Meta drops Sama after footage review

Meta just fired its data labeling partner for the Ray-Ban Meta smart glasses. The glasses capture first-person video to train Meta's AI — but contractors ended up watching people have sex, fight, and worse. Meta cut ties with Sama this week after an internal review confirmed the exposure. The fallout spotlights the messy reality of training AI on unfiltered human lives. Glasses keep selling, but Meta's scrambling to reassure users their intimate moments aren't fueling LLMs. (stuff.tv) ### What are Ray-Ban Meta glasses? These are regular Ray-Bans with built-in cameras and AI smarts from Meta. Wearers get hands-free photos, video calls, live translation, and a new "Look and Ask" feature where AI describes what you're seeing. The catch — every interaction generates video clips uploaded to Meta's cloud for processing. A white LED lights up during recording to signal others. But people still capture private stuff, from bathrooms to bedrooms. (about.fb.com) ### Why hire Sama for footage review? Meta outsources the grunt work of turning raw glasses video into AI training data. Sama, a Nairobi-based firm, employs thousands in Kenya, India, and Kenya to watch clips, label objects, actions, and contexts — think "person running" or "dog barking." It's tedious, low-paid labor essential for multimodal AI like Llama. Meta sends anonymized snippets; Sama tags them for machine learning. Turns out, filters failed — reviewers saw unredacted explicit content. (sama.com; time.com) ### What exactly did reviewers see? Sama workers viewed graphic clips: people having sex, masturbating, nude in showers, kids undressing, drug use, self-harm, even corpses. One whistleblower told Time they processed 1,000+ videos daily, 8 hours a day, with little mental health support. Violence popped up too — stabbings, road rage beatings. Meta's rules banned sensitive content, but it slipped through post-capture. Reviewers got vague instructions like "flag nudity," but volume overwhelmed safeguards. (time.com) ### Did the glasses' light prevent this? The white recording LED aims to get consent — it blinks obviously when active. But clips get reviewed later for AI improvement, not real-time. Users might forget it's on during private moments, or angles catch bystanders. Meta anonymizes data by blurring faces and stripping metadata, but full context remains for labeling accuracy. Privacy advocates call it a "surveillance nightmare" since footage feeds central servers. (theverge.com) ### Why drop Sama now? Reports surfaced in 2023 via Time, with Sama employees unionizing over trauma. Meta audited and found Sama breached policies by exposing workers to prohibited content without proper safeguards. Contract ended April 2024 — no more glasses data to Sama. Meta insists it never used explicit clips for training and paused new uploads during review. But backlash builds as glasses sales boom. (mezha.ua; time.com) ### What's Meta doing to fix it? Meta's tightening data pipelines: better pre-screening AI filters, no human review of flagged sensitive clips, and in-house labeling alternatives. They're emphasizing opt-in uploads and easier deletion. CEO Mark Zuckerberg downplayed it as an "industry challenge." Critics want hardware limits, like no cloud upload without explicit user tap. EU probes loom under GDPR for biometric data handling. (engadget.com) ### Is this a Sama or Meta problem? Both. Sama's faced heat before — OpenAI ditched them in 2022 over Kenyan workers labeling violent CSAM for pennies. Meta knew the risks but prioritized scale. Now, it threatens AI glasses adoption. Competitors like Google and Snap watch closely — their wearables avoid heavy cloud reliance. Bottom line: turning eyeglasses into AI brains demands ironclad privacy, or trust evaporates fast. Expect lawsuits and regs to tighten. (restofworld.org) Word count: 578 ```