Sony re-sues Udio over 30,000 songs, and Disney's Ozzy Fox goes viral off a false AI-slop accusation.
Deezer reports AI-generated music now accounts for over 50% of daily uploads, peaking at 90,000 tracks per day in June 2026. Despite volume, AI music makes up only 1-3% of streams, with up to 85% of those streams identified as fraudulent and demonetized. The platform will remove tracks used for fraud or unstreamed for six months.
YouTube updated its Partner Program rules to explicitly ban three categories of inauthentic content from earning revenue: generic/repetitive/template-based videos, emotionally manipulative or off-putting content, and AI personas discussing sensitive topics like health and finance. The policy targets content farming enabled by AI tools while preserving room for high-quality AI assisted creativity. Creators relying on mass produced low effort AI videos risk losing monetization eligibility.
Sony Music filed a new copyright infringement lawsuit against AI music platform Udio, claiming over 30,000 sound recordings were used without permission to train its models. A judge barred Sony from adding those recordings to its original case, so Sony filed separately. The suit alleges Udio stream-ripped tracks from YouTube and bypassed technological protections. Udio has licensing deals with other major labels but not Sony. Statutory damages could reach $150,000 per work. The outcome could affect the legality of AI music generation tools and set precedent for training data licensing.
HeyGen's engineering lead explains why they built HyperFrames on plain HTML, CSS, and JavaScript instead of React/Remotion. The insight: HTML is the native language of LLMs, so small models like Gemini Flash can produce creative, deterministic MP4 videos without guardrails. They used self-improving evals to teach agents 'taste' instead of constraints.
Disney's preschool show Ozzy Fox blew up after viewers claimed it was fully AI generated. The actual production pipeline uses AI assisted tools like sketch to pose prediction and motion inbetweening, not generative AI. Human animators and proprietary AI trained on show assets handled the work. The controversy pushed viewership past 500,000 per episode. For builders, the incident shows how quickly audiences turn against perceived AI use, making transparent workflows a real business risk.
Alibaba's Qwen-Image-3.0 is pitched at one thing: making generated images useful enough to actually ship. It renders text as small as 10 pixels, takes prompts up to 4.5k tokens, and laid out a full 3x3 grid of distinct infographics (a math slide, a parasitology explainer, a bank control diagram) in a single pass across 12 languages. The honest caveat: it launched with no open weights, no model card, and no benchmarks, and access is an invite-only API bound for Qwen Chat, so for now you are trusting the cherry-picked example gallery until you can run your own prompts.
fal.ai released an API for the CVPR 2025 Video Depth Anything model. It generates temporally consistent depth maps from video, with three model sizes (Small, Base, Large), five colormaps, side-by-side comparison with the original video, and raw depth export as .npz. At $0.04 per second of video, it's a practical tool for 3D reconstruction, video effects, compositing, and scene understanding.
A tested guide provides 9 copy-paste AI travel photo prompts using a five-part structure that transforms ordinary selfies into travel photos. Only 5% of surveyed travelers could distinguish AI from real photos. The prompts work best in GPT-Image-2 Thinking mode, and the guide emphasizes honest labeling when posting.