AI’s next product feature is proof of permission

A Twitch streamer can now stop Amazon from training future generative AI models on their streams, VODs, clips, chat, images and channel text, while the platform’s AI captions, recommendations and safety tools keep running.

That split is the whole story.

The obvious read is that this is another defensive platform move after creator anger about scraped content. I think the more interesting read is product architecture: AI systems are starting to separate using AI on the platform from using the platform to train AI.

That distinction sounds bureaucratic. It is not. It is the beginning of AI permission as a visible user experience.

Permission is becoming a product surface

For the first wave of generative AI, the magic trick was output. Type a prompt, get an image, song, video, voice or paragraph. The second wave is less glamorous: who consented, what was generated, what was captured by a camera, what can be recommended, what must be labelled, and who gets paid.

Spotify’s move points in the same direction. TechCrunch reported that Spotify will mark AI-generated artist identities with an “AI Persona” profile tag and keep their music out of editorial and algorithmic recommendations. The important part is not the badge by itself. Badges are cheap. The important part is distribution.

Spotify is saying synthetic artist profiles may exist, but they do not automatically get access to the same recommendation machinery as human performers. That turns AI disclosure from a passive label into a ranking policy. For product teams, that is the line that matters. Trust is rarely restored by a tooltip. It is restored when the system behaves differently because of what it knows.

Apple’s reported work on iPhone photo authenticity adds the provenance layer. The Verge reported that Apple is exploring stronger proof that iPhone photos came from a real camera rather than a generator, using image authenticity and metadata. This is a consumer version of a problem newsrooms and courts have cared about for years: can you prove where this image came from?

If Apple makes authenticity proof feel native to taking a photo, provenance stops being a specialist workflow. It becomes part of the default camera experience. That is a much bigger shift than another AI image detector, because detectors always arrive after the mess. Provenance starts at capture.

Then there is the licensing side. BMG announced a global alliance with Suno covering BMG’s recordings and publishing repertoire, tied to a planned music model developed with the music industry. BMG says participating artists and songwriters will have rights protected and be compensated, and the agreement also settles prior use of BMG recordings and publishing works.

That framing matters. AI music is being pulled away from the pirate-radio metaphor and into something closer to rights management. Messy, negotiated, probably imperfect, but recognisable.

The cross-domain parallel is land ownership. Economies do not scale because everyone is morally aligned; they scale because property boundaries, registries and transfer rules become legible enough for strangers to transact. Creative AI has been operating with the cultural equivalent of disputed land: amazing buildings, unclear title.

The next AI product moat may be less about model quality and more about proof of permission. Can your platform show that the data was approved for training? Can it prove the media was captured, not generated? Can it label synthetic identity without burying it? Can it route revenue to the right people?

Builders should treat this as infrastructure, not compliance theatre. The winning AI media products will not be the ones that generate the most stuff. They will be the ones users can trust to know where the stuff came from, what it is allowed to do, and who agreed to it.

---