Three new limits on what you can publish, and three receipts on what you can build.
YouTube reorganized its inauthentic content policy into three categories targeting low-effort, repetitive, misleading, or harmful content. AI tools for scripts, voiceovers, visuals remain fine; the risk is content that is low-effort or deceptive. Creators following faceless-channel playbooks should pay attention.
The EU AI Act's Article 50 transparency rules took effect on 2 August 2026, requiring providers to embed machine-readable AI content labels and deployers to disclose deepfakes and certain AI-generated text. But research on the 'implied truth effect' argues that labeling only some content makes unlabeled content look more trustworthy, while the 'liar's dividend' lets bad actors dismiss genuine footage as fake. Most businesses are deployers, not providers, and face penalties up to €15m or 3% of turnover, with a grace period for machine-readable marking until December 2026.
The MPA and ByteDance signed an MOU codifying AI guardrails across TikTok, Seedance, CapCut, and Dreamina. The agreement formalizes technical safeguards already shipped in March: face-blocking, copyrighted-character filters, visible C2PA watermarks, and invisible watermarking. Enforcement details, technical thresholds, and a U.S. rollout timeline remain undisclosed, and the practical effect for creators is that generating real faces or clearly copyrighted characters is now explicitly blocked and watermarked, though likeness-adjacent outputs still fall through the cracks.
A hands-on teardown of editing an AI song with Google's Flow Music (Lyria 3.5) shows the chorus survives edits, but runtime shifts. The author changed intro length and lead instrument from a single baseline; the six-word chorus persisted across all edits (6/6), yet both successful edits cut runtime by about 2.5 seconds. The article offers a practical acceptance framework: accept when the target change is audible and protected elements survive, retry on weak or ambiguous edits, and regenerate only when the core hook is at risk.
Keegan McCallum argues the interesting axis in generative video has shifted from quality to efficiency. uRun's Helios model, a distillation of a 14B-param open model, costs about $10 for three hours of continuously generated video. Real-time generation means you can steer live and use webcam magic mirrors, while slower high-quality clips cost ~100x more. The bottleneck is serving: globally positioned GPUs, WebRTC with ICE/TURN, and a synchronized streaming pipeline.
A reusable prompt template for Black Forest Labs' FLUX 3 generates 20-second VHS-style found footage clips. The recipe opens mid-action with no setup, names physical defects like 'a tracking bar drifting up frame' instead of saying 'VHS aesthetic', uses timestamped beats with hard cuts, budgets dialogue at 2.5 words per second, spells out audio layers, and puts negatives last. A worked example (a sitcom clip about a man protecting a lobster) worked on the first try. Settings include 4:3 aspect ratio and 720p, with advice to iterate one layer per pass and avoid on-screen text.