Plus: Google labels the AI ads made with its own tools and leaves everyone else self-reporting.
Google's AI ad labels went live July 9 across Search, YouTube and Discover, in the My Ad Center panel. Build the ad with Google's own AI tools and the disclosure is added for you. Use any other tool and it is a control you can choose to switch on, so third party AI runs on the honour system. No requirement, no deadline, whatever the coverage says.
Netflix told its Q2 call that partners used generative AI on 300 titles this year, mostly in post. The receipt behind that number: 17 minutes of The American Experiment, the docuseries with Martin Sheen voicing George Washington, came in twice as fast and at half the cost of previous options. Sarandos says the savings get reinvested into more content.
Source code from the Suno breach shows 2,013,545 clips scraped from YouTube Music alone, with Deezer, Genius, Pond5, IMSLP and Jamendo named as further sources. The scrape ran through Bright Data proxies and separately circumvented YouTube's rolling cipher. That second detail is the expensive one: labels can seek up to $150,000 per work plus $2,500 per act of circumvention, putting exposure past $9 billion. The hacker also took customer emails, phone numbers and Stripe details.
Reuters ran 40 images made with Meta's Muse Image generator back through Meta's Content Seal detector. At full size it flagged all 40; cropped to roughly a third or half, with no other edit, it missed 55% of them. Not a screenshot, not an adversarial attack, just a crop, the most routine edit there is. Meta's own wording is that the signal may be lost if an image is heavily cropped.
Adam Mosseri said in his weekly Q&A that Instagram's AI image and video generation, today rate limited by daily caps, will move to a paid tier. His actual words: eventually you're going to be able to subscribe to be able to get access to more. No price, no date, no tier detail, so budget for it rather than act on it.
Vids now takes a prompt plus a reference image, then lets you refine by talking to it, swapping a background or adjusting lighting without regenerating the shot. Avatars are the catch: only of yourself, 18 plus, region limited, SynthID watermarked. It is gated behind Google AI Pro or Ultra, or an eligible Workspace business plan, and Google publishes no clip length, resolution or quota figures.
aicameramovements.com lists 46 camera moves across 7 categories, each with a reference clip and a copy-paste prompt structured as camera, movement, speed, framing, end. The structure is the useful part: camera instruction stays separate from scene content, so a prompt survives a change of first frame. Two caveats. Nothing claims the prompts were tested against any model, and every turn into video button is an undisclosed Higgsfield affiliate link, so read the library as Higgsfield shaped.
🎬 Cool Video of the Week
This week's pick comes from Alex Patrascu: CHERRY COLA, a music video that went from a morning music prompt to a finished cut in about six hours. He prompted the track first, then built the video around it. The frame grid below is every shot in the piece, and the look holds across all of them. Tap the image to watch the clip and read his full process thread on X.
Clip by Alex Patrascu on X. Tap the image to watch.
The takeaway
Picture speaks a thousand words, same goes with AI. A detailed image prompt as a reference works much better than just words.