outil<00:00:54.280> n'a<00:00:54.480> pas<00:00:54.640> vraiment<00:00:54.940> de<00:00:55.120> point<00:00:55.220> de<00:00:55.360> vue.

00:00:55.940 --> 00:00:58.060
Mais<00:00:56.000> ce<00:00:56.100> qui<00:00:56.260> se<00:00:56.500> joue<00:00:56.580> en<00:00:56.720> dessous<00:00:57.120> est<00:00:57.280> beaucoup<00:00:57.500> moins<00:00:57.660> visible.

00:00:58.780 --> 00:01:03.910
Pour<00:00:58.800> voir<00:00:59.000> d'où<00:00:59.140> viennent<00:00:59.300> les<00:00:59.500> biais,<00:01:00.000> il<00:01:00.020> faut<00:01:00.200> remonter<00:01:00.520> à<00:01:00.660> la<00:01:00.840> source:<00:01:01.440> la<00:01:01.460> manière<00:01:01.760> dont<00:01:01.940> un<00:01:02.200> grand<00:01:02.360> modèle<00:01:02.660> de<00:01:02.780> langage<00:01:03.260> apprend.

00:01:04.260 --> 00:01:14.340
Un<00:01:04.400> LLM,<00:01:05.060> ça<00:01:05.280> ne<00:01:05.440> lit<00:01:05.480> pas<00:01:05.640> 3<00:01:05.900> articles<00:01:06.280> et<00:01:06.400> 2<00:01:06.580> CV,<00:01:07.140> ça<00:01:07.260> est<00:01:07.440> entraîné<00:01:07.840> sur<00:01:08.080> des<00:01:08.240> quantités<00:01:08.780> massives<00:01:09.460> de<00:01:09.660> textes:<00:01:10.480> des<00:01:10.520> pages<00:01:10.740> web,<00:01:11.280> des<00:01:11.420> livres,<00:01:11.940> des<00:01:12.080> articles<00:01:12.320> de<00:01:12.660> presse,<00:01:13.120> des<00:01:13.180> transcriptions.

00:01:14.340 --> 00:01:19.880
Il<00:01:14.480> y<00:01:14.640> repère<00:01:14.860> des<00:01:15.020> régularités,<00:01:15.940> des<00:01:16.040> façons<00:01:16.320> d'enchaîner<00:01:16.820> les<00:01:17.020> mots,<00:01:17.560> des<00:01:17.660> tournures<00:01:17.980> typiques,<00:01:18.740> des<00:01:18.820> associations<00:01:19.380> fréquentes.

00:01:20.250 --> 00:01:23.130
Mais<00:01:20.670> ces<00:01:20.810> textes,<00:01:21.330> vous<00:01:21.450> l'imaginez<00:01:21.970> bien,<00:01:22.390> ne<00:01:22.430> sont<00:01:22.590> pas<00:01:22.730> neutres.

00:01:23.390 --> 00:01:27.390
Ils<00:01:23.430> reflètent<00:01:23.790> des<00:01:23.950> habitudes<00:01:24.330> de<00:01:24.490> langage,<00:01:25.190> des<00:01:25.250> stéréotypes,<00:01:26.190> des<00:01:26.270> écarts<00:01:26.550> de<00:01:26.690> visibilité.

00:01:27.730 --> 00:01:31.000
Ils<00:01:27.850> reflètent<00:01:28.230> aussi<00:01:28.530> les<00:01:28.690> inégalités<00:01:29.390> déjà<00:01:29.730> là<00:01:29.990> dans<00:01:30.110> la<00:01:30.250> société.

00:01:31.130 --> 00:01:37.840
Et<00:01:31.210> le<00:01:31.310> modèle,<00:01:31.890> lui,<00:01:32.390> ne<00:01:32.490> fait<00:01:32.650> pas<00:01:32.790> spontanément<00:01:33.370> la<00:01:33.470> différence<00:01:34.010> entre<00:01:34.270> une<00:01:34.450> phrase<00:01:34.730> qui<00:01:34.910> décrit<00:01:35.230> un<00:01:35.370> fait<00:01:35.890> et<00:01:36.110> une<00:01:36.250> phrase<00:01:36.590> qui<00:01:36.750> véhicule<00:01:37.150> un<00:01:37.290> préjugé.

00:01:38.230 --> 00:01:41.580
Il<00:01:38.270> apprend<00:01:38.550> à<00:01:38.670> partir<00:01:38.990> du<00:01:39.250> tout,<00:01:39.950> sans<00:01:40.110> filtre<00:01:40.530> naturel.

00:01:41.710 --> 00:01:46.520
Résultat:<00:01:42.590> un<00:01:42.750> biais<00:01:42.990> ne<00:01:43.110> vient<00:01:43.270> pas<00:01:43.450> seulement<00:01:43.750> d'une<00:01:44.150> grosse<00:01:44.430> erreur<00:01:44.830> ponctuelle<00:01:45.390> qu'on<00:01:45.610> repère<00:01:45.870> facilement.

00:01:46.672 --> 00:01:50.709
Il<00:01:46.816> peut<00:01:46.992> être<00:01:47.216> déjà<00:01:47.393> inscrit<00:01:48.097> silencieusement<00:01:48.898> dans<00:01:49.011> les<00:01:49.139> matériaux<00:01:49.539> d'apprentissage<00:01:50.356> eux-mêmes.

00:01:51.382 --> 00:01:54.634
Ramenez<00:01:51.702> ça<00:01:51.958> au<00:01:52.022> contexte<00:01:52.439> RH<00:01:53.080> et<00:01:53.224> tout<00:01:53.384> devient<00:01:53.865> beaucoup<00:01:54.105> plus<00:01:54.265> concret.

00:01:55.146 --> 00:02:06.169
Si<00:01:55.547> dans<00:01:55.627> les<00:01:55.787> textes<00:01:56.044> utilisés<00:01:56.508> pour<00:01:56.684> entraîner<00:01:56.989> le<00:01:57.069> modèle,<00:01:57.886> on<00:01:57.950> parle<00:01:58.190> beaucoup<00:01:58.431> plus<00:01:58.591> souvent<00:01:58.911> de<00:01:59.071> certains<00:01:59.472> profils,<00:02:00.113> de<00:02:00.193> certains<00:02:00.513> parcours<00:02:01.330> ou<00:02:01.394> de<00:02:01.555> certaines<00:02:01.875> manières<00:02:02.195> de<00:02:02.308> raconter<00:02:02.708> la<00:02:02.836> réussite<00:02:03.157> professionnelle,<00:02:04.358> le<00:02:04.438> système<00:02:04.919> va<00:02:05.031> retrouver<00:02:05.560> ces<00:02:05.720> tendances.

00:02:06.601 --> 00:02:10.895
Il<00:02:06.681> va<00:02:06.793> mieux<00:02:07.002> reconnaître<00:02:07.562> ce<00:02:07.723> qui<00:02:07.819> est<00:02:07.883> très<00:02:08.123> présent<00:02:08.443> dans<00:02:08.604> ces<00:02:08.764> données<00:02:09.341> et<00:02:09.357> moins<00:02:09.597> bien<00:02:09.805> traiter<00:02:10.126> ce<00:02:10.286> qui<00:02:10.398> apparaît<00:02:10.766> peu.

00:02:11.327 --> 00:02:17.623
Certains<00:02:11.728> groupes,<00:02:12.208> certains<00:02:12.545> métiers,<00:02:13.330> certaines<00:02:13.650> situations,<00:02:14.451> certaines<00:02:14.852> régions<00:02:15.573> peuvent<00:02:15.813> tout<00:02:15.973> simplement<00:02:16.550> être<00:02:16.854> moins<00:02:17.094> représentés.

00:02:17.880 --> 00:02:21.310
Et<00:02:18.160> là,<00:02:18.840> le<00:02:18.900> problème<00:02:19.240> n'est<00:02:19.440> pas<00:02:19.600> toujours<00:02:19.940> une<00:02:20.040> réponse<00:02:20.460> choquante.

00:02:21.620 --> 00:02:23.310
Parfois,<00:02:21.880> c'est<00:02:22.120> beaucoup<00:02:22.420> plus<00:02:22.660> discret.

00:02:23.440 --> 00:02:27.800
Une<00:02:23.520> réponse<00:02:23.920> peut<00:02:24.160> être<00:02:24.340> un<00:02:24.480> peu<00:02:24.580> moins<00:02:24.820> précise,<00:02:25.560> un<00:02:25.660> peu<00:02:25.760> moins<00:02:26.020> nuancée,<00:02:26.760> un<00:02:26.840> peu<00:02:26.960> moins<00:02:27.200> utile.

00:02:28.280 --> 00:02:33.840
Selon<00:02:28.340> le<00:02:28.560> profil,<00:02:29.300> selon<00:02:29.420> la<00:02:29.580> situation,<00:02:30.660> la<00:02:30.720> qualité<00:02:31.060> peut<00:02:31.300> varier<00:02:31.720> sans<00:02:32.020> que<00:02:32.120> ça<00:02:32.280> se<00:02:32.580> voit<00:02:32.620> au<00:02:32.780> premier<00:02:33.080> coup<00:02:33.240> d'œil.

00:02:34.360 --> 00:02:39.060
C'est<00:02:34.460> un<00:02:34.640> point<00:02:34.800> clé:<00:02:35.320> le<00:02:35.480> biais<00:02:35.840> ne<00:02:36.020> se<00:02:36.420> niche<00:02:36.421> pas<00:02:36.660> seulement<00:02:37.000> dans<00:02:37.140> le<00:02:37.300> contenu<00:02:37.640> visible<00:02:38.040> d'une<00:02:38.300> phrase.

00:02:39.060 --> 00:02:42.790
Il<00:02:39.300> peut<00:02:39.480> aussi<00:02:39.780> se<00:02:40.200> loger<00:02:40.280> dans<00:02:40.540> la<00:02:40.700> qualité<00:02:41.380> inégale<00:02:41.960> des<00:02:42.140> réponses.

00:02:43.020 --> 00:02:47.920
La<00:02:43.180> même<00:02:43.400> demande<00:02:43.840> formulée<00:02:44.400> pour<00:02:44.600> différents<00:02:45.040> profils<00:02:45.540> peut<00:02:46.060> ne<00:02:46.260> pas<00:02:46.400> recevoir<00:02:46.780> le<00:02:47.120> même<00:02:47.280> niveau<00:02:47.500> de<00:02:47.700> richesse<00:02:47.860> ou<00:02:47.880> de<00:02:47.900> fiabilité.

00:02:49.410 --> 00:02:54.540
La<00:02:49.630> même<00:02:49.830> question<00:02:50.350> posée<00:02:50.710> dans<00:02:50.890> 2<00:02:51.110> langues<00:02:51.290> différentes<00:02:51.970> peut<00:02:52.350> donner<00:02:52.590> des<00:02:52.810> résultats<00:02:53.310> très<00:02:53.590> différents.

00:02:55.690 --> 00:02:58.070
Et<00:02:55.770> là,<00:02:56.270> la<00:02:56.350> question<00:02:56.630> des<00:02:56.830> langues<00:02:57.030> devient<00:02:57.370> centrale.

00:02:58.390 --> 00:03:01.520
Les<00:02:58.470> corpus<00:02:58.750> d'entraînement<00:02:59.250> ne<00:02:59.470> couvrent<00:02:59.710> pas<00:02:59.970> toutes<00:03:00.210> les<00:03:00.430> langues<00:03:00.570> de<00:03:00.710> la<00:03:00.830> même<00:03:00.970> façon.

00:03:01.670 --> 00:03:03.850
Quelques<00:03:01.830> langues<00:03:02.030> dominent,<00:03:02.730> surtout<00:03:02.990> sur<00:03:03.150> le<00:03:03.350> web.

00:03:04.010 --> 00:03:07.590
D'autres<00:03:04.090> sont<00:03:04.410> moins<00:03:04.650> présentes,<00:03:05.430> moins<00:03:05.629> détaillées,<00:03:06.410> moins<00:03:06.570> documentées.

00:03:07.590 --> 00:03:14.690
Conséquence<00:03:07.830> directe:<00:03:08.710> les<00:03:08.810> performances<00:03:09.370> peuvent<00:03:09.670> changer<00:03:09.990> selon<00:03:10.250> la<00:03:10.530> langue<00:03:10.690> utilisée,<00:03:11.610> mais<00:03:11.730> aussi<00:03:12.070> selon<00:03:12.310> le<00:03:12.490> contexte<00:03:13.030> culturel<00:03:13.590> associé<00:03:14.090> à<00:03:14.250> cette<00:03:14.390> langue.

00:03:14.940 --> 00:03:19.560
Dans<00:03:15.060> un<00:03:15.260> cadre<00:03:15.480> professionnel,<00:03:16.620> cette<00:03:16.720> différence<00:03:17.260> n'est<00:03:17.520> pas<00:03:17.760> anecdotique,<00:03:18.720> elle<00:03:18.860> peut<00:03:19.060> peser.

00:03:19.820 --> 00:03:27.730
Un<00:03:19.880> même<00:03:20.100> système<00:03:20.520> peut<00:03:20.720> paraître<00:03:21.120> solide,<00:03:22.120> très<00:03:22.260> convaincant<00:03:22.740> dans<00:03:22.880> un<00:03:23.040> contexte<00:03:23.720> abondamment<00:03:24.100> documenté,<00:03:25.260> puis<00:03:25.480> devenir<00:03:25.940> plus<00:03:26.300> fragile<00:03:26.600> dans<00:03:26.840> un<00:03:26.980> autre.

00:03:28.400 --> 00:03:31.780
À<00:03:28.680> cela<00:03:28.940> s'ajoute<00:03:29.360> un<00:03:29.520> autre<00:03:29.840> niveau:<00:03:30.440> les<00:03:30.540> choix<00:03:30.760> de<00:03:31.060> conception.

00:03:31.780 --> 00:03:36.260
Le<00:03:31.880> pré-entraînement,<00:03:32.480> lui,<00:03:33.300> apprend<00:03:33.400> surtout<00:03:33.740> une<00:03:33.900> chose:<00:03:34.720> prédire<00:03:35.020> la<00:03:35.200> suite<00:03:35.420> d'un<00:03:35.680> texte.

00:03:36.260 --> 00:03:40.670
Pas<00:03:36.440> distinguer<00:03:36.940> ce<00:03:37.120> qui<00:03:37.300> est<00:03:37.440> juste<00:03:37.800> ou<00:03:38.120> injuste,<00:03:38.960> acceptable<00:03:39.400> ou<00:03:39.920> problématique.

00:03:40.966 --> 00:03:48.307
Ensuite,<00:03:41.783> on<00:03:41.847> règle<00:03:42.184> le<00:03:42.328> système<00:03:42.745> avec<00:03:42.969> des<00:03:43.130> consignes,<00:03:43.851> des<00:03:43.931> préférences<00:03:44.508> humaines,<00:03:45.294> des<00:03:45.374> critères<00:03:45.855> pour<00:03:46.015> juger<00:03:46.352> une<00:03:46.496> réponse<00:03:47.073> utile<00:03:47.778> ou<00:03:47.794> appropriée.

00:03:48.756 --> 00:03:51.754
Et<00:03:48.820> là<00:03:48.916> encore,<00:03:49.702> ces<00:03:49.862> choix<00:03:50.279> influencent<00:03:50.824> le<00:03:50.904> comportement<00:03:51.465> final.

00:03:52.187 --> 00:04:03.600
Dans<00:03:52.363> un<00:03:52.523> outil<00:03:52.844> RH,<00:03:53.565> une<00:03:53.709> consigne<00:03:54.271> comme<00:03:54.751> valoriser<00:03:55.264> le<00:03:55.361> leadership<00:03:56.371> ou<00:03:56.595> privilégier<00:03:57.268> l'efficacité<00:03:58.519> peut<00:03:58.599> sembler<00:03:58.999> très<00:03:59.240> raisonnable,<00:04:00.122> mais<00:04:00.298> elle<00:04:00.458> oriente<00:04:00.763> déjà<00:04:01.083> le<00:04:01.196> système<00:04:01.725> vers<00:04:01.885> certaines<00:04:02.366> normes<00:04:02.847> plutôt<00:04:03.167> que<00:04:03.247> d'autres.

00:04:03.905 --> 00:04:06.261
Et<00:04:03.985> on<00:04:04.033> n'utilise<00:04:04.530> pas<00:04:04.690> tous<00:04:04.931> cet<00:04:05.267> outil<00:04:05.492> de<00:04:05.604> la<00:04:05.684> même<00:04:05.892> manière.

00:04:06.694 --> 00:04:15.244
Le<00:04:06.774> mobiliser<00:04:07.415> pour<00:04:07.560> reformuler<00:04:08.153> un<00:04:08.297> texte<00:04:08.858> ou<00:04:08.954> explorer<00:04:09.419> des<00:04:09.595> idées,<00:04:10.381> ce<00:04:10.413> n'est<00:04:10.621> pas<00:04:10.701> la<00:04:10.782> même<00:04:10.942> chose<00:04:11.343> que<00:04:11.375> l'intégrer<00:04:11.904> dans<00:04:12.080> une<00:04:12.224> chaîne<00:04:12.545> de<00:04:12.625> décision<00:04:12.866> sur<00:04:13.348> des<00:04:13.460> candidatures<00:04:14.408> ou<00:04:14.473> des<00:04:14.650> évaluations.

00:04:15.678 --> 00:04:27.137
Dès<00:04:15.839> que<00:04:16.000> le<00:04:16.080> système<00:04:16.498> agit<00:04:16.803> près<00:04:16.964> d'une<00:04:17.205> décision<00:04:17.687> professionnelle,<00:04:18.892> les<00:04:19.053> biais<00:04:19.471> hérités<00:04:19.857> des<00:04:20.017> données<00:04:20.435> et<00:04:20.500> des<00:04:20.660> choix<00:04:20.901> de<00:04:20.982> conception<00:04:21.946> peuvent<00:04:22.187> se<00:04:22.300> transformer<00:04:22.926> en<00:04:23.087> écarts<00:04:23.553> très<00:04:23.794> concrets<00:04:24.437> pour<00:04:24.550> des<00:04:24.678> personnes<00:04:25.080> réelles,<00:04:25.803> pour<00:04:25.964> des<00:04:26.044> trajectoires<00:04:26.623> de<00:04:26.768> carrière.

00:04:27.812 --> 00:04:32.843
C'est<00:04:28.069> exactement<00:04:28.632> à<00:04:28.696> partir<00:04:29.018> de<00:04:29.178> là<00:04:29.419> que<00:04:29.500> la<00:04:29.660> vigilance<00:04:30.239> n'est<00:04:30.464> plus<00:04:30.721> une<00:04:30.801> option,<00:04:31.444> elle<00:04:31.589> devient<00:04:32.151> nécessaire.