• AI KATANA
  • Posts
  • Meta plans to replace humans with AI to assess privacy and societal risks

Meta plans to replace humans with AI to assess privacy and societal risks

Also: OpenAI’s Third Core Device Signals a New Era of Everyday AI

Author

AI KATANA
June 01, 2025

Hello!

In today’s edition, the AI landscape is evolving at a dizzying pace, with major players making bold—and sometimes controversial—moves. Meta is overhauling its internal oversight systems by replacing human judgment with AI for privacy and risk evaluations, raising serious concerns about the future of ethical guardrails. Meanwhile, Google’s Veo 3 is generating buzz and alarm in equal measure with its near-cinematic AI-generated video clips that blur the line between real and synthetic. OpenAI and Microsoft are tightening their enterprise dominance through an intricate mix of partnership and rivalry, even as regulatory pressure mounts. In the open-weight AI space, DeepSeek’s R1-0528 release challenges proprietary incumbents with improved performance and transparency. And OpenAI is venturing into ambient computing with a new screenless AI device, aiming to redefine how we interact with technology in everyday life. The momentum continues across AI infrastructure and VC funding, underscoring a future where innovation must be balanced with responsibility, usability, and governance.

Sliced just for you:

  • 🤖 Meta plans to replace humans with AI to assess privacy and societal risks

  • 🎬 Google’s Veo 3 AI video generator is unlike anything you’ve ever seen. The world isn’t ready.

  • 💼 OpenAI and Microsoft Partnership: Enterprise AI Showdown

  • 🧠 DeepSeek R1-0528: How the Latest Open-Weight Update Challenges Closed-Weight AI Giants

  • 🆕 OpenAI’s Third Core Device Signals a New Era of Everyday AI

Meta is overhauling its internal risk assessment process by replacing most human-led reviews with AI systems, automating up to 90% of decisions that previously required scrutiny for privacy, youth safety, misinformation, and content integrity. While the shift is touted internally as a way to speed up product development and reduce bureaucratic delays, it has raised concerns among current and former employees who warn that removing human judgment from complex risk evaluations could amplify real-world harm. Despite Meta’s claim that human expertise will still be applied to novel or high-risk issues, internal documents reveal plans to automate assessments in sensitive areas such as AI safety and violent content. Critics argue the move undermines safeguards built over years in response to regulatory scrutiny, notably by the FTC, and suggest it reflects a broader trend toward rapid iteration and unfiltered speech, aligned with recent political shifts. The change also leaves risk calls increasingly in the hands of product teams not trained in privacy or ethics, potentially turning crucial reviews into perfunctory checklists rather than robust defenses against platform abuse.

Google’s newly released Veo 3 is redefining the boundaries of AI-generated video content, producing 8-second clips with hyperrealistic visuals, synchronized audio, dynamic dialogue, and consistent characters—ushering in a new creative era that’s both exhilarating and alarming. Artists are already leveraging the tool to create short films like “Influenders,” which showcase Veo 3’s capability to simulate real-world events and interactions with startling accuracy. Available through Google’s Gemini and Flow platforms, Veo 3 significantly reduces production barriers but also raises urgent questions about misinformation and digital trust. Despite Google’s integration of digital watermarks and safety protocols, the ease with which deepfakes and synthetic news clips can now be produced has ignited widespread concern among researchers and creators alike. As the technology rapidly evolves, the public’s ability to discern real from synthetic content is being tested like never before.

OpenAI and Microsoft’s strategic partnership has redefined the enterprise AI landscape by integrating OpenAI’s advanced models with Microsoft’s Azure cloud and productivity tools, creating a dominant enterprise offering used by over 60% of Fortune 500 firms. While Microsoft enjoys exclusive cloud distribution rights and embeds AI in products like Microsoft 365 Copilot, OpenAI simultaneously competes with its own SaaS offerings like ChatGPT Enterprise—creating both collaboration and competition. Regulatory scrutiny is intensifying, especially in the EU, over concerns of platform concentration and exclusivity. Meanwhile, enterprises face pricing pressure around a $30-per-user benchmark and are adapting to governance standards like the NIST AI RMF. As demand grows for customizable, multi-model deployments and on-prem options using Azure Stack HCI, Microsoft positions itself as the “AI ringleader” by hosting third-party models such as Claude and Gemini. Yet, OpenAI’s rapid innovation pipeline, including GPT-5, may eventually outpace Azure’s infrastructure, hinting at a future where today’s allies could become rivals. CIOs are advised to plan for agility, vendor neutrality, and ongoing antitrust developments to remain competitive in a shifting AI ecosystem.

DeepSeek’s R1-0528 update marks a major stride for open-weight AI, narrowing the performance gap with proprietary leaders like OpenAI o3 and Gemini 2.5 Pro. The release significantly boosts reasoning and math benchmarks, reduces hallucination rates by 21%, and introduces a distilled 8B-parameter model that runs on a single consumer GPU—enabling low-cost, edge, and on-prem deployments. Native support for JSON and function calling improves developer efficiency, while transparency around model weights fosters third-party auditing and safety testing. However, with no access to training data or pipeline, R1-0528 is technically open-weight, not open-source—a distinction with implications for reproducibility, policy, and trust. As global scrutiny intensifies over AI governance, DeepSeek’s release challenges the closed-weight status quo by offering high performance without a usage fee or opaque API, making it a compelling choice for start-ups, enterprises, and academic researchers seeking flexibility, cost control, and auditability in deployment.

OpenAI’s latest move into hardware signals a bold attempt to define a new “third core device” beyond phones and laptops—this time, a screen-free, voice- and gesture-controlled AI companion. Enabled by its $6.5 billion acquisition of Jony Ive’s hardware startup, the upcoming device is expected in 2026 and aims to offer always-on contextual awareness without the friction of screens or keyboards. Early details suggest the device could serve as a personal memory aid and real-time assistant while staying discreet and minimal, echoing Ive’s history of consumer-friendly design. Success hinges on solving battery life, privacy, and user clarity challenges—avoiding pitfalls that doomed earlier wearables like the Humane AI Pin. Meanwhile, Google is countering with Gemini-powered smart glasses that leverage Android’s ecosystem, betting on seamless distribution and fashion-forward design. As the race heats up, the future may not belong to a single interface but a blended landscape where phones coexist with ambient, glance-free AI companions—transforming how users interact with technology in everyday moments.

🛠️ AI tools updates

AWS has open-sourced MCP servers for Lambda, ECS, EKS and Finch, giving AI code assistants live hooks into serverless and container metadata. By exposing resource graphs, logs and deployment events via a standard gRPC endpoint, MCP lets agents craft infrastructure-aware suggestions—think generating IAM policies or CICD fixes without human context-dumping. Early adopters report 25 % faster troubleshooting in staging clusters. AWS positions MCP as a neutral layer that could work with rival assistants, but critics wonder if it nudges devs deeper into the AWS ecosystem.

Google quietly added Gemma 3n, a 3-billion-parameter multimodal model, to AI Studio and Vertex AI. Despite its small size, Gemma 3n handles image, text, audio and short video, running in real time on recent laptops and even premium phones. Benchmarks show it matching larger competitors on WebDev Arena tasks while sipping a fraction of the power. Developers can fine-tune with LoRA adapters, and Google promises to open-source weights once hardware-specific optimizations stabilize. The release underscores a trend toward lightweight, edge-friendly models that still play nice with cloud pipelines when extra muscle is needed.

💵 Venture Capital updates

Writing assistant Grammarly has obtained a $1 billion non-dilutive financing facility, earmarked for evolving from grammar checker to full-stack workplace copilot. The company plans to hire domain experts, integrate spreadsheet and presentation drafting, and expand its enterprise analytics dashboard. The sizable credit line avoids equity dilution while giving Grammarly room to chase corporate contracts before giants like Microsoft bundle similar features deeper into Office.

Cyber-security firm Horizon3.ai is lining up a $100 million round that would value it near $750 million. Its flagship NodeZero platform uses reinforcement-learning agents to probe corporate networks, rank exposures and auto-generate remediation scripts. The fresh capital will double R&D staff and push into OT and IoT environments—lucrative but historically underserved niches. Investors view automated red-teaming as a must-have as corporate attack surfaces balloon with cloud micro-services and edge devices.

🫡 Meme of the day

⭐️ Generative AI image of the day