All articles tagged with #ai training
Ford admitted its AI-driven design effort faltered as experienced workers left before their know-how could be embedded, triggering a rehiring/training push that brought back about 350 veterans to fix the AI and expand testing, even as the company contends with reputational damage and ongoing quality concerns.
The New York Times moves to amend its copyright lawsuit against OpenAI and Microsoft, arguing Microsoft actively aided training AI on NYT works via a bespoke supercomputer and that a recent Supreme Court ruling sharpens the standard for contributory infringement; two other claims are being dropped, while the core accusation that the training harmed NYT’s business and was conducted without permission remains, with OpenAI defending fair use.
A federal judge denied Meta’s bid to dismiss Strike 3 Holdings’ lawsuit, allowing claims that Meta coordinated torrenting of Strike 3’s adult videos to train its AI models to proceed. Evidence shows 47 Meta IP addresses torrented thousands of Strike 3 videos (about 81 terabytes) via Anna’s Archive from 2018–2025, with files having coordinated naming patterns. The court rejected Meta’s argument that only rogue employees were involved and held that the act of torrenting itself violated copyright, enabling direct, vicarious, and contributory infringement claims. The ruling signals potential liability for platforms that scrape copyrighted adult content for AI training.
Meta’s three-month-old Applied AI unit, now about 6,500 engineers and product managers, faces serious internal strain as workers describe being involuntarily drafted into the group to build AI training data, performing tasks like generating puzzles and coding problems. A livestream meltdown during an employee presentation highlighted the tensions, and more than 1,600 Meta staff reportedly signed a petition over monitoring clicks and keystrokes for AI training. CEO Mark Zuckerberg acknowledged distress in an internal memo and emphasized Meta’s aim to retain top talent, but morale remains extremely low as leadership reshapes theAI effort.
Meta will let some employees pause its keystroke- and mouse-tracking for up to 30 minutes to handle personal matters, while the majority remain tracked. A limited opt-out is available for remote workers with bandwidth issues, those dealing with sensitive material, or where power supply is a constraint. The tracking, part of the Model Capability Initiative to train AI, has faced protests, and Meta says battery usage has improved.
Meta is facing employee backlash over its Model Capability Initiative (MCI), a tool reportedly tracking U.S. workers’ mouse movements, keystrokes, and other activity to train AI agents. Critics say the scope could extend to code changes, device sleep/wake cycles, and clipboard contents, and may run afoul of GDPR if EU data is implicated; Meta says MCI is installed only on U.S. computers with safeguards and informed non-U.S. employees, but Reuters reports and internal concerns have raised questions about transparency and data usage.
Kotaku reports leaked audio in which Mark Zuckerberg says Meta's AI is trained by watching ‘really smart people’ at the company as about 8,000 employees are laid off (10% of the workforce) and thousands more reassigned to AI initiatives, amid criticism over employee device tracking and Meta’s ongoing AI pivot.
Five major publishers and author Scott Turow filed a class-action suit against Meta and Mark Zuckerberg, accusing Meta of training its generative AI platform Llama on millions of pirated books and articles without authorization, threatening publishers’ and authors’ livelihoods; Meta says such AI training can be fair use and vowed to fight the case in SDNY.
Ars Technica reports that Google's Gemini AI, embedded in Gmail and Drive, processes data for isolated tasks and may feed outputs into AI training, though Google says private content isn't used to train foundational models. Opting out of training is hard and buried in Gemini's Activity controls; disabling it often sacrifices useful features, a setup critics call dark patterns that undermine user autonomy. The article argues defaults push data collection and AI integration, raising ongoing privacy concerns amid antitrust scrutiny.
LinkedIn is testing an 'AI labor marketplace' that pays humans up to $150/hour to train AI systems across roles from coding to nursing, signaling a new gig category and placing LinkedIn in competition with startups like Mercor and Surge AI. The initiative highlights rapid growth in AI training work and a range of roles—from Excel/finance experts to nurses and linguists—while also underscoring cybersecurity risks in the sector, including notable data breaches at industry players.
The Register exposes a sprawling network of ad-tech vendors whose cookie and non-cookie trackers collect vast user data—IP addresses, device identifiers, location (precise and non-precise), profiles, and browsing activity—often for long durations and sometimes under legitimate-interest justifications. The piece highlights how this data ecosystem, plus storage beyond cookies, could be used to train AI models and influence ad targeting, raising privacy concerns over how much online activity is captured and stored across dozens of vendors.
Across the globe, thousands are monetizing everyday data—videos, ambient audio, even private chats—to train AI via marketplaces like Kled AI, Silencio, and Neon Mobile. Contributors in places like Cape Town, Ranchi, and Chicago earn small sums, often through broad, irrevocable licenses that allow boundless use and derivative works, with pay sometimes in USD but little recourse. While the money can help close gaps for people in economic hardship, experts warn the practice risks privacy harms, deepfakes, identity theft, and a precarious, wage-driven boom that mainly benefits platforms in wealthier countries as AI data needs outpace scraping from the open web.
Britannica and Merriam-Webster filed suit against OpenAI alleging that GPT-4 memorized substantial portions of their copyrighted content and outputs near-verbatim passages to train its AI, potentially diverting traffic away from Britannica; the case adds to a wave of publisher lawsuits over AI training and follows similar actions against OpenAI and a prior Anthropic settlement.
A Svenska Dagbladet investigation reveals Meta's Ray‑Ban smart glasses capture footage that is reviewed by human contractors for AI training, exposing bystanders to nude or sensitive content despite anonymization, highlighting ongoing privacy risks with wearable devices.
Swedish reports say offshore Meta workers in Kenya are labeling intimate and disturbing videos captured by Ray‑Ban smart glasses to train AI, including footage from bathrooms and clips showing personal data. Working under Sama, they describe being required to review such material, raising privacy and consent concerns as Meta pushes live AI features and potential facial recognition, amid related moderator‑work lawsuits.