Meta glisse des opérateurs humains derrière l'agent Muse
TL;DR
- Reuters obtained internal Meta posts showing employees raised privacy alarms before 404 Media published, including a documented case where a contractor made a racist reference during a call.
- A Meta VP overseeing the product acknowledged 'it was a miss' to start testing contractor-placed calls without user disclosure and confirmed the feature was rolled back.
- The trial mirrors Facebook's 2015 M assistant, where roughly 70 percent of responses were human-generated behind an AI interface, the same deception architecture reused a decade later.
Meta a déployé sur son agent vocal Muse une couche d'opérateurs humains chargés de boucler les appels — réservations, rendez-vous — que le produit est censé passer de manière autonome, rapporte 404 Media sur la foi de communications internes de l'entreprise.
Dans un message aux salariés, Meta indique avoir « added a human agent layer for calls to get completed » et déclare que « Muse human agent calls is ready for company dogfooding ». Le pitch officiel du produit, signé par l'ingénieur principal Ryan Fox, décrit une IA qui « calls a business on your behalf, handles the conversation, completes your request » avant d'ajouter que « Muse is now able to hand requests to a trained human agent, who places the call and works it through ».
En interne, l'annonce est mal passée. Un employé prévient: « This is absolutely going to create a ton of outcry if we publicly launch this. » Un autre juge l'approche « bad bad idea ». Selon un compte-rendu de test cité par 404 Media, un utilisateur « was not made aware of the caller being human and was only told after ».
Interrogée, une porte-parole de Meta répond que l'entreprise « will only roll it out when it's ready and with the proper disclosures ». La révélation tombe au terme d'une semaine chargée pour le produit: Wardle venait d'exposer un zero-day macOS livrant le token de Muse, et Nat Friedman a reconnu que Muse s'inspirait « fortement » d'OpenClaw.
Ce qu'en disent les autres médias
-
Reuters Lire →
Original exclusive sourced from internal company posts; documents employee opposition, the racist contractor incident, a VP rollback admission, and the Facebook M historical parallel.
It's baffling to me why we think this feature is worth the risk.
-
Crypto Briefing Lire →
Explains that Meta's sandboxed VM protections did not extend to the human handoff layer, framing the pause as a structural redesign problem rather than a policy adjustment.
Privacy concerns prompted Meta to pause the trial.
-
Inside AI Lire →
Frames the disclosure gap against EU AI Act requirements for transparency when humans participate in AI decision-making, adding regulatory exposure that U.S.-focused outlets omitted.
We will only roll it out when it's ready and with the proper disclosures.
Shared on Bluesky by 6 AI experts (top 5 by trust)
-
Scoop: Meta's Muse AI can make phone calls. Internally, Meta has told employees that many of these calls will be made by human beings in a call center: www.404media.co/meta-tests-m...
View on Bluesky → -
lol www.404media.co/meta-tests-m...
View on Bluesky →
Article original publié par 404media.co
Lire l'article original →Titre original : 404 Media : les appels vocaux Muse de Meta sont en fait passés par des humains en centre d'appels